Artificial intelligence for topic modelling in Hindu philosophy:

mapping themes between the Upanishads and the Bhagavad

Gita
1‡ 2‡
Rohitash Chandra , Mukul Ranjan ,

1 Transitional Artificial Intelligence Research Group, School of Mathematics and

Statistics, UNSW, Sydney, Australia
arXiv:2205.11020v1 [cs.CL] 23 May 2022

2 Department of EEE, Indian Institute of Technology Guwahati, Assam, India

‡These authors also contributed equally to this work.

* [email protected]

Abstract
A distinct feature of Hindu religious and philosophical text is that they come from a
library of texts rather than single source. The Upanishads is known as one of the oldest
philosophical texts in the world that forms the foundation of Hindu philosophy. The
Bhagavad Gita is core text of Hindu philosophy and is known as a text that summarises
the key philosophies of the Upanishads with major focus on the philosophy of karma.
These texts have been translated into many languages and there exists studies about
themes and topics that are prominent; however, there is not much study of topic
modelling using language models which are powered by deep learning. In this paper, we
use advanced language produces such as BERT to provide topic modelling of the key
texts of the Upanishads and the Bhagavad Gita. We analyse the distinct and
overlapping topics amongst the texts and visualise the link of selected texts of the
Upanishads with Bhagavad Gita. Our results show a very high similarity between the
topics of these two texts with the mean cosine similarity of 73%. We find that out of
the fourteen topics extracted from the Bhagavad Gita, nine of them have a cosine
similarity of more than 70% with the topics of the Upanishads. We also found that
topics generated by the BERT-based models show very high coherence as compared to
that of conventional models. Our best performing model gives a coherence score of 73%
on the Bhagavad Gita and 69% on The Upanishads. The visualization of the low
dimensional embeddings of these texts shows very clear overlapping among their topics
adding another level of validation to our results.

Author summary
Introduction
Philosophy of religion [1, 2] is a field of study that covers key themes and ideas in
religions and culture that relate to philosophical topics such as ethics and metaphysics.
Hindu philosophy [3–5] consists of schools developed for thousands of years which focus
on themes such as ethics [6], consciousness [4], karma [7, 8], logic and ultimate reality [5].
Hindu philosophy is at times referred as Indian philosophy [9, 10]. The philosophy of
karma and reincarnation are central to Hindu philosophy [10]. The Upanishads form the
key texts of Hindu philosophy and seen as the conclusion of the Vedas [11–15]. Hindu

May 24, 2022 1/39

 philosophy [9] consists of six major theistic (Astika) schools include Vedanta [14],
Samkhya [16], Nyāya [17], Vaisheshika [18], Mı̄māmsā [19], and Yoga [20]. Moreover,
Jain [21], Buddhist [22, 23], Carvaka [24] and Ājı̄vika [25] philosophy are the major
agnostic and atheistic (Nastika) schools of Hindu philosophy. There has been a lot of
interest of Hindu philosophy, particularly in the west, with a large list bibliography of
translations of key texts such as the Upanishads [26]. Moreover, Hindu and Buddhist
philosophy have parallels with development of specific themes in Greek philosophy [27].
The Upanishads and Bhagavad Gita are the foundational texts for Hindu philosophy.
These texts have been written much later in verse form in Sanskrit language, they have
been sung and remembered for thousands of years in the absence of a writing
system [11]. The Bhagavad Gita is is part of the Mahabharata which is known as the
one of oldest and largest epics written in verse in the Sanskrit language [28–30]. The
Bhagavad Gita is known as a concise summary of Hindu philosophy [10] with major
attribute which is the philosophy of karma [31–33]. The Upanishads is a collection of
philosophical texts of ancient India which marks the foundation in the history of
philosophy [34]. There are 108 key Upanishads of which some were lost in time. There
are 12 are prominent and have been well studied by western scholars [26, 35].
Nowadays, natural language processing (NLP) methods, that focuses in processing
and modelling language [36–38] are typically implemented via deep learning. NLP
considers tasks such as topic modelling, language translation, speech recognition,
semantic and sentiment analysis [37]. Sentiment analysis provides an understanding of
human emotions and affective states [39–41]. Recurrent neural networks such as
long-short term memory (LSTM) network models have been prominently used as
language models due to their capability to model temporal sequences [42]. LSTM
models have been improved for language modelling using attention based
mechanisms [43], and encoder-decoder LSTM framework with attention
(Transformer) [44, 45]. Bidirectional encoder representations from Transformer
(BERT) [46] model is a state-of-art pre-trained language model that features more than
300 million model parameters for language modeling tasks. Topic models help us better
understand a text corpus by extracting the hidden topics. Traditional topic model such
as linear discriminant analysis (LDA) [47] assumes that documents are a mixture of
topics and each topic is a mixture of words with certain probability score. Sentence
BERT (S-BERT) [48] improves BERT model by reducing computational time to to
derive semantically meaningful sentence embedding. Recent topic modelling frameworks
use S-BERT for embedding in combination with clustering methods [49–54].
BERT-based models have shown promising results for topic modelling [50, 54–56], which
motivates their usage in our study.
Religious linguistics refer to the study of religious sentences and utterances [57].
Major aim of the religious linguistic research is to create an analysis of various subject
matters related to religious sentences which include God, miracles, redemption, grace,
holiness, sinfulness along with several other philosophical interpretations [58–60]. Most
translations of the Bhagavad Gita and related texts come with interpretations and
commentary regarding philosophy and how the verses relate to issues at present [61].
Stein [62] presented a study about multi-worded expressions by extracting local
grammars based on semantic classes in the Spanish translation of the Bhagavad Gita
and found it to be promising for understanding religious texts and their literary
complexity. The role of multi-word expressions (MWE) could be a way to better
understand metaphorical and lyrical style of the Bhagavad Gita. Rajendran [63]
presented a study on metaphors in Bhagavad Gita using text analysis based on
conceptual metaphor theory (CMT). The analysis identified the source and target
domains for the metaphors, and traced the choice of metaphors to physical and cultural
experiences. The metaphors have been inspired by the human body and ancient India,

May 24, 2022 2/39

 which resonate with modern times. Rajput et al. [64] provided a statistical study of the
word frequency and length distributions prevalent in the translations of Bhagavad Gita
in Hindi, English and French from the original composition in Sanskrit. The Shannon
entropy-based measure estimated the vocabulary richness with Sanskrit as the highest,
and word-length distributions also indicated Sanskrit having longest word length.
Hence, the results demonstrated the inflectional nature of Sanskrit. Dewi [65] studied
metaphorical expressions and the conceptual expression underlying in them by
reviewing 690 sentences related to metaphor of life from Bhagavad Gita and analyzed
them using some conceptual metaphor theory. It was reported that the Bhagavad Gita
featured 24 conceptual metaphors among which life is an entity, life is a journey and
life is a continuous activity are the most frequent ones. Bhuwak [66] examined specific
ideas from Bhagavad Gita such as cognition, emotion, and behaviour by connecting
them with the context of human desire. It was reported that desires lead to behaviour
and achievement or non-achievement of desire lead to positive and negative emotions
which can be managed in a healthy way by self-reflection, contemplation and the
practice of karmayoga (selfless action). In our earlier work, BERT-based language
model framework was used for sentiment and semantic analysis as a means to compare
three different Bhagavad Gita translations where it was found that although the style
and vocabulary differ vastly, the semantic and sentiment analysis shows similarity of
majority of the verses [67]
Although the Bhagavad Gita and Upanishads have been translated into a number of
languages and studies about their central themes and topics have been prominent, there
is not much work in utilising latest advancements from artificial intelligence, such as
topic modelling using language models – powered by deep learning. In this paper, we
use advanced language models such as BERT in a framework to provide topic modelling
of the key texts of the Upanishads and the Bhagavad Gita. We analysis the distinct and
overlapping topics amongst the texts and visualise the link of selected texts of the
Upanishads with Bhagavad Gita. Our major goal is to map the topics in the Bhagavad
Gita with the Upanishads; since it is known that the Bhagavad Gita summarizes the
key messages in the Upanishads and there are studies about the parallel themes in both
texts [68]. We also provide a comparison of the proposed framework with LDA which
have been prominent for topic modelling.
The rest of the paper is organised as follows. In Section 2, we provide further details
about background behind the Bhagavad Gita and Upanishads. Section 3 presents the
methodology that highlights model development for topic modelling. Section 4 presents
the results and Section 5 provides a discussion and future work.

1 Background
1.1 BERT language model
BERT is an attention-based Transformer model [44] for learning contextualized
language representation where the vector representation of the every input token is
dependent on the context of its occurrence in a sentence. The Transformer model [44]
has been developed by using long short-term memory (LSTM) recurrent neural
networks [42, 69] with an an encoder-decoder architecture [70]. Transformer models
implement the mechanism of attention by weighting the significance of each part of the
input data which has been then prominent for language modelling tasks [44, 71].
BERT is first trained to understand the language (called pre-training phase) and the
context after that it is fine-tuned to learn the specific task such as neural machine
translation (NMT) [46, 72–76], question answering [77–82],and sentiment
analysis [83–87]. The pre-training phase of BERT involve two different NLP tasks such

May 24, 2022 3/39

 as masked language modelling (MLM) [46, 88, 89] and next sentence prediction
(NSP) [46]. MLM and NSP are semi-supervised learning tasks. In MLM, 15% words in
each input sequence is randomly replaced with a mask token and the model is trained
to predict these randomly masked input sequences based on the context provided by the
neighbouring non-masked words. In NSP, the BERT model learns to predict if two
sentences are adjacent to each other. In this way a BERT model is trained
simultaneously to minimize the combined loss function, and hence learn the
contextualized word embedding. In the fine tuning-phase one or more fully connected
layers are added on the top of final BERT layer based upon the applications. Since
BERT is pre-trained, it can be more easily trained further with datasets for specific
applications. In our earlier works, BERT-based framework has been used for sentiment
analysis of COVID-19 related tweets during the rise of novel cases in India [90]. Similar
framework using BERT was used for modelling US 2020 presidential elections with
sentiment analysis from Tweets in order to predict the state-wise winners [91].
Based upon the number of transformer blocks BERT [46] is available with two
variants: 1.) BERTBASE consists of 12 transformer blocks stacked on top of each other
with a hidden dimension embedding of 768 and 12 Attention heads, on the other hand
2.) BERTLARGE consists of of 24 transformer blocks with a hidden dimension
embedding of 1024 and 16 attention heads. BERTBASE has a total of 110 Million
parameters while BERTLARGE has a total of 340M parameters. BERT takes into
account the context for each occurrence of a given word, in comparison to context-free
models such as word vectors (word2vec) [92] and global vector (GloVe) [93] generate a
single word embedding representation for each word in the vocabulary.

1.2 Document embedding models

The universal-sentence-encoder [94] is a sentence embedding model that encodes
sentences into high-dimensional embedding vectors that can be used for various natural
language processing tasks. The model takes a variable length English text as an input
and gives 512-dimensional output vector.The model is trained with deep averaging
networks (DANs) [95] encoder, which simply takes the average of the input embeddings
for words and bi-grams and then pass them through one or more deep neural networks
to get the sentence embeddings. Sentence-BERT (S-BERT) [48] extends the BERT
model and Siamese and triplet network [96] to generate the sentence embeddings.
S-BERT uses BERT embeddings with a pooling layer to get the sentence-embedding (u
and v) of two sentences. S-BERT has been fine tuned with objective functions such as
triplet loss function and cosine similarity between u and v.

1.3 Clustering techniques

Clustering is a type of unsupervised machine learning that groups unlabelled data based
on a given similarity measure for a given dataset x(1) , ..., x(n) , where x(i) ∈ Rd is a
d-dimensional data point from the dataset. The goal of clustering is to assign each
data-point a label or a cluster identify. A large number clustering algorithms exits in
literature and we used two of them explained below for this work. Xu et al. [97]
presented an exhaustive list of different groups of clustering algorithms that includes:
1.) centroid based algorithms such as k-means clustering [98], regards the centroid of
data point as the centroid of the corresponding clusters; 2.) hierarchical based
algorithms such as agglomerative clustering [99] which creates a hierarchical relationship
among the data points in order to cluster them; 3.) density based algorithms that
connects an area with high density into clusters [100]; 4.) distribution based clustering
such as Gaussian mixture model [101] that assumes that data generated from same
distribution belongs to the same clusters.

May 24, 2022 4/39

 K-means clustering [102] clusters n-data points into k-clusters, where each data
point belongs to the cluster with the nearest mean. The k-means algorithm can be
explained in the three steps. First step involves initialization of k-centroid
corresponding to each clusters. In the second step a point is assigned to a the closest
cluster centroid. In the third step, centroid for each cluster is recalculated based on new
assigned data points and step 2 and 3 is repeated till convergence.
Hierarchical density based spacial clustering of application with noise
(HDBSCAN ) [103, 104] is a density-based hierarchical clustering algorithm that defines
clusters as highly dense regions separated by sparse regions. The goal of the algorithm
is to find high probability density regions which are our clusters. It starts with
estimating the probability density of the data by using the distance of the k th nearest
neighbors, defined as the core distance corek (x). If a region is dense, then the distance
of k th nearest neighbor will be less since more data point will fit in the region of small
radius. Similarly, for the sparse region,a larger radius would be used. A distance metric
called mutual-reachability-distance between two points a and b is defined in order to
formalize this idea of density and is given by Equation 1.
dmreach-k (a, b) = max{corek (a), corek (b), d(a, b)} (1)
where d(a, b) gives the euclidean distance between point a and b. This mutual
reachability distance is used to find the dense areas of the data but since the dense
areas are relative and different clusters (dense areas) can have different densities, the
entire data points can be modelled as a weighted graph with weight dmreach−k (a, b) of
edge between nodes a and b.

