THE ABILITY OF WORD EMBEDDINGS TO CAPTURE WORD SIMILARITIES

International Journal on Natural Language Computing (IJNLC) Vol.9, No.3, June 2020
DOI:10.5121/ijnlc.2020.9302 25
THE ABILITY OF WORD EMBEDDINGS TO
CAPTURE WORD SIMILARITIES
Martina Toshevska, Frosina Stojanovska and Jovan Kalajdjieski
Faculty of Computer Science and Engineering, Ss. Cyril and Methodius University,
Skopje, Macedonia
ABSTRACT
Distributed language representation has become the most widely used technique for language
representation in various natural language processing tasks. Most of the natural language processing
models that are based on deep learning techniques use already pre-trained distributed word
representations, commonly called word embeddings. Determining the most qualitative word embeddings is
of crucial importance for such models. However, selecting the appropriate word embeddings is a
perplexing task since the projected embedding space is not intuitive to humans.In this paper, we explore
different approaches for creating distributed word representations. We perform an intrinsic evaluation of
several state-of-the-art word embedding methods. Their performance on capturing word similarities is
analysed with existing benchmark datasets for word pairs similarities. The research in this paper conducts
a correlation analysis between ground truth word similarities and similarities obtained by different word
embedding methods.
KEYWORDS
Word Embeddings, Distributed Word Representation, Word Similarity
1. INTRODUCTION
Representing natural language has become one of the main concerns in a wide range of Natural
Language Processing (NLP) tasks. Some of the simplest techniques for representing natural
language, when the data is textual, are one-hot encoding, Bag of Words (BoW), term-frequency,
along with others [1]. These methods have a deficiency of incorporating the necessary
information in the mapped space of the representations. Representations such as one-hot
encoding create vector representations with enormous size (the size of the vocabulary) that do not
preserve the relative closeness of the words, and representations such as BoW do not capture the
context and order of the words.
Distributed representations of textual data as vectors have been proven useful in many
algorithms. Each word vector in a given corpus is represented based on the mutual information
with other words in the corpus [2]. Vector representation can be computed at different levels
including: characters [3], [4], words [5], [6], [7], phrases [8], [9], sentences [10], document
[11],etc. Word embeddings are representations of words in a continuous Rn space. Each word
isprojected in n-dimensional vector space as a real-valued vector by keeping the information
aboutthe meanings and similarities of the words. Words with similar meanings will be mapped
closer in the projected space. Word vectors capture syntactic and semantic regularities in
language, such that the operation “King - Man + Woman” results in a vector close to “Queen”
[5], [12].

26
Nowadays, several word embeddings methods incorporate different techniques for word
representation. Training the word embeddings requires large corpus, extensive computational
power, and time. Consequently, the preferred approach is to use the idea of transfer learning, i.e.
using already pre-trained embeddings and then training the NLP model with these embeddings.
However, selecting the appropriate word embeddings is a perplexing task since the projected
embedding space is not intuitive to humans. Nevertheless, deciding on the pre-trained word
embeddings used as input in the NLP model is the first and crucial step, since the quality of the
word representation has a significant influence on the global performance of the model.
Therefore, offering comparative experimental results for the methods can be valuable to
researchers for determining the most appropriate embeddings for their model based on the
comparative analysis. The evaluation of word embeddings is not unified. Generally, the
evaluation methods can be mainly grouped into two different approaches: an intrinsic and
extrinsic evaluation [13]. Intrinsic evaluations test the quality of a representation independent of
specific natural language processing tasks, directly testing the syntactic or semantic relationships
between words. Human-created benchmark datasets generated with a particular task of
comparing words by similarity, relatedness, or similar, are of crucial importance in the comparing
process of intrinsic evaluation. Extrinsic evaluation of word vectors is the evaluation of the
embeddings on a real NLP task, chosen as an evaluation method, for example, part-of-speech
tagging [14], named-entity recognition [15], sentiment analysis [16] and neural machine
translation [17].
Determining the most qualitative word embeddings is the main focus when executing the
evaluations. Additionally, the dimensionality of the word embeddings is an important segment of
the method capability to encode notable information in the projected space. Bigger
dimensionality means more space for incorporating more information in the space, but after
reaching some point, the marginal gain will diminish [5]. Typically, the dimensionality of the
vectors is set to be between 50 and 1,000, where 300 is the most frequently used. Consequently,
evaluating the word embeddings requires additional investigation for finding the most suitable
dimensionality.
In this paper, we focus on different pre-trained word embeddings used in state-of-the-art models
for some NLP tasks and analyse their performance on capturing word similarities with existing
benchmark datasets for word pairs similarities and correlation1
. The datasets WordSim353 [18],
SimLex999 [19], SimVerb3500 [20] and RG65 [21] are used in the experiments to compare the
state-of-the-art word embeddings. We also analysed the rare words dataset - RW2034 [22] to
investigate whether rare words impact the similarity results.
The rest of the paper is organized as follows. Section 2 describes the word embedding methods
used for comparative analysis. Section 3 gives an overview of the related research for the
problem. The comparative analysis along with the discussion of the results of the experiments are
included in Section 4. Finally, Section 5 concludes the paper and gives future work ideas.
2. WORD EMBEDDINGS
Mainly, word embeddings methods can be divided into three groups according to the technique of
creating the vectors: neural network based, word matrix based and a combination of the previous
two groups with an additional strategy of supplementary information fusion called ensemble
methods.
1
The code is available at https://github.com/mtoshevska/Comparing-Embeddings

