Social Media Sentiment Analysis Using Twitter Dataset

Introduction
In today's digital landscape, individuals communicate their thoughts primarily through blogs,
online forums, and social media platforms like Facebook and Twitter. Millions share their
experiences, forming dynamic online communities where consumer feedback influences
decisions. This surge of emotional data from tweets to reviews presents significant opportunities
for businesses to engage effectively with customers. As consumers increasingly rely on user-
generated content for purchasing decisions, the need for automated sentiment analysis (SA)
becomes clear. SA helps determine whether products meet expectations and allows businesses to
tailor offerings to customer needs. While traditional data analysis focuses on objective facts,
sentiment analysis explores subjective opinions and emotions, paving the way for new
applications. For example, recommendation systems can improve accuracy by incorporating SA
to predict product recommendations based on user sentiment.

Problem Statement

The rapid growth of user-generated content on Twitter presents a challenge in effectively

analyzing the vast array of sentiments expressed in tweets. Despite the importance of
understanding public opinion for businesses and consumers, the unstructured and varied nature
of this data complicates manual analysis. This research aims to develop and implement
automated sentiment analysis techniques that can accurately classify and interpret the sentiments
expressed in Twitter data. By addressing issues such as noise, ambiguity, and the dynamic nature
of language used on social media, the goal is to provide actionable insights that enhance
decision-making for businesses and improve user experience.

Objectives

Data Collection and Preprocessing: To gather a diverse dataset of tweets and preprocess the
data to remove noise, such as irrelevant content, hashtags, and links, ensuring a clean dataset for
analysis.

Sentiment Classification: To develop and implement algorithms for classifying tweets into
distinct sentiment categories (e.g., positive, negative, neutral) using machine learning and natural
language processing techniques.

Feature Extraction: To identify and extract relevant linguistic features from tweets, such as
sentiment-laden words, phrases, and contextual cues that contribute to accurate sentiment
detection.
Model Evaluation and Optimization: To evaluate the performance of various sentiment
analysis models and optimize them for higher accuracy, precision, and recall in classifying
sentiments.

Insights and Visualizations: To generate meaningful insights from the analyzed data and
present the findings through visualizations that illustrate trends, patterns, and public sentiment
over time

Methodology
Requirement Identification
Literature Review
The paper [1] (Kaur, 2015) describe regarding geographic area flood information set collected
from twitter and realize the opinion of individuals. They used Naive Bayes formula for the
classification of information and result they got 67% accuracy. They need collected several
resolution from the individuals that are useful for each government and non-government
organization to handle such scenario in an exceedingly higher manner. These strategies simpler
than lexicon-based formula.
In 2019, Saad and Yang [2] have aimed for giving a complete tweet sentiment analysis on the
basis of ordinal regression with machine learning algorithms. The suggested model included pre-
processing tweets as first step and with the feature extraction model, an effective feature was
generated. The methods such as SVR, RF, Multinomial logistic regression (SoftMax), and DTs
were employed for classifying the sentiment analysis. Moreover, twitter dataset was used for
experimenting the suggested model. The test results have shown that the suggested model has
attained the best accuracy, and also DTs were performed well when compared over other
methods.
In 2018, Fang et al. [3] have suggested multi-strategy sentiment analysis models using the
semantic fuzziness for resolving the issues. The outcomes have demonstrated that the proposed
model has attained high efficiency
Requirement Analysis
Feasibility study
i)Technical
Data Access: Twitter provides APIs for data collection, enabling access to large volumes of
tweets for sentiment analysis.
Tools and Technologies: Utilize established libraries and frameworks (e.g., Python, NLTK,
TensorFlow) for natural language processing and machine learning.
Infrastructure: Assess the computational requirements for data processing and model training;
cloud-based solutions can provide scalable resources.
Algorithm Development: Existing machine learning algorithms (e.g., SVM, LSTM) are well-
suited for sentiment classification, ensuring a solid foundation for development.