1.4 Dimentionality reduction techniques

Uniform manifold approximation and projection(UMAP) [105] for dimension reduction is
a non-linear dimensionality reduction technique which is constructed from the
theoretical framework based on Riemannian geometry and algebraic topology. The
detailed theoretical explanation of the algorithm is out of scope of this paper and can be
seen in the paper of McInnes et el. [105]. UMAP can be used in a way similar to
t-distributed stochastic neighbor embedding (t-SNE) [106] and principal component
analysis (PCA) [107] for dimensionality reduction and to visualize high dimensional
data.
Latent dirichlet allocation (LDA) [47] is a generative probabilistic model for the
topic modelling of the corpus based on word frequency. The basic idea behind the
model is that, each document is generated by a statistical generative process and hence
each document can be modelled as a random mixture of latent topics and each topic is
mixture of words characterised its distribution. A word denoted by w and indexed from
1 to the vocabulary size V and a document is given by w = {w1 , w2 , ..., wN }, where wi
is the ith word in the sequence [47]. The generative process involved in the algorithm
can be summarized as 1.) fix the number of topic and hence the dimensionality of the
Dirichlet distribution and that of the topic variable z and sample θ(per-document topic
proportion) from a Dirichlet prior Dir(α) 2.) sample a topic zn from a multinomial
distribution p(θ; α) and then 3.) sample a word wn from multinomial probability
distribution conditioned on zn , p(wn |zn , β). Overall probability of document w
containing N words can be given by Equation 2.
Z N X k
!
Y
p(w) = p (wn | zn ; β) p (zn | θ) p(θ; α)dθ (2)
θ n=1 zn =1

Given a corpus of M documents D = {w1 , ..., wM }, the EM algorithm can be used

to learn the parameters of an LDA model by maximizing a variational bound on p(D),

May 24, 2022 5/39

 Texts Translator Year
The Bhagavad Gita [109] Eknath Easwaran 1985
The Upnishads [110] Eknath Easwaran 1987
The Ten Principal Upanishads [111] Shri Purohit Swami & W.B. Yeats 1938
108 Upanishads [113] The International Gita Publication –
Table 1. Details of the texts used for topic modelling.

as seen in Equation 3.
M
X
log p(D) ≥ Eqm [log p(θ, z, w)] − Eqm [log qm (θ, z)] (3)
m=1

LDA has been used for several language modeling tasks that include the study of the
relationship between two corpora using topic modeling [108] which is also the focus of
our study.

2 Methodology
2.1 Datasets
We evaluated a number of prominent translations of the Bhagavad Gita and the
Upanishads. In order to maintain the originality of the themes and ideas of these two
classical Indian texts, we used the older and the most popular translations for this work.
We chose Eknath Eashwaren’s translation since he directly translated from Sanskrit to
English and translated both texts [109, 110], hence it would be not be creating a
translation bias for topic modelling and comparison of the topics between the texts.
Eknath Easwaran ( 1910 – 1999) was a professor of English literature in India and later
moved to the United States where he translated these texts. In addition, we chose the
translation by Shri Purohit Swami and William Butler Yeats [111] for further
comparison. W. B Yeats (1865 – 1939) was Irish poet, dramatist, prose writer and
known as one of the foremost figures of 20th-century literature. Shri Purohit Swami
(1882 – 1941) was a Hindu teacher from Maharashtra, India. The translation of the
Upanishads by them is special since it has been done jointly by prominent Indian and
Irish scholars and captures Eastern and Western viewpoints. Table 1 provides further
details of the texts. Note that Shri Purohit Swami also translated the Bhagavad
Gita [112] which can be used in future analysis, and not used in this work.
The Bhagavad Gita consist of 18 chapters which features a series of questions and
answers between Lord Krishna and Arjuna that range with a range of topics that
includes the philosophy of Karma. The Mahabharata war lasted for 18 days [114];
hence, the organisation of the Gita is symbolic.
The Upanishads [110] translated by Eknath Eashwaren provides a commentary and
translation of the 11 major and 4 minor Upanishads. The 108 Upanishads [113] is a
collection of the translation and commentary of all 108 Upanishads in a single book
compiled by the Gita Society. The translation and commentary is done by a group of
spiritual teachers who have tried to recover the Upanishads which have believed to be
lost earlier; however, there are not much details about how they have recovered
them [113]. The Chandogya Upanishad has highest number of words followed by the
Katha Upanishad and the Brihadaranyaka Upanishad. The Ten Principal
Upanishads [111] consists of the translation of the 10 major Upanishads. This text does
not have a separate explanation for each Upanishad unlike the Upanishads by Eknath
Easwaran. The Brihadaranyaka Upanishad consists of the highest number of words
followed by the Chandogya Upanishad and Katha Upanishads. The Chandogya

May 24, 2022 6/39

 Upanishad is one of the largest Upanishads consisting of 8 chapters which can be
divided into 3 natural groups according to the philosophical ideas [115]. The first group
(Chapter 1 and Chapter 2) deals with the structure and different aspects of the
languages and its expression, particularly with the syllable ”Om” that is used to
describe Brahman and beyond. The second group (Chapter 3-5) consists of the ideas of
universe, life, mind and spirituality. The third group (Chapter 6-8) deals with the more
metaphysical questions such as nature of reality and Self [115]. Since first five chapters
are intermixed with rituals, Shri Purohit Swami omitted them from in his
translation [111] along with some passages from the Brihadaranyaka Upanishad. Other
authors also state that some of the passages of the Brihadaranyaka Upanishad has been
omitted due to the repetitions [111]. Brihadaranyaka Upanishad, consisting of 6
chapters discusses about different philosophical ideas including one of the earliest
formulation of the Karma doctrine (Verse 4.4.5), ethical ideas such as self-restraint
(Damah), charity (Danam) and compassion (Daya) and also other metaphysica topics
related to philosophy of Advaita Vedanta. Eknath Easwaran [110] translated this
chapter as the Forest of Wisdom which starts with the one of Vedic theories of the
creation of the Universe and then the dialogue between a great sage, Yajnavalkya, and
his wife Maitreyi which is a deep spiritual discussion about death, possession, self,
Brahman (God) and the Atman (Self). It contains one of the earliest psychological
theories relating the human body, mind, ego and the Self. The Katha Upanishad is one
of the legendary story of a small boy Nachiketa who met Yama (the god of Death) asks
with him different questions about the nature of life, death, man, knowledge, Atman
and Moksha (liberation). The Katha Upanishad consists of 2 chapters each consisting of
3 sections.

2.2 Framework
Our major goal is to map the topics in the Bhagavad Gita with Upanishads. We begin
by selecting 12 prominent Upanishads (Isha, Katha, Kena, Prashna, Munda, Mandukya,
Taittiri, Aitareya, Chandogya, Brihadaranyaka, Brahma, Svetasvatara) from the text
translated by Eknath Easwaran [110]. The major reason that we selected both by the
same author for this task is to eliminate any bias in translation for topic modelling.
However we also considered other translations as mentioned in the table 1 and found
that these bias does affect the similarity matrix. For example when we compared the
similarity between the Upanishads by Eknath Easwaran and the Bhagavad Gita by the
same authors, average similarity score is 3% better than that of the Bhagavad Gita by
Eknath Easwaran and the Upanishads by Shri Purohit Swami. Finally, we also presents
the visualization of the topics space of 108 Upanishads and its different part divided
based on the original Vedas these Upanishads are originated from..
Next, we present a framework that employs different machine learning methods for
topic modelling. Figure 1 presents the complete framework for the analysis and topic
modelling of the respective texts given in Table 1. In Figure 1, the first stage consists of
conversion of PDF files and text pre-processing as discussed in the previous section. In
the second stage, we use two different sentence embedding models 1.) universal sentence
encoder (USE) and 2.) Sentence-BERT(SBERT) for generating the word and
documents embedding which is later passed thorough the topic extraction pipeline to
generate the topic vector and finally we compared our results with the classical topic
modelling algorithm LDA [47] across different corpus. Our framework to generate topics
is similar to Top2Vec [52], however we also used other clustering algorithms like
K-Means. First, USE and SBERT are used to generate the joint semantic embeddings
of documents and words. Since these embeddings are generally in higher dimension
which is very sparse, we need to reduce the dimension of the embeddings to get the
dense areas. We use dimensionality reduction techniques like UMAP and PCA for

May 24, 2022 7/39

 LDA

SBERT
Preprocessing Topic Extraction
PDF files Text Files Final Topics Topic Comparison
& cleaning Pipeline

USE

Create Reduce Topics

Find Low
Semantic Find Dense Find Topic Using
Dimensional
Embedding of Clusters of Vectors and Hierarchical
Document
words and Documents Topic words Topic
Embedding
Documents Reduction

Fig 1. Framework for semantic similarity and Topic Modelling

reducing the high dimensional embedding vectors generated by the S-BERT and the
USE. Next, we find dense clusters of topics in the document vectors of the corpus using
clustering algorithms like HDBSCAN and K-Means. These clusters are represented by
the centroid of document vectors in the original dimension, which is called as topic
vectors [52]. Next, we find top N(N = 50 in our case) nearest words to the topic vectors
which represent our final topic. Topic vectors also allows us to group the similar topics
and hence reduce the number of topics using Hierarchical Topic Reduction [52].
Most of the topic modelling research [52, 116, 117] involves the bench-marking model
results on pre-existing datasets such as the 20 News Groups dataset [118], the Yahoo
Answers dataset [119, 120], Web Snippets dataset [121], W2E datasets [122]. These
datasets have been prepared to be used for the algorithms bench-marking tasks and
consists of the fixed number of documents and words. The 20 News Groups Datasets for
example consists of 15,465 documents and 4,159 words [116]. Tweets have also been
used for topic modelling tasks [123–125]. Jonsson et al. [123] for example, collected
tweets from Twitter to prepare a datasets of 129,530 tweets and used LDA [47],
Biterm-Topic-Model (BTM) [124] and a variation of LDA algorithms for topic modelling
to compare their performance. In case of Twitter based topic modelling datasets, a
tweet is considered as Document, though Jonsson et al. [123] aggregate documents to
form pseudo-documents and found that it solves the poor performance of LDA on
shorter documents. Murakami et al. [126] used research papers published in the journal
Global Environmental Change (GEC) from the first volume (1990/1991) to Volume 20
(2010) as the corpus for the topic modelling. They divided the a paper into several
paragraph blocks and modelled them as a documents of the corpus.
Our dataset can be seen as similar to Murakani et al. [126]. The Bhagavad Gita and
Upanishads are written in verse form and to maintain the originality of the texts, most
of the translations also preserve the numbering of the verses. Other than the verses, the
translations also contain commentary by the translator of the texts. While creating the
datasets, we first created documents based on the verse number in the texts, i.e a verse

May 24, 2022 8/39

 Original Documents Transformed Documents
II-5(a). What winds up empirical life is (its) appearance
what winds up empirical life is its appearance as unreal.
as unreal.
”What discipline is required to know, \u2018this is a pot,
”what discipline is required to know this is a pot except
except the adequacy of the means of right \u2019 \n
the adequacy of the means of right knowledge.”
knowledge ?”
Lord, have we not prophesied in thy name? and in thy Lord have we not prophesied in your name and in
name have cast out \n devils? and in thy name done your name have cast out devils and in your name
many wonderful works? done many wonderful works.
Table 2. Processed text after removing special characters and transforming archaic words into modern English

is considered as a document of the corpus, where the numbering are clearly mentioned.
In other cases when verse numbers are not mentioned clearly, we considered one
paragraph as one documents. In case of the commentary, we split the commentary into
smaller parts to make them a document as done by Murakami et al. [126]. The statistics
in terms of number of documents, number of words (# words), average number of words
(avg # words), and number of verses (# verses) of different corpus (text files) and their
details can be found in Table 3.

2.3 Text data extraction and processing

In order to process the files given in printable document format (PDF), we converted
them into text files. Most of the PDF files were generated from the scanned images of
the printed texts, hence we used optical character recognition (OCR) based open-source
library1 . This conversion from PDF to text file gave us a raw dataset consisting of all
the texts shown in Table 1. Next, pre-processing done on the entire datasets, which
consists of the following steps.