27
2.1. Neural Network Based Methods
2.1.1. Word2Vec
Word2Vec includes two different models: Continuous Bag of Words (CBOW) and Skip-gram
[5], [6]. Both of these methods are neural networks with a hidden layer with N neurons, where N
is the dimensionality of the created word embeddings. The first method CBOW is a neural
network where the context of the words is the input of the network. The task is to predict the
current word as the output of the network [5]. The second method Skip-gram is a neural network
where the input is the one-hot encoding of the word, and the output is the predicted context of the
word, i.e. the surrounding words [5]. To make these algorithms more efficient techniques such as
Hierarchical Softmax and Skip-Gram Negative Sampling (SGNS) are used.
The most popular pre-trained Word2Vec word embeddings are 300-dimensional vectors
generated with negative sampling training on the Google News corpus (that contains 100 billion
words)2
[6]. These pre-trained embeddings are used in the experiments.
2.1.2. Fast Text
The Fast Text model [23], [24] is directly derived from the Skip-gram model of Word2Vec. The
authors claim that by using a distinct vector representation for each word, the Skip-gram model
ignores the internal structure of the words. For that purpose, they suggested a different scoring
function that takes into account the internal structure. Their subword model represents each word
as a bag of character n-gram. Special symbols < and > are added at the beginning and the end of
words to distinguish prefixes and suffixes from other character sequences. The word is also
included in the set of its n-grams to be able to learn a better representation for each word. The
authors suggest extracting all n-grams for n greater than or equal to 3 and smaller than or equal to
6.
This simple approach provides extraction of all prefixes and suffixes of a given word. When all
the n-grams of a specific word are extracted, a vector representation is assigned to every n-gram.
The final word is represented as the sum of the vector representations of all of its n-grams. This
model allows sharing the representations across words, thus allowing to learn a reliable
representation for rare words.
To bound the memory requirements, a hashing function - FNV-1a3
is used to map the n-grams to
integer numbers. There are also proposed approaches for producing compact architectures [25].
In our experiments, we have used pre-trained models both trained with subword information on
Wikipedia 2017 (16B tokens) and trained with subword information on Common Crawl (600B
tokens)4
.
2
https://code.google.com/p/word2vec/, last visited: 14.05.2020
3
http://www.isthe.com/chongo/tech/comp/fnv/, last visited: 14.05.2020
4
https://fasttext.cc/docs/en/english-vectors.html, last visited: 14.05.2020

28
2.2. Word Matrix Based Methods
2.2.1. GloVe
GloVe (Global Vectors for Word Representation) [7] is a log-bilinear regression model for
unsupervised learning of word representations that combines the advantages of two model
families: global matrix factorization and local context window methods. The general idea is that
co-occurrence ratio of any two words, that is the frequency of the words occurring in each other’s
context, encodes information about the words. It captures meaningful linear substructures,
efficiently leveraging global word-word co-occurrence matrix statistics. The model is optimized
so that the dot product of any pair of vectors is equal to the co-occurrence ratio of the
corresponding words.
There are several pre-trained GloVe vectors available for public use5
, trained on different datasets
such as Twitter and Wikipedia. In our experiments we include 50-dimensional and 200-
dimensional word vectors pre-trained on the Twitter dataset (27B tokens), and 50-dimensional
and 300-dimensional word vectors pre-trained on the Wikipedia dataset (6B tokens).
2.2.2. LexVec
LexVec [26] is based on the idea of factoring the PPMI matrix [27] using a reconstruction loss
function. This loss function does not weigh all errors equally, unlike SVD, but penalizes errors of
frequent co-occurrence more heavily while still treating negative co-occurrence, unlike GloVe.
The authors propose keeping the SGNS [6] weighting scheme by using window sampling and
negative sampling but explicitly factoring PPMI matrix rather than implicitly factorizing the
shifted PMI matrix. The minimization on the two terms of the loss function is done using two
approaches: mini-batch, which executes gradient descent in the same way as SGNS, and
stochastic, where every context window is extended with k negative samples and iterative
gradient descent is then run on pairs for each window. The authors state that the evaluation of
word similarity and analogy tasks shows that LexVec compares to and often outperforms state-
of-the-art methods on many of these tasks.
Several pre-trained LexVec vectors are available for public use6
, and the experiments in this
paper include the model trained with subword information on Common Crawl which contained
2,000,000 words (58B tokens) and the model trained on Wikipedia which contained 368,999
words (7B tokens).
2.3. Ensemble Methods
2.3.1. ConceptNetNumberbatch
ConceptNetNumberbatch represents an ensemble method that produces word embeddings from
large, multilingual vocabulary with a combination of the GloVe and Word2Vec embeddings and
additional structured knowledge from the semantic networks ConceptNet [28] and PPDB [29].
This method uses the extended retrofitting technique [30] to adjust the pre-trained word
embedding matrices with knowledge graphs. The objective is to learn a new embeddings matrix
such that the word embeddings are close by some distance metric to their counterparts in the
primary word embedding matrix and adjacent vertices in the knowledge network.
5
https://nlp.stanford.edu/projects/glove/, last visited: 14.05.2020
6
https://github.com/alexandres/lexvec, last visited: 14.05.2020