ii)Operational
Ensure the availability of team members with expertise in data science, machine learning, and
natural language processing and developing workflows for data collection, preprocessing,
analysis, and reporting that can be integrated into existing business operations.Establish
processes for continuous monitoring and updating of models to adapt to evolving language and
sentiment trends on Twitter.

iii)Economic

Sentiment analysis can be defined as an automated process of extracting attitudes, opinions,

opinions, and emotions from text, speech, tweets and database sources through language
processing. natural language (NLP). Sentiment analysis involves classifying opinions in the text
into categories such as "positive" or "negative" or "neutral". It is also known as subjective
analysis, polling and evaluative extraction.
The words opinion, feel, opinion, and belief are used interchangeably, but there is a difference
between them.
Opinion A controversial conclusion (because different experts have different opinions)
View Subjective opinion
Belief Intentional acceptance and intellectual consent
Sentiment Opinions represent one's feelings

Sentiment analysis is a term that encompasses many tasks such as emotion mining, sentiment
classification, subjective classification, opinion synthesis, or opinion spam detection, among
others. It aims to analyze feelings, attitudes, opinions, emotions, etc. of people related to items
such as products, individuals, objects, organizations, and services. Mathematically, we can
represent an opinion as a percentile (o, f, so, h, t), where: o = object; f = feature of object o; so =
orientation or polarity of opinion about the f characteristic of object o; h = opinion holder; t =
time the opinion was expressed.
Object An entity can be a person, event, product, organization, or subject
Feature An attribute (or part) of the object on which the evaluation is performed.
Opinion orientation The orientation of an opinion about a characteristic f indicates whether
or polarity the opinion is positive, negative, or neutral.
Opinion holder An opinion holder is a person or organization or entity expressing an
 opinion.