1. Removing unicode characters generated in the text files due to noise in the PDF
files;
2. Normalizing(assigning uniform verses from each text) verse numbering in the
Upanishads and the Bhagavad Gita;
3. Replacing the archaic English words such as ”thy” and ”thou” with modern
English words like your and you;
4. Removing the punctuation, extra spaces, and lower-casing;
5. Removing repetitive and redundant sentences such as ”End of the Commentary”.
Examples of selected text from the original document along with the processed text
is shown in Table 2. The original text and processed text has been given in the Github
repository 2 . In topic modelling literature, word is the basic unit of data which is
defined to be an items from vocabulary indexed by {1, ..., V }, where V is the
vocabulary size. A Document is a collection of N words represented by
w = {w1 , w2 , ..., wN }, where wi is the ith word in the sequence. The corpus is
considered as a collection of M documents denoted by D = {w1 , w2 , ..., wM } [47].
1 https://github.com/writecrow/ocr2text
2 https://github.com/sydney-machine-learning/topicmodelling_vedictexts/

May 24, 2022 9/39

 Corpus # Documents # Words Avg # words # Verses
The Upanishads(Eknath Easwaran) 862 40737 47.26 708
The Bhagavad Gita(Eknath Easwaran) 700 20299 27.50 700
Ten principal Upanishads 1267 27492 21.70 1267
108 Upanishads 6191 405559 65.50 6191
Table 3. Dataset Statistics

2.4 Technical details

In our framework, S-BERT and USE are used for the task of generating sentence
embeddings. We used pre-trained S-BERT3 , which has been trained on a large
multilingual corpus. The model uses DistilBERT [127] as the base transformer model,
then its output is pooled using an average pooling layer and a fully connected (dense)
layer is used finally to give a 512 dimensional output. We used different combination of
dimensionality reduction techniques and clustering algorithms with the pre-trained
semantic embeddings to get the final topics for each corpus.
The embedding dimension is reduced to the 5-dimension using the selected
dimensionality reduction techniques i.e UMAP and PCA. UMAP uses two important
parameters, n neighbors and min dist in order to control the local and global structure
of the final projection. We fine-tuned these parameters to optimize the topic-coherence
metric and use the final UMAP model with the default min dist value of 0.1,
n neighbors value of 10 and the n components value of 5, which is the final dimension of
the embeddings. We set the random-state to 42 and use cosine-similarity as the
distance metric.
After getting the embedding of the documents in the reduced dimensions we used
two different clustering algorithms - HDBSCAN [103, 104] and K-Means [102]
algorithms to cluster the documents where each clusters represent a topic. We
fine-tuned different parameters of HDBSCAN to get the optimal value of topic
coherence metric(Topic coherence is discussed with great details in next section), which
represents how good our generated topics are.We chose the number of topics obtained at
the optimal value of topic coherence metric as the optimal number of topics and used
the same number as the value of K for K-Means Clustering algorithms. The
min cluster size defines the smallest grouping size to be considered as cluster, we set it
to 10. Finally, in the remaining two parameters, we use metric = euclidean and
min samples = 5. The k-means algorithm is trained for the 300 iterations, with the
num clusters parameters same as the number of labels found using HDBSCAN.

3 Results
3.1 Data Analysis
We begin by reporting key features of the selected texts (datasets) as shown in Table 3.
The Upanishads by Eknath Easwaran contains 862 documents, 40737 words and 705
verses. Since this text contains explanation by the authors as well so the number of
documents is more than the number of verses for this text. Ten Principal Upanishads
by W.B. Yeats and Shri Purohit Swami Consists of 1267 documents and same number
of verses as well. The corpus also consists of 27492 words with an average of 21.70
words per documents. The Bhagavad Gita by Eknath Easwaran consists of 700 verses
and the same number of documents along with 20299 words with an average of 21.70
words per documents.
3 https://huggingface.co/sentence-transformers/

May 24, 2022 10/39

Fig 2. Chapter Wise Word count for different texts in the dataset

May 24, 2022 11/39

 Figure 2 shows the chapter-wise word count of the respective corpus. The Bhagavad
Gita consists of 18 chapter and we see that Chapter 2 has highest number of words,
followed by Chapter 18 and Chapter 11. This is because these chapters contains
relatively more number of verses and explains much deeper topics of Hindu philosophy.
Chapter 18 of the Bhagavad Gita contains the highest number of verses(78) followed by
the chapter 2 which contain 72 verses and chapter 11 containing 55 verses. Chapter 2 of
the Bhagavad Gita discuss about the Samkhya and Yoga School of Hindu
Philosophy [109, 128, 129]. It teaches about the cosmic wisdom (Brahm Gyan) and the
methods of it attainment along with along with the notion of qualia (Atman/self), duty,
action (karma), selfless action (karma yoga), rebirth, afterlife, and the qualities of
self-realized individuals (muni) [129]. Eknath Easwaran [109] claimed this chapter as an
overview for the remaining sixteen chapters of the Bhagavad Gita. Chapter 11 is also
called as the ”Vishwa Roopa Darshana Yoga” [128] which has been translated as ”The
Cosmic Vision” by Eknath Easwaran [109], and ”The Yoga of the Vision of the
Universal” Form [128] by Swami Chinmayananda. This chapter talks about the supreme
vision of Lord Krishna which made Arjuna experience deep peace and joy of Samadhi
(enlightenment) along with the feeling of being terrified at the same time [109, 129].
When terrified, Arjuna asks about the identity of the cosmic vision of God. Lord
Krishna replies in verse-32 of Chapter 11 that came into Robert Oppenheimer’s mind
when he saw the atomic bomb explode over Trinity in the summer of 1945 [109, 130]. He
mentioned, ”A few people laughed, a few people cried. Most people were silent. I
remembered the line from the Hindu scripture, the Bhagavad Gita; Vishnu is trying to
persuade the Prince that he should do his duty and, to impress him, takes on his
multi-armed form and says, Now I am become Death, the destroyer of worlds.”
The n-gram [131] is typically used to provide basic statistics of a text using a
continuous sequence of words or other elements. Bi-grams and tri-grams are typical
examples of n-grams. Figure 4 shows the count of top 10 bigrams and trigram along
with the top 20 words for the Upanishads by Eknath Easwaran. In the case of the
Upanishads by Eknath Easwaran, (lord, love) is the most frequent bigram which has
occurred more than 60 times followed by (realize, self) and (go, beyond). In the same
corpus, when we look at the trigram’s bar plot we find that (united, lord, love), (self,
indeed, self) and (inmost, self, truth) to be the top 3 trigrams of the corpus. Similarly
Figure 5 shows unigrams, bigrams and the trigrams of the Ten Principal Upanishads.
Although these n-grams just state the frequency of occurrence of the continuous
sequence of words, they give a rough idea about the themes and topics discussed in the
corpus. This can be seen in Figure 13 that a lot of topics do contain these words. We
can see that ’self’ is one of the predominant word of topic 4 and topic 8 of the ten
principal Upanishads. In fact when we observe these topics carefully, we find that the
entire topic is related to the theme of ’self’. Similarly, we find words ’lords’, ’god’ and
’sage’ to be predominant words in topic 1 and topic 3 of the ten Principal Upanishads.
Figure 3 shows the bigrams, trigrams and word count for the Bhagavad Gita. We see
that ’arjuna’, ’self’, ’krishna’, ’action’ and ’mind’ are top 5 words of the Bhagavad Gita.
Among the bigrams and trigrams, we see that (every, creature), (supreme, goal) and
(selfless, service) are top 3 bigrams while (attain, supreme, goal), (beginning, middle,
end) and (dwells, every, creature) are the top 3 trigrams of the text. Since Arjuna and
Krishna are the protagonists of the text, it is obvious for them to be among the top
words of the text. We see that other than these, ’self, action, and mind’ are the
prominent words that give us a basic idea about the themes that can be verified from
the topics presented in Figure 12. Topic 1 of the Bhagavad Gita shown in Figure 12
shows all the names of the Hindu spiritual entities (deities). We find that Krishna and
Arjuna are one of the major one among them. This topic also includes other entities
and deities such as Jayadratha, Vishnu and Bhishma that have been mentioned by the

May 24, 2022 12/39

 Lord Krishna in the text. The words related to the ’self’ can be seen in topic 2 of Figure
12; hence, we can conclude that themes related to self are present in Topic 2 identified
by our framework. We can also see that topic 13 of the Bhagavad Gita contains the
words related to ’action’ (karma) which is also one of the top 5 words of the texts.
In terms of the individual word frequency, we find that ”self” is one of the most
occurred word in all the three corpus which is a major theme of Hindu Philosophy. The
”self” is the translation from the Sanskrit word ”Atman”, which refers to the spirit, and
more precisely ”qualia” as known in the definition pertaining to the hard problem of
consciousness [132]. The ”Atman” is also often translated as consciousness and there
are schools of thought (Advaita Vedanta [133]) that sees the Atman as Brahman (loosely
translated to the concept of God or super-consciousness) [4, 134]. Often, it is wrongly
translated to the term soul which is an Abrahamic religious concept, where humans only
to have the soul and the rest of life forms do not [135]. Atman, on the other hand, is the
core entities of all life forms and also of non-life forms in Hindu philosophy. Not only in
Upanishads but it has been explained in The Bhagavad Gita as well with a great details.
Finally, ”attain supreme goal” is the most occurred trigram of the Bhagavad Gita which
suggests that the Bhagavad Gita talks about attaining supreme goal with a great details
along with the other philosophical topics. The Bhagavad Gita is also known as the
Karma Upanishad or the text that focuses on philosophy of karma (action/work) [8].
The major focus of the Bhagavad Gita is karma philosophy given a conflicting situation
and the path to self realisation as the goal of life, and hence, it has also been recognised
as a book of leadership and management [136, 137], and psychology [138].

('every', 'creature') arjuna

('supreme', 'goal') self
('selfless', 'service') krishna
('selfish', 'desire') action
('arjuna', 'krishna') mind
('pleasure', 'pain') one
Bigrams and Trigrams

('selfish', 'attachments') see

('spiritual', 'wisdom') supreme

Top 20 Words
('sense', 'objects') every
('attain', 'supreme') among
('attain', 'supreme', 'goal') wisdom
('beginning', 'middle', 'end') without
('dwells', 'every', 'creature') lord
('senses', 'mind', 'intellect') path
('sattva', 'rajas', 'tamas') selfish
('lifes', 'supreme', 'goal') attain
('gone', 'beyond', 'gunas') knowledge
('whose', 'consciousness', 'unified') world
('sanjaya', 'spoken', 'words') body
('giving', 'self', 'discipline') spiritual
0 5 10 15 20 0 20 40 60 80 100 120
Counts Counts
Fig 3. Visualisation of top 20 words, and top 10 bigrams and trigrams for the
Bhagavad Gita.

3.2 Modelling and Predictions

3.2.1 Topic Coherence
Quantitative evaluation of the topic model is one of its major challenge. Initially, topic
models were evaluated with held-out-perplexity but it does not correlate with the
human evaluation [139]. A topic can be said to be coherent if all or most of the words of
the topic support each other or are related [140]. Human evaluation of topic coherence
is done in two ways: 1.) rating, where human evaluators rate the topic quality on a
three point topic quality score, and 2.) intrusion, where each topic is represented by its
top words along with an intruding word which has very low probability of belonging to
the topic since it does not belong in the topics uncovered. It is a behavioral way to
judge topic coherence and measured by how well a human evaluator can detect the

May 24, 2022 13/39

 ('lord', 'love') self
('realize', 'self') one
('go', 'beyond') life
('self', 'supreme') world
('let', 'us') lord
('birth', 'death') us

Bigrams and Trigrams

('joy', 'one') death
('know', 'self') upanishads

Top 20 Words
('dear', 'one') brahman
('united', 'lord') mind
('united', 'lord', 'love') body
('self', 'indeed', 'self') know
('inmost', 'self', 'truth') love
('truth', 'self', 'supreme') beyond
('self', 'supreme', 'shvetaketu') supreme
('come', 'everything', 'inmost') without
('everything', 'inmost', 'self') heart
('nothing', 'come', 'everything') see
('go', 'beyond', 'death') fire
('self', 'truth', 'self') joy
0 10 20 30 40 50 60 0 100 200 300 400 500
Counts Counts
Fig 4. Visualisation of top 20 words, and top 10 bigrams and trigrams for the
Bhagavad Gita. Upanishads by Eknath Easwaran.

('yadnyawalkya', 'said') self

('bright', 'eternal') said
('janaka', 'said') spirit
('eternal', 'self') man
('said', 'narada') life
('self', 'lives') one
Bigrams and Trigrams

('lord', 'said') body

('wise', 'man') light
Top 20 Words

('prajapati', 'said') mind

('shwetaketu', 'explain') may
('bright', 'eternal', 'self') food
('yes', 'said', 'sanatkumar') knows
('eternal', 'self', 'lives') fire
('woven', 'warp', 'woof') water
('may', 'please', 'provided') everything
('honey', 'beings', 'beings') world
('please', 'provided', 'worships') know
('lord', 'said', 'shwetaketu') gods
('nothing', 'spirit', 'anything') lives
('said', 'narada', 'yes') nothing
0 5 10 15 20 25 30 35 0 50 100 150 200 250 300
Counts Counts
Fig 5. Visualisation of top 20 words, and top 10 bigram and trigrams for the ten principal Upanishads.