29
The pre-trained embeddings version ConceptNetNumberbatch 19.087
[31] is used in the
experiments in this paper.
3. RELATED WORK
There are several attempts to compare different methods for creating word embeddings on
different datasets for different purposes. Mainly the experiments are conducted on datasets
containing English words, but some are on other languages as well. Berardi et al. [32] have
conducted a comparison between the Skip-gram model of Word2Vec [5], [6] and GloVe [7] in
the Italian language. The models are trained on two datasets: the entire dump of the Italian
Wikipedia and a collection of 31,432 books (mostly novels) written in Italian. The authors
believe that by using two very different datasets, both by purpose and style, they will investigate
the impact of the training data on the two models. In this research, they conclude that
Word2Vec’s Skip-gram model outperforms GloVe on both datasets. However, it does not have
the same accuracy as trained on English datasets, which may be a sign of a higher complexity of
the Italian language.
Comparison between different models for word embeddings in different languages is also
conducted in [33], where the authors first compare Hungarian analogical questions to English
questions by training a Skip-gram model on the Hungarian Webcorpus. They obtain similar
results in the morphological questions, but the semantic questions are dominated by the English
model. They also conduct a proto dictionary generation and comparison using CBOW Word2Vec
and GloVe models on Hungarian/Slovenian/Lithuanian languages to English.
In contrast, there have been quite a few studies evaluating word embeddings in quantitatively
representing word semantics in the English language, mostly comparing the word embeddings
generated by different methods. Baroni et al. [34] conducted a comparison between four models:
Word2Vec’s CBOW8
, DISSECT9
, Distributional Memory model10
and Collobert and Weston11
.
The evaluation has been done using a corpus of 2.8 billion tokens. In their research, they
concluded that Word2Vec’s CBOW model outperformed other methods for almost all the tasks.
Ghannay et al. [35] evaluated the word embeddings generated by CSLM word embeddings [36],
dependency-based word embeddings [37], combined word embeddings and Word2Vec’s Skip-
gram model on multiple NLP tasks. The models were trained on the Gigaword corpus composed
of 4 billion words, and the authors found that the dependency-based word embeddings gave the
best performance. They also concluded that significant improvement can be obtained by a
combination of embeddings. Authors in [13] compared Word2Vec’s CBOW model, GloVe,
TSCCA [38], C&W embeddings [39], Hellinger PCA [40] and Sparse Random Projections [41]
and concluded that Word2Vec’s CBOW model outperformed the other encodings on 10 out of
the 14 test datasets. Apart from comparisons of embeddings in the general NLP domain, there
have also been comparisons of word embeddings on many specific domains such as biomedical
domains [42], [43].
7
https://github.com/commonsense/conceptnet-numberbatch, last visited: 14.05.2020
8
https://code.google.com/p/word2vec/, last visited: 14.05.2020
9
http://clic.cimec.unitn.it/composes/, last visited: 14.05.2020
10
http://clic.cimec.unitn.it/dm/, last visited: 14.05.2020
11
http://ronan.collobert.com/senna/, last visited: 14.05.2020