In recent years, several researchers have done a lot of work in the field of "Twitter Sentiment
Analysis". In its early days, it was used for binary classifications that assigned opinions or ratings
to dipole classes, such as positive or negative.
[1] proposed a model to classify tweets as objective, positive, and negative. They created a
Twitter database by collecting tweets using the Twitter API and automatically captioning those
tweets with emojis. Using this dataset, they developed a sentiment classifier based on the
polynomial Naive Bayes method using features such as N-gram tags and POS. The training set
they used was less effective because it only contained tweets with emojis.
[2] deployed two models, a Naive Bayes bigram model and a Maximum Entropy model to
classify tweets. They found that the Naive Bayes classifier performed much better than the
Maximum Entropy model.
[3] proposed an emotion analysis solution for Twitter data using remote monitoring, where
their training data consisted of tweets with emojis acting as big tags. They build models using
Naive Bayes, MaxEnt, and Support Vector Machines (SVM). Their feature spaces include
unigram, bigram, and store. They conclude that SVM is superior to other models and unigram is
more efficient in terms of features.
[4] designed a two-stage automated sentiment analysis method for ranking tweets. They
classify tweets as objective or subjective, then in the second stage, subjective tweets are
classified as positive or negative. The feature space used includes retweets, hashtags, links,
punctuation, and exclamation points along with features like pre-word polarization and selling
points.
[5] used Twitter streaming data provided by the Firehouse API, making every user post
public in real time. They experimented with naive polynomial Bayes, stochastic gradient descent,
and Hoeffding trees. They concluded that the SGD-based model, when used with an appropriate
learning rate, is better than the rest used.
[6] developed a 3-dimensional model to classify sentiments into positive, negative, and
neutral classes. They experimented with models such as: unigram model, feature-based model,
and tree-core model. For the tree-core-based model, they represent tweets as a tree. The feature-
based model uses 100 features, and the unigram model uses more than 10,000 features. They
concluded that the features that combine the anterior polarities of words with their pronunciation
tag (pos) are the most important and play a key role in the classification task. The tree-core-
based model outperforms the other two.
[7] proposed an approach to use user-defined hashtags in tweets as sentiment classification
using punctuation, single word, n-gram and patterns as feature types are then combined into a
single feature vector for sentiment classification. They used the K-Nearest Neighbor strategy to
label sentiment by constructing a feature vector for each example in the training and testing sets.
[8] used Twitter API to collect Twitter data. Their training data falls into three different
categories (camera, film, mobile). The data is labeled as positive, negative and no opinion. Posts
containing filtered comments. The Unigram Naive Bayes model was implemented, and the
simplified Naive Bayes independence assumption was used. They also remove unnecessary
features using mutual information and chi-square feature extraction methods. Finally, the
direction of a tweet is predicted. i.e., positive, or negative.
[9] presented variations of the Naive Bayes classifier to detect the polarization of tweets in
English. Two different variations of the Naive Bayes classifier were built, namely Baseline
(trained to classify tweets as positive, negative, and neutral) and Binary (using polar and
categorical vocabulary) are positive and negative. Neutral tweets are ignored). Features taken
into account by taxonomists are lemma (noun, verb, adjective, and adverb), polar and multiple
lexicons from different sources, and valence shifts.
[10] used the bag-of-words method for sentiment analysis, where the relationship between
words is not considered at all and a document is represented as a set of simple words. To
determine the sentiment for the entire document, the sentiment of each word was determined,
and these values were unified with several aggregate functions.
[11] used the word Network vocabulary database to determine the emotional content of a word
in different dimensions. They developed a distance measure on the WordNet and determined the
semantic poles of adjectives.
[12] used a synthetic framework for sentiment classification achieved by combining different
feature sets and classification techniques. In their work, they used two types of feature sets
(partial speech information and word relationships) and three basic classifiers (Naive Bayes,
Maximum Entropy and Support Vector Machines). They applied synthetic approaches such as
fixed matching, weighted matching and meta-classifier matching to classify sentiments and
achieve better accuracy.
[13] highlights effective challenges and techniques for extracting comments from Twitter
tweets. Spam and extremely diverse languages make it difficult to get comments in Twitter.
A general model for sentiment analysis is as follows:
 Figure – 1: Sentiment Analysis Architecture

Pre-processing of the dataset

A tweet containing multiple opinions on data expressed in different ways by different users. The
Twitter dataset used in this survey work has been labeled into two classes viz. cathode and anode
and thus sentiment analysis of the data becomes easy to observe the influence of different
characteristics. Polar raw data is prone to inconsistencies and redundancies. Tweet preprocessing
includes the following points:
• Remove all URLs (e.g., www.xyz.com), hash tags (e.g., #subject), target (@username) •
Correct spelling: Strings of repeating characters must be handled
• Replace all emojis with their emotions.
• Remove all punctuation, symbols, numbers
• Remove stop words
• Expand abbreviations (we can use a dictionary of acronyms)
• Delete non-English tweets
Table 1: Publicly Available Datasets for Twitter.
31,861 Pos tweets
HASH Tweets http://demeter.inf.ed.ac.uk
64,850 Neg tweets,
 125,859 Neu tweets
Tweets and 230,811 Pos&
EMOT http://twittersentiment.appspot.com
Emoticons 150,570 Neg tweets
1,520 Pos
tweets,200 Neg
ISIEVE Tweets www.i-sieve.com
tweets, 2,295 Neu
tweets
Columbia
university Tweets Email: [email protected] 11,875 tweets
dataset
Patient 2000 patient
Opinions http://patientopinion.org.uk
dataset opinions
Sample Tweets http://goo.gl/UQvdx 667 tweets
Stanford http://ai.stanford.edu/~amaas/data/sentim 50000 movie
Movie Reviews
dataset ent reviews
4 million tweets
http://cs.stanford.edu/people/alecmgo/trai categorized as
Stanford Tweets
ningandtestdata.zip positive and
negative
400 deceptive and
400 truthful reviews
Spam dataset Spam Reviews http://myleott.com/op_spam
in positive and
negative category.
1,000 discussions,
Sarcasm and ~390,000 posts, and
Soe dataset http://nlds.soe.ucsc.edu/iac
nasty reviews some ~ 73,000,000
words