May 24, 2022 14/39

 intruding word [139, 141]. Automated topic coherence metric based on normalized
pointwise mutual information(NPMI) correlates really well with the human evaluation
and interpretation of the topic coherence [141–144]. Röder et. al. [145] provided a
detailed study on the coherence measure and its correlation with the human topic
evaluation data. We use the topic coherence NPMI measure(TC-NPMI) [145] as a
metric to fine tune and evaluate different models on different corpus. Equation 4 gives
the NPMI for a pair of words (wi , wj ) from the top N (set to 50) words of a given topic:
 
P (w ,w )+
log P (wii)·Pj(wj )
NPMI (wi , wj ) =   (4)
− log (P (wi , wj ) + )

where, the joint probability P (wi , wj ), i.e the probability of the single word P (wi )
is calculated by the Boolean sliding window approach (window length of s set to the
default value of 110). A virtual document is created and the count of occurrence of the
word (wi ) or the word pairs (wi , wj ), and then it is divided by the total number of the
virtual documents.
We use TC-NPMI as the topic-coherence measure to evaluate different topic models
and tune different hyper-parameters of different algorithms. Table 4 shows the value of
metric for different model on different datasets. We trained the LDA model for 200
iterations with other hyper-parameters set to the default value as given in the
gensim [146] library. We fine-tuned the number of topics parameters to get the optimal
value of TC-NPMI.
Next, we evaluate different components in the BERT-based topic model presented
earlier (Figure 1 from where we have five major approaches: 1.)
SBERT-UMAP-HDBSCAN, 2.) SBERT-UMAP-KMeans, 3.) USE-UMAP-HDBSCAN,
4.) USE-UMAP-KMeans, and 5.) LDA. In Table 4, we observe that in the case of the
Bhagavad Gita, the combination of USE-UMAP-KMeans gives the best TC-NPMI score
on both the datasets with a very slight difference when compared to
USE-UMAP-HDBSCAN and SBERT-UMAP-KMeans. Note that high TC-NPMI
results indicate better results. In the case of the Upanishads, we find a similar trend.
We also observe that LDA does not perform well, even after fine-tuning the number of
topics parameters to optimize the topic coherence.
Although the use of KMeans for the clustering component gives the best result, we
choose USE-UMAP-HDBSCAN to find the topic similarity between the Upanishads and
The Bhagavad Gita in the next section. This is because HDBSCAN does not require us
to specify the number of clusters, that corresponds to the number of topics, beforehand.
USE-UMAP-HDBSCAN gave 18 topics for the corpus - the Upanishads [110] for the
optimal value of the topic coherence mentioned in Table 4. Similarly, we got 14 topics
from the Bhagavad Gita [147]. In the case of the 108 Upanishads which contains larger
number of documents as compared to the rest of the texts, we got more topics for the
optimal values of topic coherence. However, we reduced the number of topics using
hierarchical topic reduction [52] in some of cases for example, while comparing the topic
similarity of the Bhagavad Gita and the Upanishads. Since the number of documents
and words are different for different corpus as seen from Table 3, the number of topics
obtained are different for different corpus. For example, in the Ten Principal
Upanishads – there are 1267 documents and we got 28 topics for them at the optimal
value of topic coherence. Similarly for 108 Upanishads, there are 6191 documents which
gives 115 topics (Table 4) for the model SBERT-UMAP-HDBSCAN at the optimal
value of topic coherence. Also, while plotting the semantic space for the different topics
obtained by our model as shown in Figure 8, Figure 10, and Figure 11, we reduced the
number of topics to 10 in order to visualize the topic’s semantic space clearly.

May 24, 2022 15/39

 Datasets
Model Bhagavad Gita The Upanishads Ten Principal Upanishads 108 Upanishads
# Topics TC-NPMI # Topics TC-NPMI # Topics TC-NPMI # Topics TC-NPMI
SBERT-UMAP-HDBSCAN 14 0.70 18 0.67 28 0.70 115 0.63
SBERT-UMAP-KMeans 14 0.72 18 0.69 28 0.73 115 0.66
USE-UMAP-HDBSCAN 14 0.73 18 0.67 32 0.73 125 0.64
USE-UMAP-KMeans 14 0.73 18 0.69 32 0.71 125 0.67
LDA 20 0.32 20 0.29 24 0.31 140 0.39
Table 4. Value of topic coherence metric (TC-NPMI) for different Corpus

3.2.2 Topic similarity between the Bhagavad Gita and the Upanishads
There are studies that suggest that the Bhagavad Gita summarizes the key themes of
the Upanishads and various other Hindu texts [148–150]. The Bhagavad Gita along
with the Upanishads and the Brahma Sutras is known as the Prasthanatrayi [151–155],
literally meaning the three points of departure [151], or the three sources [153] ) which
makes the three foundational texts of the Vedanta school of Hindu
philosophy [13, 14, 149, 150, 156]. Sargeant et al. [148] stated that the Bhagavad Gita is
the summation of the Vedanta. Nicholson et al. [150] and Singh et al. [149] regarded the
Bhagavad Gita to be the key text of Vedanta.
Another source which discusses a direct relationship between the Bhagavad Gita and
the Upanishads is the Gita Dhayanam (also sometimes called Gita Dhyana and Dhyana
Slokas) which refers to the invocation of the Bhagavad Gita) [147, 157, 158]. We need to
note that Gita Dhayanam is an accompanying text with 9 verses used for prayer and
meditation that complements the Bhagavad Gita. These 9 verses are attributed
traditionally to Sri Madhusudana Sarasvati and are generally chanted by the students of
Gita before they start their daily studies [157]. These verses offer salutations to various
Hindu entities such as the Vyasa, Lord Krishna, Lord Varuna, Lord Indra, Lord Rudra
and the Lord of the Maruta and also characterises the relationship between the
Bhagavad Gita and the Upanishads. The 4th verse of the Gita Dhyanam states a direct
cow and milk relationship between the Upanishads and the Gita. Eknath
Easwaran [147] translated the 4th verse as ”The Upanishads are the cows milked by
Gopala, the son of Nanda, and Arjuna is the calf. Wise and pure men drink the milk,
the supreme, immortal nectar of the Gita”. Although these relationships have been
studied and retold for centuries, there are no existing studies that establishes a
quantitative measure to this relationship using modern language models. Next, we
evaluate and discuss similar relationships both quantitatively using a mathematical
formulation and also qualitatively by looking at the topics generated by our models as
shown in Tables 5, 6, and Figures 12, 13.
In order to evaluate the relationship between the Bhagavad Gita and the
Upanishads, we used the obtained topics to find a similarity matrix as shown in the
heatmap of Figure 6. The vertical axis of the heatmap shows the topics of the Bhagavad
Gita while the horizontal axis of the heatmap represent the topics of the Upanishads.
The heatmap represents the cosine similarity of the topic-vector obtained by the topic
model. Therefore, in each of the topics obtained from the Bhagavad Gita, we calculate
its similarity with all the topics of the Upanishads and then find the topic with
maximum similarity. This operation can be mathematically represented by the
Equation 5a. We represent the number of topics in Gita by Ngita and the number of
topics in Upanishads by Nupan . In each topic Tigita from the Bhagavad Gita, we explore
and find the most similar topic from Upanishads Tiupan . The topics and their similarity
score can be found in Table 5 and Table 6. We observe a very high similarity in the
topics of the Bhagavad Gita and two different texts of Upanishads (shown in Table 5

May 24, 2022 16/39

 and Table 6). These tables also show the mean similarity score which is given by the
average of all the similarity scores as shown in Equation 5b and given below:
Nupan
 
Tupan
i = arg max Sim Vigita , Vjupan (5a)
j=1
NP
gita N  
upan
max Sim Vigita , Vjupan
i=1 j=1
AvgSim = (5b)
Ngita

where Vigita and Viupan represent the ith topic vectors of the Bhagavad Gita and the
Upanishads, respectively. Sim(.) represent the similarity measure defined by equation 6,
which is cosine similarity in our case. There are various other measures of similarity
score between two vectors; however, the cosine similarity is used widely in the
literature [159–161]. One of the major reason for this is its interpretability. Value of
cosine similarity between any two vector lies between 0 and 1. A value closer to 1
represent that vectors are almost similar to each other and a value closer to 0 represent
that they are completely dissimilar.
The cosine similarity between any two vectors U and V is represented by Equation 6.
Since the topic vector contains contextual and thematic information about a topic, the
similarity score gives us extent of closeness of the themes and topics of the Bhagavad
Gita and the Upanishads.
U·V
Sim(U, V) = cos(θ) = (6)
kUkkVk

We can observe from the Table 5 that a number of the topics of Bhagavad Gita are
similar to the topics of the Upanishads with more than 70% similarity. We also find
that topic 4 of the Bhagavad Gita is similar to that of the topic 5 of the Upanishads
(Eknath Easwaran) with a similarity of 90%. We can see that both of these topics
contains almost similar words. Similarly, topic-5 of the Bhagavad Gita has a similarity
of 86% when compared with topic 8 of the Upanishads. Both of these topics are are
related to immortality and death. The similarity can be observed via Table 5; for
example, topic-1 of both Bhagavad Gita and the Upanishads (Eknath Easwaran)
consists of the words related to Hindu deities and entities such as Krishna, Arjuna,
Vishnu and Samashrava, they also have a similarity of 76%.
Figure 8 represents a visualization of the semantic space of the Bhagavad Gita and
the Upanishads with given topic labels. Although we find in Table 4 that Bhagavad Gita
and the Upanishads gave 14 and 18 topics respectively, we are only presenting 10 topics
from both of these texts in order to have a clear visualization. Each dots in the diagram
represent the two dimensional (2D) embedding of each of the documents of the corpus.
These topics can be seen in Figure 12 along with some of the most relevant documents
of the text with their source. Figure 12 represents the themes related to the deities and
the entities of the Hindu philosophy. We can also observe that documents relevant to
topic-1 have been originated form chapter 1, 3 and 10. These all are the verses
containing the name of the Hindu deities. Topic-2 of the same table encapsulate the
idea of self, worship, desire and fulfillment. A similar pattern can be observed in Table 6
which represent the topics and documents of the Ten Principal Upanishads [111].

3.2.3 108 Upanishads

Finally, we apply a selected respective topic modelling approach
(USE-UMAP-HDBSCAN) from our topic modelling framework (Figure 1) for analysis of
the complete 108 Upanishads. We note that the 108 Upanishads are also known as

May 24, 2022 17/39

 1.0
ic-1 0.76 0.36 0.59 0.48 0.49 0.49 0.5 0.5 0.4 0.39 0.36 0.63 0.43 0.34 0.38 0.32 0.32 0.38
Top
ic-2 0.55 0.45 0.68 0.38 0.67 0.62 0.56 0.48 0.7 0.58 0.76 0.47 0.59 0.74 0.46 0.56 0.42 0.57
Top
ic-3 0.59 0.43 0.75 0.38 0.66 0.59 0.65 0.5 0.62 0.5 0.61 0.52 0.48 0.64 0.55 0.55 0.4 0.51
Top 0.8
-4
Topic 0.59 0.52 0.62 0.49 0.9 0.52 0.49 0.42 0.6 0.57 0.59 0.54 0.6 0.67 0.36 0.39 0.28 0.49

ic-5 0.55 0.45 0.61 0.36 0.6 0.86 0.58 0.49 0.86 0.54 0.57 0.44 0.67 0.56 0.52 0.55 0.4 0.63
Top
Topic Vectors of Bhagavad Gita

-6
Topic 0.65 0.49 0.77 0.39 0.61 0.67 0.78 0.8 0.66 0.57 0.52 0.58 0.57 0.44 0.64 0.53 0.41 0.56 0.6
-7
Topic 0.74 0.46 0.63 0.51 0.69 0.55 0.51 0.45 0.51 0.48 0.5 0.74 0.49 0.56 0.47 0.47 0.31 0.41

ic-8 0.56 0.51 0.67 0.4 0.56 0.61 0.73 0.59 0.71 0.54 0.53 0.51 0.54 0.53 0.71 0.52 0.43 0.54
Top
ic-9 0.66 0.44 0.65 0.36 0.67 0.64 0.55 0.58 0.63 0.6 0.52 0.53 0.52 0.59 0.52 0.44 0.36 0.47 0.4
Top
0
ic-1 0.59 0.31 0.59 0.3 0.54 0.6 0.45 0.4 0.52 0.46 0.46 0.48 0.46 0.44 0.36 0.4 0.39 0.43
Top
1
ic-1 0.51 0.43 0.58 0.32 0.78 0.47 0.45 0.43 0.61 0.6 0.54 0.42 0.62 0.65 0.48 0.48 0.26 0.56
Top
12 0.47 0.27 0.2
ic- 0.5 0.24 0.53 0.43 0.4 0.29 0.52 0.47 0.51 0.36 0.45 0.55 0.33 0.37 0.38 0.43
Top
ic- 13 0.46 0.34 0.46 0.27 0.53 0.39 0.36 0.27 0.48 0.51 0.44 0.35 0.4 0.6 0.45 0.37 0.37 0.32
Top
4
ic-1 0.68 0.42 0.6 0.46 0.56 0.54 0.57 0.68 0.5 0.45 0.42 0.57 0.45 0.33 0.5 0.43 0.28 0.5
Top
0.0
ic-1 ic-2 ic-3 ic-4 ic-5 ic-6 ic-7 ic-8 ic-9 -10 -11 -12 -13 -14 -15 -16 -17 -18
Top Top Top Top Top Top Top Top Top Topic Topic Topic Topic Topic Topic Topic Topic Topic
Topic Vectors of Upanishads
Fig 6. Heatmap showing the similarity between different topics of Bhagavad Gita and Upanishads generated from a selected
approach (SBERT-UMAP-HDBSCAN).