30
4. COMPARATIVE ANALYSIS
Word vectors can be evaluated using two different approaches. Extrinsic evaluation is an
evaluation with real NLP tasks such as natural language inference or sentiment analysis. For that
purpose, word vectors are incorporated into an embedding layer of a deep neural network. The
model is then trained on a specific NLP task. If the performances are bad, itis uncertain whether
the word vectors lack to represent the meaning of the words or simply the model is not good
enough for the task.
The prior issue justifies the need for intrinsic evaluation. This type of evaluation is an
intermediate evaluation where word vectors are directly evaluated on different tasks. The intrinsic
evaluation involves tasks as word vector analogies and word similarities.
Word vector analogies are used to evaluate word vectors by how well their cosine distance after
addition captures intuitive semantic and syntactic analogy questions. For example, the operation
“King - Man + Woman” results in a vector close to “Queen”, the operation “Windows -
Microsoft + Google” results in a vector close to “Android”, etc.
Word similarities are values that measure the similarity of word pairs. These values are collected
in several human-generated datasets that are applied as a resource for comparison of the quality
of mapping the words in some space with different word representations. Different datasets
represent the similarity differently, by either coupling or segregating the word similarity with
word relatedness or word association.
4.1. Data
WordSim35312
dataset consists of 353 noun pairs with human similarity ratings [18]. The dataset
is used as a golden standard in the field of word similarity and relatedness computation. The
similarity scores of the word pairs in this resource are a representative measure for the similarity
of the words dependent on the relatedness or association. Later the dataset is annotated with
semantic relations by distinguishing the concepts of similarity and relatedness [44].
SimLex99913
has recently become a widely used lexical resource for tracking progress in word
similarity computation. This benchmark dataset is concentrated on measuring the similarity,
rather than relatedness or association [19]. It consists of 666 noun-noun pairs,222 verb-verb pairs
and 111 adjective-adjective pairs. The annotation guidelines of the dataset narrow the similarity
to synonymy [45]. SimLex999 can be used to measure the ability of the word embedding models
to capture the similarity of the words (without the influence of the relatedness or association)
while at the same time excluding the dependence of relatedness or association.
SimVerb350014
represents a dataset for verb similarity that consists of 3500 verb-verb pairs [20].
This dataset is built on the same annotation criteria as the SimLex-999, i.e. there is a difference
between similarity and relatedness. Therefore, this method is a broad coverage resource for verb
similarities.
RG65 [21] is one of the earliest created datasets for word similarities. This dataset contains 65
word pairs with a similarity score for each pair ranging between 0 and 4. The similarity scores are
a mean value of human judgements from 51 subjects. The size of this dataset is smaller than
12
http://alfonseca.org/eng/research/wordsim353.html, last visited: 14.05.2020
13
https://fh295.github.io/SimLex-999.zip, last visited: 14.05.2020
14
https://www.repository.cam.ac.uk/handle/1810/264124, last visited: 14.05.2020

31
the previously mentioned datasets. Still, it is one of the most used in the previous research work,
so we include this dataset in our experiments.
RW203415
[22] is a word similarity dataset of 2034 word pairs assessed by semantic similarity
score in a scale from 0 to 10. This dataset contains words with low occurrences (rare words). It
can be employed for testing the word representations of morphologically complex and rare
words.
4.2. Word Similarity Experiments
4.2.1. Word Similarities
Word embeddings represent words in a way that words with similar meaning are represented with
similar vectors. The experiments presented in this section evaluate the ability to capture word
similarities of different pre-trained word embeddings. For each word pair in the datasets, the
cosine similarity of the corresponding word embedding vectors is computed. Since the range of
similarity in the datasets is [0, 10], the cosine similarity values are normalized in that range.
Similarities from the RG65 dataset are also normalized in the range [0, 10]. Average cosine
similarities for each dataset are plotted in Figure 1.
Average similarity obtained by human ratings is 5.86, 4.56, 4.29, 4.69 and 6.15 for WordSim353,
SimLex999, SimVerb3500, RG65 and RW2034 dataset, respectively. The closest average cosine
similarity is obtained among GloVe and FastText word embeddings for each dataset except
RW2034. The average cosine similarities for GloVe embeddings is 5.37, 4.62, 3.79 and 4.48 for
WordSim353, SimLex999, SimVerb3500 and RG65 dataset, respectively. With the FastText
embeddings, average cosine similarity is 4.69, 4.81, 4.12 and 4.17 for WordSim353, SimLex999,
SimVerb3500 and RG65 dataset, respectively. These values direct to the conclusion that FastText
and GloVe perform better in capturing similarities between words. However, this statement does
not hold for RW2034 dataset. For this dataset, closest cosine similarity, 3.72, is obtained by
ConceptNetNumberbatch word embeddings. Since it is composed of rare words, we can conclude
that ConceptNetNumberbatch word embeddings have better ability to capture similarities
between rare words.
In terms of dimensionality, it is expected that word vectors of higher dimensionality have a more
prominent quality of the representation. This statement is confirmed for SimLex999 dataset and
slightly confirmed for SimVerb3500 dataset, when analysing the GloVe embeddings with
different dimensionalities. For SimLex999 dataset, average cosine similarity is closer to ground
truth similarity for 200-dimensional embeddings pre-trained on Twitter and 300-dimensional
embeddings pre-trained on Wikipedia. For SimVerb3500 dataset, the statement applies only for
200-dimensional embeddings pre-trained on Twitter, while the similarities for vectors pre-trained
on Wikipedia are closer to ground truth similarities when the dimensionality is 50. The findings
for WordSim353, RG65 and RW2034 dataset contradict the hypothesis about dimensionality. In
fact, for these datasets lower vector dimensionality, 50, implies closer cosine to ground-truth
similarity indicating that lower-dimensional word vectors are capable of storing significant
information about words.
15
https://nlp.stanford.edu/ lmthang/morphoNLM/, last visited 14.05.2020