Feature Extraction
The preprocessed data set has many special properties. In feature extraction method, we extract
facets from the processed data set. This aspect is then used to calculate the negative and positive
polarity in a sentence, which is useful in determining the opinions of individuals using patterns
like unigram, bigram [15]
Machine learning techniques claim to represent the main features of a text or document to
process. These key features are considered as feature vectors used for the classification task.
Some examples of features that have been reported in the documentation are:
1. Words And Their Frequencies: Unigram, bigram and n-gram samples with their frequency
counts are considered as features. There have been many studies on using word presence
rather than frequency to better describe this characteristic. [16] showed better results when
using presence instead of frequency.
2. Parts Of Speech Tags: Parts of speech such as adjectives, adverbs, and certain groups of
verbs and nouns are good indications of subjectivity and sentimentality. We can generate
syntactic dependency models by parsing or dependency trees.
3. Opinion Words and Phrases: Besides specific words, a number of emotional phrases and
expressions can be used as traits. Example: costs someone an arm and a leg.
4. Position Of Terms: The position of a term in the text can affect the difference between the
term and the overall feeling of the text.
5. Negation: Negation is an important but difficult characteristic to interpret. The presence of a
negative often shifts the polarity of opinion. For example: I am not happy
6. Syntax: Syntactic patterns such as phrases are used by many researchers as traits to learn
subjective patterns.
Supervised learning is an important technique for solving classification problems. Classifier
training facilitates future predictions for unknown data.
1. Naive Bayes Classification: It is a probabilistic classifier and can learn the pattern by
examining a set of classified documents . It compares content with word lists to classify
documents according to their correct category or class.

2. Maximum Entropy Classification: In the Maximum Entropy Classifier, no assumptions are

made about the relationships between the extracted objects from the dataset. This classifier
always tries to maximize the entropy of the system by estimating the conditional distribution
of the class label. The maximum entropy even handles feature overlap and is like the logistic
regression which finds the distribution over classes. The conditional distribution is defined as
MaxEnt makes no independent assumptions for its features, unlike Naive Bayes.

3. Support Vector Machine: Vector machines support data analysis, decision bounds
determination, and kernel use for calculations performed in the input space . The input data
are two sets of vectors of size m each. Then all the data represented by a vector is classified
into a class. Then we find a margin of between the two layers, which is far from any
document. The distance determines the classifier's margin, maximizing the margin reduces
undecided decisions. SVM also supports classification and regression, which is very useful
for statistical learning theory, and it also helps to accurately recognize the factors that need to
be considered in order to understand it successfully.

Approaches for sentiment analysis

The machine learning-based method uses the classification technique to classify the text into
classes. There are mainly two types of machine learning techniques:
1. Unsupervised learning: It does not include a category and they do not provide exact targets
with and are therefore based on grouping.

2. Supervised learning: It is based on a labeled dataset and hence the labels are fed to the model
in the process. These labeled data sets are trained to yield meaningful outputs when
encountered during decision making.
The success of these two learning methods mainly depends on selecting and extracting the
specific set of features used for affective detection. The machine learning method applied to
sentiment analysis mainly belongs to supervised classification. In machine learning
technique, two datasets are required - Training Set, Test Set.
 Several machine learning techniques have been built by to classify tweets into classes.
Machine learning techniques such as Naive Bayes (NB), Maximum Entropy (ME), and
Support Vector Machines (SVM) have achieved great success in sentiment analysis.
Machine learning starts with collecting a set of training data. Then we train a classifier on the
training data. When the supervised classification technique was chosen, an important
decision for the was to choose a feature. They can tell us how the documents are presented.
The most common features used in the sentiment classification are:

o Term presence and their frequency o Part of speech information

o Negations
o Opinion words and phrases

For supervised techniques, Support Vector Machines (SVM), Naive Bayes, Maximum Entropy
are some of the most commonly used techniques.
While semi-supervised and unsupervised techniques are introduced when it is not possible to
have an initial set of labeled documents/assessments to classify the remaining items.