Gita Upanishads
Topics of Gita Most Similar topics in Upanishads Similarity Score
Topic ID Topic ID
krishna,jayadratha,shraddha,ahamkara, sage,wisdom,devotee,sages,vishnu,mahabharata,
topic-1 topic-1 0.76
arjuna,ikshvaku,sankhya,ashvattha,kusha,vishnu devotees,samashrava,hindu,mahavakyas,theravada
selfless,selflessly,selfish,selfishly,desires, desires,happiness,eternal,selfless,beings,spiritual,
topic-2 topic-11 0.76
unkindness,desire,suffering,greed,themselves existence,spirituality,desire,joy,eternity,buddhism
worships,worship,devotion,devotees,eternal, eternal,divine,deity,eternity,lords,lord,everlasting,
topic-3 topic-3 0.75
myself,beings,eternity,spiritually,spiritual devotional,omnipotent,gods,beings,soul,beloved
meditation,meditate,spiritually,spiritual, meditation,meditating,meditates,meditate,meditated,
topic-4 topic-5 0.90
yoga,minds,asceticism,spirit,nirvana,wisdom minds,mind,spiritually,interiorize,enlightenment,spiritual
immortality,death,mortality,immortal,deathless, immortality,death,immortal,mortality,deathless,
topic-5 topic-6 0.86
eternity,eternal,dying,mortal,dead,mortals mortal,dying,mortals,eternity,deathlessness,eternal
gods,eternal,universe,beings,eternity,heavens, celestial,sun,heavens,earth,earthly,heaven,heavenly,
topic-6 topic-8 0.80
celestial,immortality,heavenly,divine,god luminous,sunrise,sky,universe,illumined,light,illumine
brahman,wisdom,devotees,brahma, sage,wisdom,devotee,sages,vishnu,mahabharata,devotees,
topic-7 topic-1 0.74
devotee,teachings,sages,worships,divine,devote samashrava,hindu,mahavakyas,theravada,hindus,buddhi
existence,universe,beings,eternal,nonexistence, universe,omnipotent,eternal,cosmos,eternity,beings,cosmic,
topic-8 topic-7 0.73
immortality,creatures,eternity,creature,cosmos immortal,gods,celestial,deity,beyondness,god,heavens
ignorance,ignorant,wisdom,delusions,darkness, meditation,meditating,meditates,meditate,meditated,
topic-9 topic-5 0.67
delusion,evils,intellects,eternal,asceticism minds,mind,spiritually,interiorize,enlightenment,spiritual
senses,sense,feeling,selflessly,selfless,selfishly, meditation,meditating,meditates,meditate,meditated,
topic-10 topic-5 0.78
feel,selfish,minds,oneself,themselves,perceive minds,mind,spiritually,interiorize,enlightenment,spiritual
enemy,enemies,conquer,defeat,fight,fighting, immortality,death,immortal,mortality,deathless,mortal,
topic-11 topic-6 0.60
conquered,battle,fought,nonviolence,dishonor dying,mortals,eternity,deathlessness,eternal,dead,deaths
forgiving,renunciation,fulfill,selfless,nonbeing, selfs,self,selfless,oneself,himself,themselves,selfish,ego,
topic-12 topic-14 0.55
unpleasant,selflessly,fulfilling,insatiable,indulging itself,egoism,yourself,independently,ourselves,autonomic
actions,act,action,acts,acting,inaction,selflessly, selfs,self,selfless,oneself,himself,themselves,selfish,ego,
topic-13 topic-14 0.60
selfless,themselves,unaffected,indifference,ignorance itself,egoism,yourself,independently,ourselves,autonomic
beings,spiritual,gods,divine,heavens,ocean, sage,wisdom,devotee,sages,vishnu,mahabharata,devotees,
topic-14 topic-1 0.68
spiritually,shudra,ahamkara,rudras,sacred,worships samashrava,hindu,mahavakyas,theravada,hindus,buddhi
Mean Similarity Score(AvgSim) 0.73
Table 5. Topics of the Bhagavad Gita(Eknath Easwaran ) with most similar topics from the Upanishads(Eknath Easwaran)

May 24, 2022 18/39

 1.0
ic-1 0.59 0.65 0.47 0.42 0.46 0.47 0.48 0.4 0.45 0.48 0.48 0.36 0.37 0.53 0.32 0.35 0.38 0.57
Top
ic-2 0.59 0.45 0.25 0.77 0.69 0.59 0.58 0.67 0.48 0.6 0.51 0.47 0.44 0.62 0.49 0.25 0.31 0.43
Top
ic-3 0.58 0.46 0.27 0.66 0.59 0.67 0.59 0.64 0.48 0.53 0.58 0.37 0.44 0.7 0.41 0.33 0.38 0.49
Top 0.8
-4
Topic 0.51 0.47 0.23 0.57 0.58 0.5 0.8 0.62 0.53 0.51 0.57 0.42 0.37 0.54 0.39 0.25 0.22 0.48

ic-5 0.66 0.5 0.24 0.6 0.89 0.62 0.54 0.59 0.48 0.64 0.58 0.46 0.5 0.77 0.67 0.33 0.34 0.46
Top
Topic Vectors of Bhagavad Gita

ic-6 0.82 0.52 0.34 0.55 0.64 0.77 0.58 0.53 0.61 0.64 0.65 0.46 0.55 0.7 0.53 0.54 0.46 0.62
Top 0.6
-7
Topic 0.55 0.61 0.36 0.56 0.57 0.58 0.58 0.55 0.54 0.5 0.67 0.43 0.45 0.65 0.34 0.32 0.39 0.53

ic-8 0.62 0.47 0.28 0.57 0.62 0.73 0.52 0.6 0.48 0.61 0.62 0.41 0.51 0.67 0.52 0.42 0.42 0.5
Top
-9 0.4
Topic 0.65 0.56 0.34 0.62 0.68 0.62 0.5 0.58 0.55 0.57 0.71 0.61 0.45 0.71 0.44 0.42 0.31 0.48

ic- 10 0.55 0.47 0.32 0.53 0.61 0.44 0.45 0.49 0.38 0.49 0.46 0.41 0.37 0.59 0.41 0.27 0.28 0.44
Top
ic- 11 0.55 0.44 0.16 0.62 0.56 0.55 0.6 0.64 0.62 0.57 0.57 0.48 0.42 0.53 0.5 0.28 0.2 0.43
Top
2 0.2
ic-1 0.42 0.43 0.27 0.64 0.55 0.44 0.47 0.53 0.36 0.45 0.38 0.39 0.36 0.52 0.41 0.16 0.22 0.35
Top
3
ic-1 0.37 0.4 0.19 0.56 0.46 0.48 0.4 0.54 0.43 0.44 0.52 0.46 0.36 0.49 0.34 0.18 0.2 0.27
Top
4
ic-1 0.76 0.57 0.34 0.45 0.52 0.59 0.54 0.42 0.57 0.54 0.54 0.35 0.48 0.58 0.41 0.53 0.44 0.61
Top
0.0
ic-1 ic-2 ic-3 ic-4 ic-5 ic-6 ic-7 ic-8 ic-9 -10 -11 -12 -13 -14 -15 -16 -17 -18
Top Top Top Top Top Top Top Top Top Topic Topic Topic Topic Topic Topic Topic Topic Topic
Topic Vectors of Upanishads
Fig 7. Heatmap showing the similarity between different topics of Bhagavad Gita(Eknath Easwaran) and the Ten Principal
Upanishads(Shri Purohit Swami) generated from a selected approach (SBERT-UMAP-HDBSCAN).

May 24, 2022 19/39

Fig 8. Visualization of the semantic space of The Bhagavad Gita and The Upanishads with topic labels.

May 24, 2022 20/39

Fig 9. Visualization of the Combined Semantic Space of The Bhagavad Gita and The Upanishads (PCA and UMAP).

May 24, 2022 21/39

 Gita Upanishads Similarity
Topics of Gita Most Similar topics in Upanishads
Topic ID Topic ID Score
krishna,jayadratha,shraddha,ahamkara,arjuna, sage,watnadewa,jaiwali,sages,told,riddle,heard,
topic-1 topic-2 0.65
ikshvaku,sankhya,ashvattha,kusha,vishnu butspirit,spoke,whoknows,thathe,speak,spirits
selfless,selflessly,selfish,selfishly,desires, himself,self,selfexistent,oneself,desires,eternal,
topic-2 topic-4 0.77
unkindness,desire,suffering,greed,themselves beings,desire,selfwilled,devotee,worship,sinful
worships,worship,devotion,devotees,eternal, immortality,immortal,immortals,heaven,eternal,
topic-3 topic-14 0.70
myself,beings,eternity,spiritually,spiritual heavenborn,heavens,heavenly,celestial,paradise
meditation,meditate,spiritually,spiritual,yoga, meditation,meditating,meditates,spiritual,spirit,
topic-4 topic-7 0.80
minds,asceticism,spirit,nirvana,wisdom meditated,meditate,meditations,spirits,butspirit
immortality,death,mortality,immortal,deathless, immortality,death,immortal,inlife,deathless,dead,
topic-5 topic-5 0.89
eternity,eternal,dying,mortal,dead,mortals mortality,life,dying,immortals,eternal,mortal,dies
gods,eternal,universe,beings,eternity,heavens, heavenborn,eternal,heavens,celestial,heaven,fire,heavenly,
topic-6 topic-1 0.82
celestial,immortality,heavenly,divine,god spirits,gods,ablaze,immortality,beings,spirit,immortal
brahman,wisdom,devotees,brahma,devotee, knowledge,whoknows,knowthe,wisdom,known,
topic-7 topic-11 0.67
teachings,sages,worships,divine,devote unknowable,knower,knowing,knew,knows,omniscient
existence,universe,beings,eternal,nonexistence, eternal,everything,beings,immortality,immortal,everybodys,
topic-8 topic-6 0.73
immortality,creatures,eternity,creature everybody,all,every,everyone,earthly,allcontaining,whatever
ignorance,ignorant,wisdom,delusions,darkness, knowledge,whoknows,knowthe,wisdom,known,
topic-9 topic-11 0.71
delusion,evils,intellects,eternal,asceticism unknowable,knower,knowing,knew,knows,omniscient
enemy,enemies,conquer,defeat,fight,fighting, immortality,death,immortal,inlife,deathless,dead,
topic-10 topic-5 0.61
conquered,battle,fought,nonviolence,dishonor mortality,life,dying,immortals,eternal,mortal,dies
senses,sense,feeling,selflessly,selfless,selfishly, self,oneself,selfevident,selfwilled,himself,selfdependent,
topic-11 topic-8 0.64
feel,selfish,minds,oneself,themselves,perceive selfcreator,selfexistent,selfborn,selfdepending,itself,selfinterest
forgiving,renunciation,fulfill,selfless,nonbeing, himself,self,selfexistent,oneself,desires,eternal,beings,desire,
topic-12 topic-4 0.64
unpleasant,selflessly,fulfilling,insatiable,indulging selfwilled,devotee,worship,sinful,worshipper,gods,godless
actions,act,action,acts,acting,inaction,selflessly, himself,self,selfexistent,oneself,desires,eternal,beings,desire,
topic-13 topic-4 0.56
selfless,themselves,unaffected,indifference,ignorance selfwilled,devotee,worship,sinful,worshipper,gods,godless
beings,spiritual,gods,divine,heavens,ocean,spiritually, heavenborn,eternal,heavens,celestial,heaven,fire,heavenly,
topic-14 topic-1 0.76
shudra,ahamkara,rudras,sacred,worships,devotees spirits,gods,ablaze,immortality,beings,burning,spirit,immortal
Mean Similarity Score(AvgSim) 0.71
Table 6. Topics of the Bhagavad Gita(Eknath Easwaran ) with most similar topics from the Ten Principal Upanishads(Shri
Purohit Swami & W.B. Yeats)

Vedas # Upanishads Upanishads

Prasna,Mundaka,Mandukya, Atahrvasiras,Atharvasikha,Brihajjabala,Devi
Nrisimhatapini,Naradaparivrajaka,Sita,Sarabha,Tripadvibhuti-Mahanarayana,
Atharva-Veda 31 Ramarahasya,Ramatapini,Sandilya,Paramahamsaparivrajaka,Annapurna,Surya,
Atma,Pasupatabrahmana,Parabrahma,Tripuratapini,Bhavana,Bhasmajabala,
Ganapati,Mahavakya,Gopalatapini,Krishna,Hayagriva,Dattatreya,Garuda
Kathavalli,Taittiriyaka,Brahma,Kaivalya,Svetasvatara, Garbha,Varaha,Akshi
Narayana,Amritabindu,Amritanada,Kalagnirudra,Kshurika,Sarvasara
Krishna-Yajur-Veda 32 Sukarahasya,Tejobindu,Dhyanabindu,Brahmavidya,Yogatattva,Dakshinamurti,
Skanda,Sariraka,Yogasikha,Ekakshara,Avadhuta,Katharudra,Rudrahridaya,
Yoga-kundalini,Panchabrahma,Pranagnihotra,Kalisamtarana,Sarasvatirahasya
Isavasya,Brihadaranyaka,Jabala,Hamsa,Paramahamsa,Paingala,Bhiksu,Tarasara
Sukla-Yajur-Veda 19 Mantrika,Niralamba,Trisikhibrahmana,Mandalabrahmana,Turiyatita,,Subala
Satyayani,Muktika,Advayataraka,Adhyatma,Yajnavalkya
Kena,Chandogya,Aruni,Maitrayani,Maitreya,Vajrasuchika,Jabali
Sama Veda 16 Rudrakshajabala, Yogachudamani,Vasudeva,Savitri,Darsana
Mahat,Sannyasa,Avyakta,Kundika
Aitareya,Kaushitakibrahmana,Nadabindu,Atmabodha,Nirvana,Mudgala,
Rig Veda 10
Akshamalika ,Tripura,Saubhagyalakshmi, Bahvricha
Table 7. Classification of Upanishads based on original Vedas it is derived from

Upanishads that fall under 4 different categories identified by the four Vedas [12] ( Rig
Veda, Samar Veda, Yajur Veda, Artha Veda) which are known as the founding texts of
Hinduism. The Rig Veda is the oldest Hindu texts written in ancient Sanskrit and
believed to be remembered orally from guru-student tradition of mantra-recital [162]
thousands of years before being written down [11]. It has been difficult to translated
and also understand significance of certain aspects of the Vedas since it has been

May 24, 2022 22/39

 Labels
Topic-1 12
Topic-2
Topic-3 11
Topic-4
Topic-5 10

UMAP Embedding (dim = 2)

Topic-6
Topic-7 9
Topic-8
Topic-9
Topic-10 8
7
6
5
4
2 3 4 5 6 7 8 9
UMAP Embedding (dim = 1)
Fig 10. Visualisation of different topics of 108 Upanishads

written in ancient Sanskrit in verse with symbolism [163]. The Upanishads are known as
the texts that explain the philosophy of the Vedas and also known as the concluding
chapters that have been added to the four Vedas [164]. Table 7 gives information about
how the 108 Upanishads have been grouped according to their historical relevance to
the respective Vedas.
Figure 11 presents visualization of the semantic space of different parts (divided by 4
Vedas as shown in Table 7) of 108 Upanishads.