32
Figure 1: Average cosine similarities. GT - ground truth similarities. F-C-300 - 300-dimensional FastText
embeddings pre-trained on Crawl dataset. F-W-300 - 300-dimensional FastText embeddings pre-trained on
Wikipedia dataset. G-T-200 - 200-dimensional GloVe embeddings pre-trained on Twitter dataset. G-T-50 -
50-dimensional GloVe embeddings pre-trained on Twitter dataset. G-W-300 - 300-dimensional GloVe
embeddings pre-trained on Wikipedia dataset. G-W-50 - 50-dimensional GloVe embeddings pre-trained on
Wikipedia dataset. L-C-300 - 300-dimensional LexVec embeddings pre-trained on Crawl dataset. L-W-300
- 300-dimensional LexVec embeddings pre-trained on Wikipedia dataset. N-C-300 - 300-dimensional
ConceptNetNumberbatch embeddings. W-G-300 - 300-dimensional Word2Vec embeddings pre-trained on
Google News dataset.

33
4.2.2. Correlation Analysis
Average similarity does not always provide good insight into word similarities since similarities
obtained by one metric can be higher than similarities obtained by another metric. Despite
different average values, both similarity distributions could be correlated. Therefore, for all pre-
trained word embedding vectors, we calculate the correlation coefficient between cosine
similarities of word vectors and ground truth similarities.
Spearman correlation coefficient, Pearson correlation coefficient and Kendall’s tau correlation
coefficient are computed for each pair of ground-truth similarities and cosine similarities. The
results are summarized in Table 1 for all datasets and are depicted in Figure 2 for the
WordSim353 dataset, Figure 3 for the SimLex999 dataset, Figure 4 for the SimVerb3500 dataset,
Figure 5 for the RG65 dataset and Figure 6 for the RW2034 dataset.
Results of the correlation analysis oppose the findings of average similarity. GloVe and FastText
word embeddings demonstrated better performance for capturing word similarities in terms of
average cosine similarity among four out of five datasets. However, they have the lowest
correlation coefficient. The value is around 0, implying no correlation between ground truth and
cosine similarities of word vectors. FastText, despite being developed to overcome weaknesses of
Word2Vec, does not show to capture word similarities better since correlation is roughly 0.1 for
SimLex999 and SimVerb3500 and approximately 0.2 for WordSim353, RG65 and RW2034.
GloVe also shows weak performance on SimLex999 and SimVerb3500, although it is used as a
starting point in many high-performance deep learning models for different NLP tasks. However,
the correlation on WordSim353 and RG65 is quite higher, with a maximum value of 0.61 for the
Spearman correlation coefficient on the WordSim353 dataset and 0.77, again Spearman
correlation, on the RG65 dataset. On the RW2034 dataset, the highest correlation coefficient for
GloVe embeddings is 0.41 with 300-dimensional vectors pre-trained on Wikipedia.
For all five datasets, the highest correlation is achieved by ConceptNetNumberbatch word
embeddings, indicating a positive correlation between similarity distributions. On the other side,
the worst performing word vectors are GloVe vectors on the SimVerb3500 dataset and FastText
vectors on all other datasets.
The second highest correlation coefficient belongs to Word2Vec for the SimVerb3500 with 0.38
Pearson correlation, and RW2034 with 0.53 Spearman correlation. For the other datasets, at the
second place are LexVec embeddings with 0.72 Spearman correlation for the WordSim353
dataset, 0.45 Pearson correlation for the SimLex999 dataset, and 0.77 Pearson correlation for
RG65 dataset. The third highest correlation is obtained with Word2Vec embeddings for the
WordSim353 (0.69 Spearman correlation), SimLex999 (0.45 Pearson correlation) and RG65
(0.77 Pearson correlation), and with LexVec embeddings for SimVerb3500 (0.31 Pearson
correlation) and RW2034 (0.51 Spearman correlation).
The overall conclusion is that ensemble methods demonstrate promising performance and,
despite being relatively new, could possibly improve the performance of other tasks in the NLP
domain. However, state-of-the-art methods based on matrix factorization and neural networks,
such as Word2Vec, still achieve comparable performance and therefore their widespread usage in
most tasks is justified.
The assumption about the better quality of word vectors with higher dimensionality corresponds
with the correlation results. Correlation coefficients for GloVe word embeddings despite being
low for the SimLex999 and SimVerb3500 dataset, show that higher dimensionality entails higher