Figure – 2: Sentiment Classification Based On Emoticons

Lexicon-Based Approaches
The lexical-based method uses a sentiment dictionary with opinion words and matches them with
data to determine polarity. They assign sentiment scores to opinion words that describe how
positive, negative, and objective the words in the dictionary are. Vocabulary-based approaches
rely primarily on affective vocabulary, that is, a collection of known, pre-compiled emotional
terms, phrases, and even idioms developed for traditional communication genres, such as
Vocabulary Seeking Opinion.
There are two sub-categories for this approach:
1. Dictionary-based: It is based on the use of commonly collected and annotated terms (seeds).
This set is enriched by looking up synonyms and antonyms from the dictionary. An example
of this dictionary is the WordNet, which was used to develop a synonym called
SentiWordNet. Cons: Can't handle domain and context specific orientations.

2. Corpus-Based: The corpus approach aims to provide dictionaries related to a particular

domain. These dictionaries are created from a set of input opinion terms developed through
the search for related words through the use of statistical or semantic techniques.

Figure – 3: Lexicon-Based Model

o Statistical Method: Latent Semantic Analysis (LSA).

o Semantic-based methods such as the use of synonyms and antonyms or word-synonym
relationships such as WordNet could also represent an interesting solution.
Following performance measures such as accuracy and recall, we provide a comparative study of
existing opinion mining techniques, including machine learning, word-based approaches, and
more. vocabulary, multi-domain and multilingual approaches, etc., as shown in Table 2.
Table 2: Publicly Available Datasets for Twitter.
Method Data Set Accuracy Author
Machine SVM Movie reviews 86.40% Pang, Lee [16]
Learning CoTraining SVM Twitter 82.52% Liu [14]
Deep learning Stanford Sentiment 80.70% Richard [15]
Treebank
Lexical based Corpus Product reviews 74.00% Turkey
Dictionary Amazon’s --- Taboada [17]
Mechanical Turk
Cross-lingual Ensemble Amazon 81.00% Wan,X. [18]
Co-Train Amazon, ITI68 81.30% Wan,X. [18]
EWGA IMDb movie review >90% Abbasi,A.
CLMM MPQA, NTCIR, ISI 83.02% Mengi
Cross-domain Active Learning Book, DVD, 80% (avg) Li, S
Thesaurus Electronics, Bollegala[22]
SFA Kitchen Pan S J [14]

Sentiment Analysis Tasks

Sentiment analysis is a challenging interdisciplinary task that includes natural language
processing, web crawling, and machine learning. This is a complex task and can be divided into
the following tasks, namely:
• Subjective Classification
• Sentiment Classification
• Complimentary Tasks o Object Holder Extraction o Object/Feature Extraction

Figure – 4: Sentiment Analysis Tasks

Levels of Sentiment Analysis

The tasks described in the previous section can be performed at several levels of detail.
 Figure – 5: Levels of Sentiment Analysis