4 Discussion
The high level of semantic and topic similarity between the Bhagavad Gita and the
different sets of the Upanishads by the respective authors is not surprising. It verifies
well known thematic similarities as pointed out by Hindu scholars such as Swami
Vivekananda [165] and western scholars [14]. Bhagavad Gita is well known as the
central text of Hinduism that summaries the rest of the Vedic corpus. The Bhagavad
Gita is a conversation between Lord Krishna and Arjuna in a situation where Arjuna
has to go to war. The Bhagavad Gita is a chapter from the Mahabharata that uses a
conflicting event to summarize philosophy of the Upanishads and the Vedic corpus. The
Mahabharata is one of the oldest and longest texts written in verse form in Sanskrit
which describes a historical event (118,087 sentences, 2,858,609 words) [166]. We note
that most of Hindu ancient texts have been written in verse so that it can be sung and
remembered through an oral tradition in an absence of a writing system.
The goal of Lord Krishna was to motivate Arjuna to do his duty (karma) and go to
war to protect ethical standards (dharma) in the society. Krishna, in the Bhagavad Gita
begins by renouncing his duties as a warrior. We note that the Mahabharata war, is
believed to take place after the Vedas and the Upanishads were composed. Note that by
composition, it does not mean that these texts were written, they were sung and verses
became key mantras that were remembered through a guru-student tradition for
thousands of years. There are accounts where the Vedas have been mentioned in the
Mahabharata. Hence, Krishna is known as a student of the Vedic corpus which also

May 24, 2022 23/39

Fig 11. Visualization of the semantic space of different parts (based on 4 Vedas) of 108 Upanishads.

May 24, 2022 24/39

 refers to the entire library of Hindu science, literature, history and philosophy.
Therefore, the topics in the Upanishads were well known by Lord Krishna and he may
have merely used some of the themes to highlight about themes of duty, ethics (dharma)
and work (karma) in order to motivate Arjuna to do his duty at the time of need,
otherwise, his side (Pandava) would lose the war to the opposition (Kaurava). The
Mahabharata war has blood relives on opposing sides of the war battleground known as
Kurushetra and hence it was difficult for Arjuna to make a decision either to fight for
dharma or renounce his duties and become a yogi (mystic).
Table 5 further compares the topics of the Bhagavad Gita with the Upanishads. We
can observe that each of the topic encapsulate some of the ideas expressed in selected
verses shown in Tables 7 and 8. If a topic of the Gita and the Upanishads have very
high similarity, this represents the fact that the ideas encapsulated by the topics of the
Gita and the Upanishads are almost same. In Table 5, we can observe that topic 4 of
the Bhagavad Gita and topic-5 of the Upanishads have a similarity of 90%, this can be
seen from the topics also they are representing the similar themes that are related to
the ideas of meditation, yoga and spirituality. Similarly, we observe that topic-5 of Gita
have a similarity score of 86% when compared with topic-6 of the Upanishads. Here, we
can also observe that both topics encapsulate similar ideas of death, mortality and
immortality. Similar ideas can be observed in Table 6 as well the topics of the Bhagavad
Gita is compared with the topics of the Upanishads.
Figure 8 depicts a representation of the semantic space of the Bhagavad Gita and
the Upanishads with topic labels. It represents the lower dimensional embedding of the
very high dimensional document vectors. In Figure 8, we represented only 10 topics in
order to retain the clarity of the diagram. Figure 9 shows the UMAP and PCA
embedding of the entire document. In order to generate this plot, we first created the
embeddings of each documents and then reduced the embedding to 2D by using PCA
and UMAP. After reducing the dimension, we assigned the labels (Gita, and the
Upanishads) based on the corpus. Figure 9 shows that low-dimensional embeddings
reveals very clear overlaps across the documents.
Even with the presence of translation bias by considering two different translations
of the Upanishads, our results demonstrate a very high resemblance between the topics
of these two texts, with a mean cosine similarity of more than 70% between the topics
of the Bhagavad Gita and those of the Ten Principal Upanishads by Shi Purohit Swami
and W.B Yeats. Eight of the fourteen topics extracted from the Bhagavad Gita have a
cosine similarity of more than 70% with the topics in the 10 Principal Upanishads, which
can also be seen in the table 6, and 3 of them have a similarity of more than 80%. When
considering the translation of both texts by same author as in the case of the Bhagavad
Gita [109] and the Upanishads [110], we see that average similarity increase to 73% with
9 out of 14 topics having more than 70% similarity and 3 of them having a similarity of
more than 80%. We also found that topics generated by the BERT based models show
very high coherence as compared to that of the LDA. Our best performing model gives
a coherence score of 73% on The Bhagavad Gita [109], 69% on The Upanishads [110],
73% on The ten Principal Upanishads [111] and 66% on the 108 Upanishads.
The major limitation is due to the translation bias, which is not present when we
take the same translator - this is why we chose the Upanishads and Bhagavad Gita by
Eknath Easwaren in order to limit the bias. However, if we consider the complete 108
Upanishads, such translation bias remain. Moreover, the style and language of the
translations not only depend on the translator but on the era. In the case of the 108
Upanishads, a group of translators have contributed which creates further biases.
However, in terms of topics uncovered, we find a consistent set of topics that well alight
with the respective texts, after manually verifying it.
Further extension can be done by taking the other translations into consideration.

May 24, 2022 25/39

 The Ten Principal Upanishads [111] published in 1938, was translated by the Irish poet
William Butler Yeats and Hindu guru Shri Purohit Swami. The translation process
occurred between the two authors throughout the 1930s and this book can has been
claimed as one of the final works of William Butler Yeats [167]. We note that Shri
Purohit Swami has also translated the Bhagavad Gita, hence this would be a good
companion with Eknath Eashwaren for the respective texts. These extensions could
help in refining the proposed framework.
Moreover, in terms of the mythological texts and epics, there are various texts such
as the Vishnu Purana, Shiv Purana out of the 18 different Puranas that have underlying
topics that are similar. In this study, we focused on philosophical texts, while in future
studies, there can be scope for topic modelling from selected texts in the Puranas. The
framework can also be used to study texts from other religions, along with n
non-religious and non-philosophical texts. Furthermore, it can be used to study themes
of modern poetry, writers, songs and be used to compare different religions and time
frames, i.e how the themes changes over different centuries, prior to or after a war or a
pandemic (such as the COVID-19).
We note that there exists specialised BERT pre-trained models such as those for
medicine and law [168–173], but there is nothing yet developed for philosophy. Hindu
philosophy is distinct and has terms and ideas that are not present in other
philosophical areas (such as western philosophy). Hence, we need specialised pre-trained
BERT model for Hindu philosophy which can provide better predictions in related
language tasks since it will have better knowledge-base. This work can further be
improved using language models for the native Sanskrit text. We intend to explore topic
models after building BERT-like language models for Indian philosophical literature
written in Sanskrit.
We note that our previous work focused on semantic and sentiment analysis of the
Bhagavad Gita translations [174]. Augmenting semantic and sentiment analysis to our
proposed topic modelling framework can provide more insights to the meaning behind
the philosophical verses. We plan to build our models in a similar fashion and
investigate their variations for texts in three different languages: Hindi, English, and
Sanskrit. Finally, post verification study is needed where Sanskrit expert and Hindu
philosophers can study the topics uncovered by the proposed framework.
The Bhagavad Gita and the Upanishads are considerably large texts in the content
of religious and philosophical texts. However, the proposed framework can be used for
larger corpus such as modelling overlapping topics around the Mahabharata and the
Puranas, which are texts that are magnitudes larger than the ones considered in this
study. However, we note that the Bhagavad Gita and Upanishads, although smaller in
size are more condensed in philosophy while the Mahabharata is an epic poem.
In future work, there can be a detailed study of the topics uncovered with a
discussion of related texts in Vedic studies that relate to morphology, lexicography,
grammar (patterns in sentences), meter (lengthy sentences), and phonology (sound
system), etc. Furthermore, we need to create processed benchmark text datasets for
Indian languages that can benefit NLP applications associated with Indian languages.

5 Conclusion and Future work

We presented a topic modeling framework for Hindu philosophy using state-of-art
deep-learning based models. The use of such technique for studying Hindu texts is
relatively novel; however, computational and statistical approaches have been used in
the past. The major goal of the study was to link the topics from the Upanishads with
the Bhagavad Gita.
The representation of the low-dimensional embeddings presented in this work reveals

May 24, 2022 26/39

 a lot of overlap between the Upanishads and the Bhagavad Gita’s topics, which adds to
our objective of demonstrating the Bhagavad Gita’s relationship with the Upanishads.
Given the importance of religious literature to a community, employing computational
models to verify any of its old and traditional philosophical principles demonstrates the
scientific nature of the literature and religion. Despite the fact that the idea of the Gita
being the essential extract of the Upanishads has been written and researched in ancient
Indian philosophical literature for generations, no attempt has ever been made to
substantiate this facts using computational and scientific methodologies. Our research
presents a novel way for applying modern deep learning-based methods to a
centuries-old philosophical narratives.

Data and Code

Python-based open source code and data can be found here 4 .

Author contributions statement

R. Chandra devised the project with the main conceptual ideas and and contributed to
overall writing, literature review and discussion of results. M. Ranjan provided
implementation and experimentation and further contributed in results visualisation
and analysis along with writing.

Acknowledgement
We thank Shweta Bindal from Indian Institute of Technology - Guwathi for contributing
to discussions about the workflow used in this work.

Appendix
References
1. Meister C. Introducing philosophy of religion. Routledge; 2009.
2. Reese WL. Dictionary of philosophy and religion: Eastern and Western thought.
1996;.

3. Bernard T. Hindu philosophy. Motilal Banarsidass Publ.; 1999.

4. Saksena SK. Nature of consciousness in Hindu philosophy. University of London,
School of Oriental and African Studies (United Kingdom); 1939.
5. Chaudhuri H. The concept of Brahman in Hindu philosophy. Philosophy East
and West. 1954;4(1):47–66.
6. Roy K. Just and unjust war in Hindu philosophy. Journal of Military Ethics.
2007;6(3):232–245.
7. Reichenbach B. The law of karma: A philosophical study. Springer; 1990.

8. Mulla ZR, Krishnan VR. Karma Yoga: A conceptualization and validation of the
Indian philosophy of work. Journal of Indian Psychology. 2006;24(1/2):26–43.
4 https://github.com/sydney-machine-learning/topicmodelling_vedictexts

May 24, 2022 27/39

Fig 12. Topics of Bhagavad Gita and the most relevant documents(Model: USE-HDBSCAN-UMAP).

May 24, 2022 28/39

Fig 13. Topics of the Ten Principal Upanishads and some of their relevant documents(Model: USE-HDBSCAN-UMAP).