34
correlation. Furthermore, we can infer that word vectors of higher dimensionality are better for
capturing word similarities.
Table 1: Correlation coefficients between ground truth similarities and word vector cosine similarities. WS
- WordSim353. SL - SimLex999. SV - SimVerb3500. RG - RG65. RW - RW2034. S - Spearman
correlation coefficient. Pr - Pearson correlation coefficient. K - Kendall’s tau correlation coefficient.
Word Embeddings
Corr
Coeff
WS SL SV RG RW
GloVe-Wikipedia-50
S 0.50 0.26 0.15 0.59 0.34
Pr 0.51 0.29 0.16 0.56 0.34
K 0.36 0.18 0.10 0.42 0.23
GloVe-Wikipedia-300
S 0.61 0.37 0.23 0.77 0.41
Pr 0.60 0.39 0.23 0.75 0.40
K 0.45 0.26 0.16 0.57 0.28
GloVe-Twitter-50
S 0.46 0.10 0.03 0.57 0.30
Pr 0.46 0.10 0.04 0.56 0.30
K 0.32 0.06 0.02 0.40 0.21
GloVe-Twitter-200
S 0.52 0.13 0.06 0.70 0.33
Pr 0.52 0.14 0.07 0.71 0.32
K 0.36 0.08 0.04 0.50 0.22
FastText-Wikipedia-300
S 0.19 0.09 0.03 0.16 0.23
Pr 0.19 0.07 0.02 0.16 0.21
K 0.13 0.06 0.02 0.10 0.15
FastText-Crawl-300
S 0.25 0.16 0.11 0.12 0.27
Pr 0.26 0.16 0.11 0.18 0.26
K 0.17 0.11 0.07 0.09 0.18
LexVec-Wikipedia-300
S 0.66 0.38 0.28 0.75 0.49
Pr 0.63 0.39 0.28 0.76 0.46
K 0.48 0.27 0.19 0.56 0.34
LexVec-Crawl-300
S 0.72 0.44 0.30 0.75 0.51
Pr 0.68 0.45 0.31 0.77 0.49
K 0.53 0.31 0.21 0.58 0.35
Word2Vec-GoogleNews-300
S 0.69 0.44 0.36 0.76 0.53
Pr 0.65 0.45 0.38 0.77 0.51
K 0.51 0.31 0.25 0.58 0.37
Numberbatch-ConceptNet-300
S 0.81 0.63 0.57 0.92 0.64
Pr 0.75 0.65 0.59 0.90 0.61
K 0.63 0.46 0.41 0.77 0.46

35
Figure 2: Correlation between ground truth similarities and similarities obtained with different word
embedding methods for the WordSim353 dataset. Gt_sim - ground truth similarities obtained by human
ratings. Cosine_sim - cosine similarities between word embeddings.

36
embedding methods for the SimLex999 dataset. Gt_sim - ground truth similarities obtained by human

37
embedding methods for the SimVerb3500 dataset. Gt_sim - ground truth similarities obtained by human

38
embedding methods for the RG65 dataset. Gt_sim - ground truth similarities obtained by human ratings.
Cosine_sim - cosine similarities between word embeddings.

39
embedding methods for the RW2034 dataset. Gt_sim - ground truth similarities obtained by human ratings.
Cosine_sim - cosine similarities between word embeddings.

40
5. CONCLUSIONS AND FUTURE WORK
In this paper, we have analysed several word embedding methods. These methods could be
generally divided into two groups according to the creation technique: neural network based and
matrix based. We further examined an ensemble method which is a compound of the two
methods.
Word embedding similarity experiments were conducted for all of the examined methods. These
experiments evaluate the ability to capture word similarities between pairs of words on five
different datasets using a cosine similarity measure. Analysing only the average similarity
information and comparing these values with the average value of the human ratings, the
conclusion was that the GloVe and FastText outperformed the other word embedding methods on
four out of the five datasets. However, for RW2034 dataset, ConceptNetNumberbatch achieved
the closest similarity results, so we can conclude that these embeddings better capture similarities
between rare words.
After conducting the word embedding similarity experiments, we also carried out correlation
analysis. In this analysis, the Spearman correlation coefficient, Pearson correlation coefficient
and Kendall’s tau correlation coefficient were computed for each pair of ground truth similarities
and cosine similarities of the word embeddings. Using the data obtained in this analysis we have
concluded that even with the high cosine similarity values of GloVe and FastText, their
correlation values are close to 0, implying no correlation between ground truth and cosine
similarities of word vectors. Moreover, ConceptNetNumberbatch word embeddings have the
highest correlation coefficients even though they did not have high average similarity values,
except for the RW2034 dataset.
The overall conclusion would be that ensemble methods show promising performance and
although relatively new could possibly improve the performance of other tasks in the NLP
domain. However, state-of-the-art methods based on matrix factorization and neural networks,
such as Word2Vec have comparable performances and therefore their widespread usage in most
tasks is justified.
A more comprehensive study of word similarities is including the context of the text of the
words. Newest word embedding methods, like ELMO [46] or BERT [47], utilize the context of
the words as they appear in the text, creating deep contextualized word representations.
Examining the difference between these different mappings of the words will be part of our
future work. Additionally, combining the different ideas from intrinsic and extrinsic evaluations
would provide a more beneficial comparison of the different word representations.
REFERENCES
[1] Manning, Christopher D., Prabhakar Raghavan, and HinrichSchütze. “Scoring, term weighting and
the vector space model.” Introduction to information retrieval 100 (2008): 2-4.
[2] Mikolov, Tomáš, Wen-tau Yih, and Geoffrey Zweig. “Linguistic regularities in continuous space
word representations.” Proceedings of the 2013 conference of the north american chapter of the
association for computational linguistics: Human language technologies. 2013.
[3] Santos, Cicero D., and Bianca Zadrozny. “Learning character-level representations for part-of-speech
tagging.” Proceedings of the 31st international conference on machine learning (ICML-14). 2014.
[4] dos Santos, Cıcero, et al. “Boosting Named Entity Recognition with Neural Character Embeddings.”
Proceedings of NEWS 2015 The Fifth Named Entities Workshop. 2015.
[5] Mikolov, Tomas, et al. “Efficient Estimation of Word Representations in Vector Space.” ICLR
(Workshop Poster). 2013.