Document level refers to marking individual documents with their sentiment. At the document
level, the entire document is classified as either positive or negative at the grade level. General
Approach is to find the poles of each phrase or word in and combine them to find the pole of
document & other approach is Complex linguistic phenomena such as resolution of co-
references, pragmatics, etc. The various tasks involved in this are: Task: Emotions Categorize
entire document Categories: Positive, negative, and neutral Assumption: Each document focuses
on a single audience (this is not true in discussion posts, blogs, etc.) and contains comments by a
single opinion holder
Sentence-level analysis deals with tagging individual sentences with their respective emotional
poles. Sentence level classification classifies sentences into levels of positive, negative, or
neutral. General Approach is to find the sentimental orientation of each word in the
phrase/sentence, then combine them to determine the sentiment of the phrase or the whole phrase
& other approach is examining the discursive structure of text, the various tasks involved in this
are Task 1: Determine Subjective/Objective Sentence Type: Objective and Subjective Task 2:
Categorize Sentiment Sentence Type: Positive and Negative Hypothesis: A sentence contains
only one opinion may not always be true.
Aspect Level or Functional Level This involves labeling each word with their emotion and
identifying the entity to which the sentiment is directed. Sentiment classification problems at the
aspect or function level with identifying and extracting product features from source data.
Techniques such as dependency parsing, and discourse construction are used for this purpose.
Various Tasks involved in this are Task 1: Identify and extract the characteristics of an object
that have been commented on by an opinion holder (e.g., reviewer) Task 2: Determine if the
opinion about the object is negative, positive, or neutral Task 3: Look up synonyms for the
object.
Most recent works have used pre-sentence-word polarization to classify sentiments at the
sentence level and the document level. Classification of feelings from mainly using adjectives as
characteristics but adverbs, the two automatic word-level sentiment annotation methods are
Dictionary-based approach & The corpus-based approach.
Data Preparation
Since this dataset is quite large, during the exploration we start by generating sub-datasets during
training, to speed up the testing steps. We also re-index the dataframe, to make it easier to add
columns later in the process.
Here we explore the data, check if it is balanced and check for patterns in missing rows. This can
usually be done automatically with pandas config. Sections in this title include:
• Basic visualization - We can display some basic statistics about the data using pandas and
also display a few entries from the dataset, to see the sample points we will be working with.
• Automated data exploration with pandas configuration - Pandas-profiling is a library used to
automate data discovery. This gives us a good overview of the dataset, which we can use to
inform our further work.
• Checking balances in output categories - We want to check the balance of the output column
(Sentiment), so we don't train a model that always predicts the output. This model can be
highly accurate, but we won't learn anything about trends in the data other than numbers from
a general point of view. You can also think about balancing test data, but keep in mind that
real world data can hardly be well balanced and experimental data is like real world data
Sentiment ratios change d' by about 0.15 to 0.3, which is usually a good balance, so we
are unlikely to see a scenario where only one class is predicted. However, we will be
watching to see if our training accuracy gets around 0.3, which could be a sign of this
problem.
We can create an indexer to convert the perception from label to index and vice versa. This is
very useful for understanding our predictions later. Then we convert the "Opinion" column in the
training data into labels, we will learn how to predict. After that we work with the text in the
"Tweet_Content" column, extract all the information we can and convert it into a format usable
for the neural network that we will train later. This involves removing words with little meaning
(keyword removal) and grouping words with the same meaning regardless of details such as time
(lemmatization). Next, we use encryption to encode the presence of words in the matrix, similar
to hot encoding. This is called the "bag of words" method. Sections in this title include:

• Stop Word Removal and Lemmatization with NLTK - Here, we first split each string into its
individual words, before checking if these strings - Contains text, yes in keyword list If no
text in word, meaning is, has only digits or punctuation (or other characters), or the word is a
stop word (words like "with", "a", "the"), the word is removed from the string. We also
perform lemma in this step, where we convert words back to their original form, so tenses
and other details can be omitted in our final model (a negative past tense statement is always
negative).
• Tokenization - We generate tokens for the most common words in the dataset, so that we can
represent the presence of words in our generated corpus (n most common words) with a list
of integers.
• Adding the Tokenized Strings to the DataFrame

The data is now almost ready for a trained model, but some final preparation will need to take
place. For example, we need to remove columns that we do not intend to use, such as the
"Tweet_Content" column, from which useful information has been extracted. We also split the
data into a training and test set so that we can evaluate the performance of our model without
touching the retained data. We do this because if we constantly examine this retained data, it
loses its usefulness as invisible “real world” data. Sections in this heading include Clear unused
data, Test stream separation.