May 24, 2022 29/39

 9. Dasgupta S. A history of Indian philosophy. vol. 2. Motilal Banarsidass Publ.;
1975.
10. Radhakrishnan S, Radhakrishnan S. Indian philosophy. vol. 2. Allen & Unwin
London; 1929.
11. Staal F. Discovering the Vedas: origins, mantras, rituals, insights. Penguin
Books India; 2008.
12. Witzel M. Vedas and upanisads. The blackwell companion to hinduism. 2003; p.
68–101.
13. Radhakrishnan S. The Vedanta philosophy and the doctrine of Maya. The
International Journal of Ethics. 1914;24(4):431–451.
14. Torwesten H, Rosset L. Vedanta: Heart of Hinduism. Grove Press; 1991.
15. Prabhu HA, Bhat P. Mind and consciousness in yoga–Vedanta: A comparative
analysis with western psychological concepts. Indian journal of psychiatry.
2013;55(Suppl 2):S182.
16. Malinar A. Narrating Sām
. khya Philosophy: Bhı̄s.ma, Janaka and Pañcaśikha at”
Mahābhārata” 12.211—12. Journal of Indian Philosophy. 2017; p. 609–649.
17. Chakrabarti KK. Classical Indian philosophy of mind: The Nyaya dualist
tradition. SUNY Press; 1999.
18. Chatterji JC. The Hindu Realism: Being an Introduction to the Metaphysics of
the Nyaya-Vaisheshika System of Philosophy. Jain Publishing Company; 2007.
19. Arnold D. Of Intrinsic Validity: A Study on the Relevance of Pūrva Mı̄mām
. sā.
Philosophy East and West. 2001; p. 26–53.
20. Maas PA. A Concise Historiography of Classical Yoga Philosophy. Periodization
and historiography of Indian philosophy. 2013;37:53–90.
21. Long JD. Jaina Philosophy. In: The Oxford Handbook of World Philosophy;
2011.
22. Coomaraswamy AK. Hinduism and Buddhism. Golden Elixir Press; 2011.
23. Gellner DN. The anthropology of Buddhism and Hinduism: Weberian themes.
Oxford University Press; 2001.
24. Bhattacarya R. Studies on the Carvaka/Lokayata. Studies on the
Carvaka/Lokayata. 2009; p. 0–0.
25. Barua BM. Ājı̄vika-what it Means. Annals of the Bhandarkar Oriental Research
Institute. 1926;8(2):183–188.
26. Renard P. HISTORICAL BIBLIOGRAPHY OF Upanishads IN
TRANSLATION. Journal of Indian Philosophy. 1995;23(2):223–246.
27. Marlow A. Hinduism and Buddhism in Greek philosophy. Philosophy East and
West. 1954;4(1):35–45.
28. Rajagopalachari C. Mahabharata. vol. 1. Diamond Pocket Books (P) Ltd.; 1970.
29. Hiltebeitel A. Ritual of Battle, The: Krishna in the Mahabharata. SUNY Press;
1976.

May 24, 2022 30/39

 30. Gandhi M. The bhagavad gita according to Gandhi. North Atlantic Books; 2010.
31. Brown DM. The Philosophy of Bal Gangadhar Tilak: Karma vs. Jnana in the
Gita Rahasya. The Journal of Asian Studies. 1958;17(2):197–206.
32. Phillips S. Yoga, karma, and rebirth: A brief history and philosophy. Columbia
University Press; 2009.
33. Muniapan B, Satpathy B. The ‘dharma’and ‘karma’of CSR from the
Bhagavad-Gita. Journal of Human Values. 2013;19(2):173–187.

34. Scharfstein BA. A comparative history of world philosophy: from the

Upanishads to Kant. SUNY Press; 1998.
35. Cross S. Turning to the East: How the Upanishads reached the West. India
International Centre Quarterly. 1998;25(2/3):123–129.

36. Indurkhya N, Damerau FJ. Handbook of natural language processing. vol. 2.

CRC Press; 2010.
37. Manning C, Schutze H. Foundations of statistical natural language processing.
MIT press; 1999.

38. Chowdhury GG. Natural language processing. Annual review of information

science and technology. 2003;37(1):51–89.
39. Liu B, Zhang L. A survey of opinion mining and sentiment analysis. In: Mining
text data. Springer; 2012. p. 415–463.
40. Medhat W, Hassan A, Korashy H. Sentiment analysis algorithms and
applications: A survey. Ain Shams engineering journal. 2014;5(4):1093–1113.
41. Hussein DMEDM. A survey on sentiment analysis challenges. Journal of King
Saud University-Engineering Sciences. 2018;30(4):330–338.
42. Hochreiter S, Schmidhuber J. Long short-term memory. Neural computation.
1997;9(8):1735–1780.

43. Wang Y, Huang M, Zhu X, Zhao L. Attention-based LSTM for aspect-level

sentiment classification. In: Proceedings of the 2016 conference on empirical
methods in natural language processing; 2016. p. 606–615.
44. Vaswani A, Shazeer N, Parmar N, Uszkoreit J, Jones L, Gomez AN, et al.
Attention is all you need. In: Advances in neural information processing
systems; 2017. p. 5998–6008.
45. Wolf T, Chaumond J, Debut L, Sanh V, Delangue C, Moi A, et al.
Transformers: State-of-the-art natural language processing. In: Proceedings of
the 2020 Conference on Empirical Methods in Natural Language Processing:
System Demonstrations; 2020. p. 38–45.
46. Devlin J, Chang MW, Lee K, Toutanova K. Bert: Pre-training of deep
bidirectional transformers for language understanding. arXiv preprint
arXiv:181004805. 2018;.
47. Blei DM, Ng AY, Jordan MI. Latent dirichlet allocation. the Journal of machine
Learning research. 2003;3:993–1022.

May 24, 2022 31/39

 48. Reimers N, Gurevych I. Sentence-bert: Sentence embeddings using siamese
bert-networks. arXiv preprint arXiv:190810084. 2019;.

49. Silveira R, Fernandes C, Neto JAM, Furtado V, Pimentel Filho JE. Topic
Modelling of Legal Documents via LEGAL-BERT. Proceedings http://ceur-ws
org ISSN. 2021;1613:0073.
50. Peinelt N, Nguyen D, Liakata M. tBERT: Topic models and BERT joining
forces for semantic similarity detection. In: Proceedings of the 58th Annual
Meeting of the Association for Computational Linguistics; 2020. p. 7047–7055.
51. Grootendorst M. BERTopic: leveraging BERT and c-TF-IDF to create easily
interpretable topics (2020). URL https://doi org/105281/zenodo;4381785.
52. Angelov D. Top2vec: Distributed representations of topics. arXiv preprint
arXiv:200809470. 2020;.

53. Sia S, Dalmia A, Mielke SJ. Tired of Topic Models? Clusters of Pretrained
Word Embeddings Make for Fast and Good Topics too! arXiv preprint
arXiv:200414914. 2020;.
54. Thompson L, Mimno D. Topic modeling with contextualized word
representation clusters. arXiv preprint arXiv:201012626. 2020;.
55. Grootendorst M. BERTopic: Neural topic modeling with a class-based TF-IDF
procedure. arXiv preprint arXiv:220305794. 2022;.
56. Glazkova A. Identifying topics of scientific articles with BERT-based approaches
and topic modeling. In: Pacific-Asia Conference on Knowledge Discovery and
Data Mining. Springer; 2021. p. 98–105.
57. Scott M. Religious Language. In: Zalta EN, editor. The Stanford Encyclopedia
of Philosophy. Winter 2017 ed. Metaphysics Research Lab, Stanford University;
2017.
58. Pandharipande RV. Language of Religion: What does it inform the field of
Linguistics?;.
59. Keane W. Religious language. Annual review of anthropology. 1997;26(1):47–71.
60. Downes W. Linguistics and the Scientific Study of Religion. Religion, language,
and the human mind. 2018; p. 89.

61. Theodor I. Exploring the Bhagavad Gita: Philosophy, structure and meaning.
Routledge; 2016.
62. Stein D. Multi-Word Expressions in the Spanish Bhagavad Gita, Extracted with
Local Grammars Based on Semantic Classes. In: LREC 2012 Workshop
Language Resources and Evaluation for Religious Texts (LRE-Rel); 2012. p.
88–94.
63. Rajandran K. From matter to spirit: Metaphors of enlightenment in
Bhagavad-Gita. Journal of Language Studies. 2017;17(2):163–176.
64. Rajput NK, Ahuja B, Riyal MK. A statistical probe into the word frequency
and length distributions prevalent in the translations of Bhagavad Gita.
Pramana. 2019;92(4):1–6.

May 24, 2022 32/39

 65. Dewi NLPS. Metaphors of Life in Bhagavad Gita As It Is.
Humanis;22(2):503–508.

66. Bhawuk DP. Anchoring cognition, emotion, and behavior in desire: A model
from the Bhagavad-Gita. Handbook of Indian psychology. 2008; p. 390–413.
67. Chandra R, Kulkarni V. Semantic and sentiment analysis of the Bhagavad Gita
translations using BERT-based language models. arXiv preprint arXiv:xxx.
2022;.

68. Haas GC. Recurrent and parallel passages in the principal Upanishads and the
Bhagavad-Gı̄tā. Journal of the American Oriental Society. 1922; p. 1–43.
69. Greff K, Srivastava RK, Koutnı́k J, Steunebrink BR, Schmidhuber J. LSTM: A
search space odyssey. IEEE transactions on neural networks and learning
systems. 2016;28(10):2222–2232.

70. Malhotra P, Ramakrishnan A, Anand G, Vig L, Agarwal P, Shroff G.

LSTM-based encoder-decoder for multi-sensor anomaly detection. arXiv
preprint arXiv:160700148. 2016;.
71. Beltagy I, Peters ME, Cohan A. Longformer: The long-document transformer.
arXiv preprint arXiv:200405150. 2020;.

72. Imamura K, Sumita E. Recycling a pre-trained BERT encoder for neural

machine translation. In: Proceedings of the 3rd Workshop on Neural Generation
and Translation; 2019. p. 23–31.
73. Yang J, Wang M, Zhou H, Zhao C, Zhang W, Yu Y, et al. Towards making the
most of bert in neural machine translation. In: Proceedings of the AAAI
Conference on Artificial Intelligence. vol. 34; 2020. p. 9378–9385.
74. Zhu J, Xia Y, Wu L, He D, Qin T, Zhou W, et al. Incorporating BERT into
Neural Machine Translation. In: International Conference on Learning
Representations; 2020.

75. Clinchant S, Jung KW, Nikoulina V. On the use of BERT for neural machine
translation. arXiv preprint arXiv:190912744. 2019;.
76. Shavarani HS, Sarkar A. Better Neural Machine Translation by Extracting
Linguistic Information from BERT. arXiv preprint arXiv:210402831. 2021;.

77. Esteva A, Kale A, Paulus R, Hashimoto K, Yin W, Radev D, et al. COVID-19

information retrieval with deep-learning based semantic search, question
answering, and abstractive summarization. NPJ digital medicine. 2021;4(1):1–9.
78. Khazaeli S, Punuru J, Morris C, Sharma S, Staub B, Cole M, et al. A Free
Format Legal Question Answering System. In: Proceedings of the Natural Legal
Language Processing Workshop 2021; 2021. p. 107–113.

79. Geva M, Khashabi D, Segal E, Khot T, Roth D, Berant J. Did Aristotle Use a
Laptop? A Question Answering Benchmark with Implicit Reasoning Strategies.
Transactions of the Association for Computational Linguistics. 2021;9:346–361.
80. Ozyurt IB, Bandrowski A, Grethe JS. Bio-AnswerFinder: a system to find
answers to questions from biomedical texts. Database. 2020;2020.

May 24, 2022 33/39

 81. Lamm M, Palomaki J, Alberti C, Andor D, Choi E, Soares LB, et al. QED: A
Framework and Dataset for Explanations in Question Answering. Transactions
of the Association for Computational Linguistics. 2021;9:790–806.
82. Kwiatkowski T, Palomaki J, Redfield O, Collins M, Parikh A, Alberti C, et al.
Natural questions: a benchmark for question answering research. Transactions
of the Association for Computational Linguistics. 2019;7:453–466.
83. Hoang M, Bihorac OA, Rouces J. Aspect-based sentiment analysis using bert.
In: Proceedings of the 22nd Nordic Conference on Computational Linguistics;
2019. p. 187–196.
84. Li X, Bing L, Zhang W, Lam W. Exploiting BERT for end-to-end aspect-based
sentiment analysis. arXiv preprint arXiv:191000883. 2019;.
85. Wu Z, Ong DC. Context-guided bert for targeted aspect-based sentiment
analysis. Association for the Advancement of Artificial Intelligence. 2020; p. 1–9.
86. Yang K, Xu H, Gao K. CM-BERT: Cross-Modal BERT for Text-Audio
Sentiment Analysis. In: Proceedings of the 28th ACM International Conference
on Multimedia; 2020. p. 521–528.
87. Du C, Sun H, Wang J, Qi Q, Liao J. Adversarial and domain-aware BERT for
cross-domain sentiment analysis. In: Proceedings of the 58th annual meeting of
the Association for Computational Linguistics; 2020. p. 4019–4028.
88. Liu Y, Ott M, Goyal N, Du J, Joshi M, Chen D, et al. Roberta: A robustly
optimized bert pretraining approach. arXiv preprint arXiv:190711692. 2019;.
89. Lan Z, Chen M, Goodman S, Gimpel K, Sharma P, Soricut R. Albert: A lite
bert for self-supervised learning of language representations. arXiv preprint
arXiv:190911942. 2019;.
90. Chandra R, Krishna A. COVID-19 sentiment analysis via deep learning during
the rise of novel cases. PloS One. 2021;16(8):e0255615.
91. Chandra R, Saini R. Biden vs Trump: Modeling US General Elections Using
BERT Language Model. IEEE Access. 2021;9:128494–128505.
92. Mikolov T, Chen K, Corrado G, Dean J. Efficient estimation of word
representations in vector space. arXiv preprint arXiv:13013781. 2013;.
93. Pennington J, Socher R, Manning CD. GloVe: Global Vectors for Word
Representation. In: Empirical Methods in Natural Language Processing
(EMNLP); 2014. p. 1532–1543. Available from:
http://www.aclweb.org/anthology/D14-1162.
94. Cer D, Yang Y, Kong Sy, Hua N, Limtiaco N, John RS, et al. Universal
sentence encoder. arXiv preprint arXiv:180311175. 2018;.
95. Iyyer M, Manjunatha V, Boyd-Graber J, Daumé III H. Deep unordered
composition rivals syntactic methods for text classification. In: Proceedings of
the 53rd annual meeting of the association for computational linguistics and the
7th international joint conference on natural language processing (volume 1:
Long papers); 2015. p. 1681–1691.
96. Schroff F, Kalenichenko D, Philbin J. Facenet: A unified embedding for face
recognition and clustering. In: Proceedings of the IEEE conference on computer
vision and pattern recognition; 2015. p. 815–823.