41
[6] Mikolov, Tomas, et al. “Distributed representations of words and phrases and their compositionality.”
Advances in neural information processing systems. 2013.
[7] Pennington, Jeffrey, Richard Socher, and Christopher D. Manning. “Glove: Global vectors for word
representation.” Proceedings of the 2014 conference on empirical methods in natural language
processing (EMNLP). 2014.
[8] Yu, Mo, and Mark Dredze. “Learning composition models for phrase embeddings.” Transactions of
the Association for Computational Linguistics 3 (2015): 227-242.
[9] Zhou, Zhihao, Lifu Huang, and Heng Ji. “Learning Phrase Embeddings from Paraphrases with
GRUs.” Proceedings of the First Workshop on Curation and Applications of Parallel and Comparable
Corpora. 2017.
[10] Kiros, Ryan, et al. “Skip-thought vectors.” Advances in neural information processing systems. 2015.
[11] Le, Quoc, and Tomas Mikolov. “Distributed representations of sentences and documents.”
International conference on machine learning. 2014.
[12] Levy, Omer, and Yoav Goldberg. “Linguistic regularities in sparse and explicit word representations.”
Proceedings of the eighteenth conference on computational natural language learning. 2014.
[13] Schnabel, Tobias, et al. “Evaluation methods for unsupervised word embeddings.” Proceedings of the
2015 conference on empirical methods in natural language processing. 2015.
[14] Ratnaparkhi, Adwait. “A maximum entropy model for part-of-speech tagging.” Conference on
Empirical Methods in Natural Language Processing. 1996.
[15] Yadav, Vikas, and Steven Bethard. “A Survey on Recent Advances in Named Entity Recognition from
Deep Learning models.” Proceedings of the 27th International Conference on Computational
Linguistics. 2018.
[16] Liu, Bing. Sentiment analysis: Mining opinions, sentiments, and emotions. Cambridge University
Press, 2015.
[17] Cho, Kyunghyun, et al. “Learning Phrase Representations using RNN Encoder–Decoder for
Statistical Machine Translation.” Proceedings of the 2014 Conference on Empirical Methods in
Natural Language Processing (EMNLP). 2014.
[18] Finkelstein, Lev, et al. “Placing search in context: The concept revisited.” Proceedings of the 10th
international conference on World Wide Web. 2001.
[19] Hill, Felix, Roi Reichart, and Anna Korhonen. “Simlex-999: Evaluating semantic models with
(genuine) similarity estimation.” Computational Linguistics 41.4 (2015): 665-695.
[20] Gerz, Daniela, et al. “SimVerb-3500: A Large-Scale Evaluation Set of Verb Similarity.” Proceedings
of the 2016 Conference on Empirical Methods in Natural Language Processing. 2016.
[21] Rubenstein, Herbert, and John B. Goodenough. “Contextual correlates of synonymy.”
Communications of the ACM 8.10 (1965): 627-633.
[22] Luong, Minh-Thang, Richard Socher, and Christopher D. Manning. “Better word representations with
recursive neural networks for morphology.” Proceedings of the Seventeenth Conference on
Computational Natural Language Learning. 2013.
[23] Bojanowski, Piotr, et al. “Enriching word vectors with subword information.” Transactions of the
Association for Computational Linguistics 5 (2017): 135-146.
[24] Joulin, Armand, et al. “Bag of Tricks for Efficient Text Classification.” Proceedings of the 15th
Conference of the European Chapter of the Association for Computational Linguistics: Volume 2,
Short Papers. 2017.
[25] A. Joulin, E. Grave, P. Bojanowski, M. Douze, H. Jegou, and T. Mikolov, “Fasttext. zip:
Compressing text classification models,” arXiv preprint arXiv:1612.03651 (2016).
[26] Salle, Alexandre, Aline Villavicencio, and Marco Idiart. “Matrix Factorization using Window
Sampling and Negative Sampling for Improved Word Representations.” Proceedings of the 54th
Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers). 2016.
[27] Bullinaria, John A., and Joseph P. Levy. “Extracting semantic representations from word co-
occurrence statistics: A computational study.” Behavior research methods 39.3 (2007): 510-526.
[28] Speer, Robyn, and Catherine Havasi. “Representing General Relational Knowledge in ConceptNet
5.” Proceedings of the Eighth International Conference on Language Resources and Evaluation
(LREC'12). 2012.
[29] Ganitkevitch, Juri, Benjamin Van Durme, and Chris Callison-Burch. “PPDB: The paraphrase
database.” Proceedings of the 2013 Conference of the North American Chapter of the Association for
Computational Linguistics: Human Language Technologies. 2013.