• Dropping Unused Data - Here we remove unnecessary columns from the DataFrame. They
have no use value (Tweet ID) or useful information has been extracted (Tweet Content).
We're also removing "y" or the dependent variable here, so we don't accidentally train on it.
• Test-Train Split - Here, we use SciKit-Learn's built-in functionality to split our data into a
test set and a train set, with the appropriate labels. We use a constant random state to make
this repeatable.

Model Construction and Training

Finally, it's time to build our model. In this case, we use a neural network built with Keras. Then
we train it with our data in the training dataset and validate using the test dataset.
• Model Construction - Here we define the neural network that we will train to predict the
output. This model is built with the following classes:
o Embedding - The embedding layer is a very important part of this process, as it allows us
to learn the meaning of words, based on the context of other words around them. In
effect, this class places each word in a vector space, in which we will learn where and
what words can be used around it.
o Bidirectional - These layers are improvements to cyclic neural networks (RNNs), where
data flows in a certain order so that we can learn from the order of the text. In two-
dimensional layers, this happens in both directions, so we can try to find out the meaning
of words in context both forward and backward. We use LSTMs (long-term memory
nodes) in this network, so our network can "remember" the previous context. LSTMs
perform better than RNNs for applications such as word processing because they solve
the problem of gradients disappearing by allowing gradients to pass through unchanged.
o Dense - The last layer of the network is a dense layer. It just means that all the nodes in
this layer are connected to all the nodes in the previous layer, so the input of any LSTM
unit can
be included. Here we use softmax exit function and 4 nodes to generate the probability
for each potential exit (each possible sentiment).
• Training - We then adapted this model to our data, using backpropagation, over 15 epochs.
We can see an increase in model accuracy over different epochs, both on training and test
datasets. Now that we have trained the model, we can see its accuracy with a confusion
matrix. This allows us to see predictions for Tweets with different real values. From that, we
can see that we predict some classes better than others. Now that we are satisfied with our
model, we can train using the full dataset and predict the retained test data. This involves
performing all of our transformations on both this training dataset and the retained test data.
Fortunately, we can reuse the above code to achieve this, so a bit more explanation is needed.
There are a few potential reasons for this. First, we don't really know how the Tweets in the
workout data are classified. Maybe a bag-of-words classifier was used, which means we can just
reproduce it. Next steps here might be to look at the training and test data to see if that's the case
and see if the models perform better on the changed data.
In addition, we need to know the limitations of our model. Two-way LSTM models often require
longer pieces of text to learn the context of words, which is not possible in a Tweet. Where
bidirectional LSTM models really shine is in longer pieces of text and perform much better than
wordfor-word classifiers for tasks like evaluation classification. As a result, we generally don't
expect great performance from this type of model on Tweets, where word bags can often work
very well, due to this lack of context.
Conclusion
It turns out that we can classify Tweets pretty well with a bag of "standard" word classifications.
Here we can assume that this is largely due to a lack of context in Tweets, meaning that more
complex sentence structures are difficult to display.
If we have a connection to any of the entities mentioned in the classifier, we may want to