May 24, 2022 34/39

 97. Xu D, Tian Y. A comprehensive survey of clustering algorithms. Annals of Data
Science. 2015;2(2):165–193.

98. Steinley D. K-means clustering: a half-century synthesis. British Journal of

Mathematical and Statistical Psychology. 2006;59(1):1–34.
99. Gowda KC, Krishna G. Agglomerative clustering using the concept of mutual
nearest neighbourhood. Pattern recognition. 1978;10(2):105–112.
100. Ester M, Kriegel HP, Sander J, Xu X, et al. A density-based algorithm for
discovering clusters in large spatial databases with noise. In: kdd. vol. 96; 1996.
p. 226–231.
101. Reynolds DA. Gaussian mixture models. Encyclopedia of biometrics.
2009;741(659-663).
102. Lloyd S. Least squares quantization in PCM. IEEE transactions on information
theory. 1982;28(2):129–137.
103. Campello RJ, Moulavi D, Sander J. Density-based clustering based on
hierarchical density estimates. In: Pacific-Asia conference on knowledge
discovery and data mining. Springer; 2013. p. 160–172.

104. McInnes L, Healy J. Accelerated hierarchical density based clustering. In: 2017
IEEE International Conference on Data Mining Workshops (ICDMW). IEEE;
2017. p. 33–42.
105. McInnes L, Healy J, Melville J. Umap: Uniform manifold approximation and
projection for dimension reduction. arXiv preprint arXiv:180203426. 2018;.

106. Van der Maaten L, Hinton G. Visualizing data using t-SNE. Journal of machine
learning research. 2008;9(11).
107. Wold S, Esbensen K, Geladi P. Principal component analysis. Chemometrics
and intelligent laboratory systems. 1987;2(1-3):37–52.

108. Ding J, Jin W. Overlaplda: A generative approach for literature-based discovery.

In: 2019 IEEE International Conference on Bioinformatics and Biomedicine
(BIBM). IEEE; 2019. p. 362–369.
109. Easwaran E. The Bhagavad Gita. Tomales; 2007.
110. Easwaran E. The Upanishads. Nilgiri Press Tomales, CA; 1987.

111. Swami SP, Yeats WB. The Ten Principal Upanishads. Rupa Publications India
Pvt. Ltd.; 2012. Available from:
https://books.google.co.in/books?id=f1KeAwAAQBAJ.
112. Swami SP. The Holy Geeta. Jaico Publishing House; 1935.

113. The International Gita Society. 108 Upanishads;.

https://gita-society.com/wp-content/uploads/PDF/108upanishads.pdf.
114. Thapar R. War in the Mahabharata. PMLA/Publications of the Modern
Language Association of America. 2009;124(5):1830–1833.

115. Witz KG. The supreme wisdom of the Upanis.ads: an introduction. Motilal
Banarsidass Publ.; 1998.

May 24, 2022 35/39

 116. Doan TN, Hoang TA. Benchmarking Neural Topic Models: An Empirical Study.
In: Findings of the Association for Computational Linguistics: ACL-IJCNLP
2021; 2021. p. 4363–4368.

117. Lisena P, Harrando I, Kandakji O, Troncy R. ToModAPI: A Topic Modeling

API to Train, Use and Compare Topic Models. In: Proceedings of Second
Workshop for NLP Open Source Software (NLP-OSS); 2020. p. 132–140.
118. Lang K. The 20 news groups data set. http://people csail mit
edu/jrennie/20Newsgroups/. 2008;.

119. Yin W, Hay J, Roth D. Benchmarking zero-shot text classification: Datasets,

evaluation and entailment approach. arXiv preprint arXiv:190900161. 2019;.
120. Zhang X, Zhao J, LeCun Y. Character-level convolutional networks for text
classification. Advances in neural information processing systems.
2015;28:649–657.
121. Ueda N, Saito K. Parametric mixture models for multi-labeled text. In:
Advances in neural information processing systems; 2003. p. 737–744.
122. Hoang TA, Vo KD, Nejdl W. W2e: A worldwide-event benchmark dataset for
topic detection and tracking. In: Proceedings of the 27th ACM International
Conference on Information and Knowledge Management; 2018. p. 1847–1850.
123. Jónsson E, Stolee J. An evaluation of topic modelling techniques for twitter.
University of Toronto. 2015;.
124. Yan X, Guo J, Lan Y, Cheng X. A biterm topic model for short texts. In:
Proceedings of the 22nd international conference on World Wide Web; 2013. p.
1445–1456.
125. Sridhar VKR. Unsupervised topic modeling for short texts using distributed
representations of words. In: Proceedings of the 1st workshop on vector space
modeling for natural language processing; 2015. p. 192–200.

126. Murakami A, Thompson P, Hunston S, Vajn D. ‘What is this corpus about?’:

using topic modelling to explore a specialised corpus. Corpora.
2017;12(2):243–277.
127. Sanh V, Debut L, Chaumond J, Wolf T. DistilBERT, a distilled version of
BERT: smaller, faster, cheaper and lighter. arXiv preprint arXiv:191001108.
2019;.
128. Chinmayananda S. The Holy Geeta. Central Chinmaya Mission Trust; 1996.
129. Yogananda P. God talks with Arjuna. Diamond Pocket Books (P) Ltd.; 2007.
130. Bera RK. Oppenheimer Quotes the Bhagavad Gita. Available at SSRN 3215156.
2018;.
131. Sidorov G, Velasquez F, Stamatatos E, Gelbukh A, Chanona-Hernández L.
Syntactic n-grams as machine learning features for natural language processing.
Expert Systems with Applications. 2014;41(3):853–860.
132. Stubenberg L. Consciousness and qualia. J. Benjamins Pub.; 1998.

133. Indich WM. Consciousness in advaita vedanta. Motilal Banarsidass Publ.; 1995.

May 24, 2022 36/39

 134. Woodhouse MB. Consciousness and Brahman-Atman. The Monist.
1978;61(1):109–124.

135. Ackerman RW. The Debate of the Body and the Soul and Parochial
Christianity. Speculum. 1962;37(4):541–565.
136. Sharma S. Corporate Gita: lessons for management, administration and
leadership. Journal of Human Values. 1999;5(2):103–123.
137. Nayak AK. Effective leadership traits from Bhagavad Gita. International
Journal of Indian Culture and Business Management. 2018;16(1):1–18.
138. Reddy M. Psychotherapy-insights from bhagavad gita. Indian journal of
psychological medicine. 2012;34(1):100–104.
139. Chang J, Gerrish S, Wang C, Boyd-Graber JL, Blei DM. Reading tea leaves:
How humans interpret topic models. In: Advances in neural information
processing systems; 2009. p. 288–296.
140. Syed S, Spruit M. Full-text or abstract? examining topic coherence scores using
latent dirichlet allocation. In: 2017 IEEE International conference on data
science and advanced analytics (DSAA). IEEE; 2017. p. 165–174.

141. Morstatter F, Liu H. In search of coherence and consensus: measuring the

interpretability of statistical topics. Journal of Machine Learning Research.
2018;18(169):1–32.
142. Newman D, Lau JH, Grieser K, Baldwin T. Automatic evaluation of topic
coherence. In: Human language technologies: The 2010 annual conference of the
North American chapter of the association for computational linguistics; 2010. p.
100–108.
143. Newman D, Noh Y, Talley E, Karimi S, Baldwin T. Evaluating topic models for
digital libraries. In: Proceedings of the 10th annual joint conference on Digital
libraries; 2010. p. 215–224.

144. Newman D, Bonilla EV, Buntine W. Improving topic coherence with regularized
topic models. Advances in neural information processing systems.
2011;24:496–504.
145. Röder M, Both A, Hinneburg A. Exploring the space of topic coherence
measures. In: Proceedings of the eighth ACM international conference on Web
search and data mining; 2015. p. 399–408.
146. Řehůřek R, Sojka P. Software Framework for Topic Modelling with Large
Corpora. In: Proceedings of the LREC 2010 Workshop on New Challenges for
NLP Frameworks. Valletta, Malta: ELRA; 2010. p. 45–50.
147. Easwaran E. The Bhagavad Gita for Daily Living: Commentary, Translation,
and Sanskrit Text, Chapters 13 Through 18. Nilgiri Press; 1984.
148. Sargeant W, Chapple CK. The Bhagavad Gita: Twenty-fifth–Anniversary
Edition. SUNY Press; 2009.
149. Singh K. The Sterling Book of Hinduism. Sterling Publishers Pvt. Ltd; 2011.

150. Nicholson A. Unifying Hinduism. Columbia University Press; 2010.

May 24, 2022 37/39

 151. Meena A. The View of Sustainability and Inclusiveness in Prasthanatrayi.
Jahnavi Sanskrit E-Journal. 2014;.

152. Rao M. A Brief History of the Bhagavad Gita’s Modern Canonization. Religion
Compass. 2013;7(11):467–475.
153. Lattanzio NG. I Am that I Am: Self-Inquiry, Nondual Awareness, and Nondual
Therapy as an Eclectic Framework. Argosy University/Schaumburg (Chicago
Northwest); 2020.

154. Mahanti P. The Concept of Purusa and Purusottama in the Commentary of

Sankara on the Prasthanatrayi. 1986;.
155. Nrugham L. Suicide in Ancient Hindu Scriptures: Condemned or Glorified?
Handbook of Suicidal Behaviour. 2017; p. 23–37.
156. Isayeva N. Shankara and Indian philosophy. SUNY Press; 1993.

157. Chinmayananda S. Srimad Bhagawad Geeta Chapter I & II. Central Chinmaya
Mission Trust; 2014.
158. Ranganathananda S. Universal Message of the Bhagavad Gita: An exposition of
the Gita in the Light of Modern Thought and Modern Needs. Advaita Ashrama
(A Publication House of Ramakrishna Math, Belur Math); 2000.
159. Salicchi L, Lenci A. PIHKers at CMCL 2021 Shared Task: Cosine Similarity and
Surprisal to Predict Human Reading Patterns. In: Proceedings of the Workshop
on Cognitive Modeling and Computational Linguistics; 2021. p. 102–107.
160. Thongtan T, Phienthrakul T. Sentiment classification using document
embeddings trained with cosine similarity. In: Proceedings of the 57th Annual
Meeting of the Association for Computational Linguistics: Student Research
Workshop; 2019. p. 407–414.
161. Gunawan D, Sembiring C, Budiman MA. The implementation of cosine
similarity to calculate text relevance between two documents. In: Journal of
physics: conference series. vol. 978. IOP Publishing; 2018. p. 012120.
162. Yelle RA. Explaining mantras: Ritual, rhetoric, and the dream of a natural
language in Hindu Tantra. Routledge; 2004.
163. Aurobindo S. Secret of the Veda. Lotus Press; 2018.

164. Pandit MP. Mystic Approach to the Veda and the Upanishad. Lotus Press;
1974.
165. Vivekananda S. Essentials of Hinduism. Advaita Ashrama (A publication
branch of Ramakrishna Math, Belur Math); 1937.

166. Das D, Das B, Mahesh K. A computational analysis of Mahabharata. In:

Proceedings of the 13th International Conference on Natural Language
Processing; 2016. p. 219–228.
167. MSS 0126, Special Collections, University of Delaware Library, Newark,
Delaware. William Butler Yeats Papers;. https://library.udel.edu/
special/findaids/view?docId=ead/mss0126.xml;tab=notes.

May 24, 2022 38/39

 168. Lee J, Yoon W, Kim S, Kim D, Kim S, So CH, et al. BioBERT: a pre-trained
biomedical language representation model for biomedical text mining.
Bioinformatics. 2020;36(4):1234–1240.

169. Tai W, Kung H, Dong XL, Comiter M, Kuo CF. exBERT: Extending
pre-trained models with domain-specific vocabulary under constrained training
resources. In: Findings of the Association for Computational Linguistics:
EMNLP 2020; 2020. p. 1433–1439.
170. Beltagy I, Lo K, Cohan A. SciBERT: A pretrained language model for scientific
text. arXiv preprint arXiv:190310676. 2019;.
171. Chalkidis I, Fergadiotis M, Malakasiotis P, Aletras N, Androutsopoulos I.
LEGAL-BERT: The muppets straight out of law school. arXiv preprint
arXiv:201002559. 2020;.

172. Rasmy L, Xiang Y, Xie Z, Tao C, Zhi D. Med-BERT: pretrained contextualized

embeddings on large-scale structured electronic health records for disease
prediction. NPJ digital medicine. 2021;4(1):1–13.
173. Müller M, Salathé M, Kummervold PE. Covid-twitter-bert: A natural language
processing model to analyse covid-19 content on twitter. arXiv preprint
arXiv:200507503. 2020;.

174. Chandra R, Kulkarni V. Semantic and sentiment analysis of selected Bhagavad

Gita translations using BERT-based language framework. IEEE Access.
2022;doi:doi.org/10.1109/ACCESS.2022.3152266.

May 24, 2022 39/39