42
[30] R. Speer and J. Chin, “An ensemble method to produce high-quality word embeddings.” arXiv
preprint arXiv:1604.01692 (2016).
[31] Speer, Robyn, Joshua Chin, and Catherine Havasi. “Conceptnet 5.5: An open multilingual graph of
general knowledge.” Thirty-First AAAI Conference on Artificial Intelligence. 2017.
[32] Berardi, Giacomo, Andrea Esuli, and Diego Marcheggiani. “Word Embeddings Go to Italy: A
Comparison of Models and Training Datasets.” IIR. 2015.
[33] Makrai, Márton, et al. “Comparison of distributed language models on medium-resourced languages.”
XI. Magyar SzámítógépesNyelvészetiKonferencia (MSZNY 2015) (2015).
[34] Baroni, Marco, Georgiana Dinu, and Germán Kruszewski. “Don’t count, predict! a systematic
comparison of context-counting vs. context-predicting semantic vectors.” Proceedings of the 52nd
Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2014.
[35] Ghannay, Sahar, et al. “Word embedding evaluation and combination.” Proceedings of the Tenth
International Conference on Language Resources and Evaluation (LREC'16). 2016.
[36] Schwenk, Holger. “CSLM-a modular open-source continuous space language modeling toolkit.”
INTERSPEECH. 2013.
[37] Levy, Omer, and Yoav Goldberg. “Dependency-based word embeddings.” Proceedings of the 52nd
Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers). 2014.
[38] Dhillon, Paramveer S., et al. “Two step CCA: a new spectral method for estimating vector models of
words.” Proceedings of the 29th International Conference on International Conference on Machine
Learning. 2012.
[39] Collobert, Ronan, et al. “Natural language processing (almost) from scratch.” Journal of machine
learning research 12.Aug (2011): 2493-2537.
[40] Lebret, Rémi, and Ronan Collobert. “Word Embeddings through Hellinger PCA.” Proceedings of the
14th Conference of the European Chapter of the Association for Computational Linguistics. 2014.
[41] Li, Ping, Trevor J. Hastie, and Kenneth W. Church. “Very sparse random projections.” Proceedings of
the 12th ACM SIGKDD international conference on Knowledge discovery and data mining. 2006.
[42] Pakhomov, Serguei VS, et al. “Corpus domain effects on distributional semantic modeling of medical
terms.” Bioinformatics 32.23 (2016): 3635-3644.
[43] Wang, Yanshan, et al. “A comparison of word embeddings for the biomedical natural language
processing.” Journal of biomedical informatics 87 (2018): 12-20.
[44] Agirre, Eneko, et al. “A Study on Similarity and Relatedness Using Distributional and WordNet-based
Approaches.” Proceedings of Human Language Technologies: The 2009 Annual
Conference of the North American Chapter of the Association for Computational Linguistics. 2009.
[45] Kliegr, Tomáš, and OndřejZamazal. “Antonyms are similar: Towards paradigmatic association
approach to rating similarity in SimLex-999 and WordSim-353.” Data & Knowledge Engineering
115 (2018): 174-193.
[46] Peters, Matthew E., et al. “Deep Contextualized Word Representations.” Proceedings of the 2018
Conference of the North American Chapter of the Association for Computational Linguistics: Human
Language Technologies, Volume 1 (Long Papers). 2018.
[47] Devlin, Jacob, et al. “BERT: Pre-training of Deep Bidirectional Transformers for Language
Understanding.” Proceedings of the 2019 Conference of the North American Chapter of the
Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and
Short Papers). 2019.

THE ABILITY OF WORD EMBEDDINGS TO CAPTURE WORD SIMILARITIES

More Related Content

What's hot (19)

Similar to THE ABILITY OF WORD EMBEDDINGS TO CAPTURE WORD SIMILARITIES (20)

More from kevig (20)

Recently uploaded (20)

THE ABILITY OF WORD EMBEDDINGS TO CAPTURE WORD SIMILARITIES