configure it to automatically detect people's opinions about our products or services. This can
provide very valuable insights into what people are thinking without having to manually extract
data anymore.
Testing more advanced methods, such as a pre-trained Bidirectional Encoder Representation
from a Transformer (BERT) model may yield better results than what we get here
Reference
[1] A.Pak and P. Paroubek. „Twitter as a Corpus for Sentiment Analysis and Opinion Mining".
In Proceedings of the Seventh Conference on International Language Resources and
Evaluation, 2010, pp.1320-1326
[2] R. Parikh and M. Movassate, “Sentiment Analysis of User- Generated Twitter Updates using
Various Classi_cation Techniques",CS224N Final Report, 2009
[3] Go, R. Bhayani, L.Huang. “Twitter Sentiment Classification Using Distant Supervision".
Stanford University, Technical Paper,2009
[4] L. Barbosa, J. Feng. “Robust Sentiment Detection on Twitter from Biased and Noisy Data".
COLING 2010: Poster Volume,pp. 36-44.
[5] Bifet and E. Frank, "Sentiment Knowledge Discovery in Twitter Streaming Data", In
Proceedings of the 13th International Conference on Discovery Science, Berlin, Germany:
Springer,2010, pp. 1-15.
[6] Agarwal, B. Xie, I. Vovsha, O. Rambow, R. Passonneau, “Sentiment Analysis of Twitter
Data", In
Proceedings of the ACL 2011Workshop on Languages in Social Media,2011 , pp. 30-38
[7] Dmitry Davidov, Ari Rappoport." Enhanced Sentiment Learning Using Twitter Hashtags and
Smileys". Coling 2010: Poster Volume pages 241{249, Beijing, August 2010
[8] Po-Wei Liang, Bi-Ru Dai, “Opinion Mining on Social MediaData", IEEE 14th International
Conference on Mobile Data Management, Milan, Italy, June 3 - 6, 2013, pp 91-96, ISBN:
978-1494673-6068-5, http://doi.ieeecomputersociety.org/10.1109/MDM.2013.
[9] Pablo Gamallo, Marcos Garcia, “Citius: A Naive-Bayes Strategyfor Sentiment Analysis on
English Tweets", 8th International Workshop on Semantic Evaluation (SemEval 2014),
Dublin, Ireland, Aug 23-24 2014, pp 171-175.
[10] P. D. Turney, “Thumbs up or thumbs down?: semantic orientation applied to
unsupervised classification of reviews,” in Proceedings of the 40th annual meeting on
association for computational linguistics, pp. 417–424, Association for Computational
Linguistics, 2002.
[11] J. Kamps, M. Marx, R. J. Mokken, and M. De Rijke, “Using wordnet to measure
semantic orientations of adjectives,” 2004.
[12] R. Xia, C. Zong, and S. Li, “Ensemble of feature sets and classification algorithms for
sentiment classification,” Information Sciences: an International Journal, vol. 181, no. 6, pp.
1138–1152, 2011.
[13] ZhunchenLuo, Miles Osborne, TingWang, An effective approach to tweets opinion
retrieval", Springer Journal on World Wide Web, Dec 2013, DOI: 10.1007/s11280-013-0268-
7.
[14] Liu, S., Li, F., Li, F., Cheng, X., &Shen, H.. Adaptive co-training SVM for sentiment
classification on tweets. In Proceedings of the 22nd ACMinternational conference on
Conference on information & knowledge management (pp. 2079-2088). ACM,2013.
[15] Socher, Richard, et al. "Recursive deep models for semantic compositionality over a
sentiment Treebank." Proceedings of the Conference on Empirical Methods in Natural
Language Processing(EMNLP). 2013.
[16] Pang, B.and Lee, L. “A sentimental education: Sentiment analysis using subjectivity
summarization based on minimum cuts”. 42nd Meeting of the Association for Computational
Linguistics[C] (ACL-04). 2004, 271-278.
[17] Taboada, M., Brooke, J., Tofiloski, M., Voll, K., &Stede, M..“Lexicon based methods for
sentiment analysis”. Computational linguistics, 2011:37(2), 267-307.
[18] Wan, X..“A Comparative Study of Cross-Lingual Sentiment Classification”. In
Proceedings of the 2012 IEEE/WIC/ACMInternational Joint Conferences on Web
Intelligence and Intelligent Agent Technology-Volume 01 (pp. 24-31). IEEE Computer
Society.2012
[19] Bollegala, D., Weir, D., & Carroll, J.. Cross-Domain Sentiment Classification using a
Sentiment Sensitive Thesaurus. Knowledge and Data Engineering, IEEE Transactions on,
25(8), 17191731,2013