Source-linked AI summary
Sentiment Analysis of Twitter Data: A Survey of Techniques
Vishal. A. Kharde, Prof. Sheetal. Sonawane
TL;DR
The paper addresses how to analyze large volumes of heterogeneous Twitter opinions. It surveys machine-learning and lexicon-based techniques and compares them with evaluation measures, reporting 77.73 accuracy for its best-tested SVM feature set.
Problem
Large volumes of heterogeneous Twitter opinions require sentiment analysis to classify positive, negative, or neutral content.
Method
The paper surveys machine-learning and lexicon-based sentiment approaches, using preprocessing, feature extraction, classifier training, and semantic analysis on a labeled Stanford Twitter dataset.
Results
77.73 accuracy was obtained with SVM using unigram and bigram features with stopwords removed; Naive Bayes reached 76.44 and MaxEnt 74.93 in the reported comparison.
Takeaways & Limitations
The survey concludes that SVM and Naive Bayes provide high-accuracy baseline methods, while lexicon-based methods can be effective with fewer human-labeled documents.
Abstract
from arXiv · showhide
With the advancement of web technology and its growth, there is a huge volume of data present in the web for internet users and a lot of data is generated too. Internet has become a platform for online learning, exchanging ideas and sharing opinions. Social networking sites like Twitter, Facebook, Google+ are rapidly gaining popularity as they allow people to share and express their views about topics,have discussion with different communities, or post messages across the world. There has been lot of work in the field of sentiment analysis of twitter data. This survey focuses mainly on sentiment analysis of twitter data which is helpful to analyze the information in the tweets where opinions are highly unstructured, heterogeneous and are either positive or negative, or neutral in some cases. In this paper, we provide a survey and a comparative analyses of existing techniques for opinion mining like machine learning and lexicon-based approaches, together with evaluation metrics. Using various machine learning algorithms like Naive Bayes, Max Entropy, and Support Vector Machine, we provide a research on twitter data streams.General challenges and applications of Sentiment Analysis on Twitter are also discussed in this paper.
1. INTRODUCTION
Sentiment analysis addresses the challenge of extracting and classifying opinions, sentiments, and emotions from the large volume of user-generated online content. Twitter research has explored binary and three-way polarity classification using diverse machine-learning, feature-based, and lexicon-based methods.
- Motivation: Online platforms generate vast amounts of sentiment-rich user content that is difficult for individuals to analyze manually.This motivates automating sentiment analysis for reviews, tweets, comments, and other online sources.
- Scope: Sentiment analysis mines attitudes, opinions, views, and emotions from text and classifies them as positive, negative, or neutral.It is also associated with subjectivity analysis, opinion mining, and appraisal extraction.
- Opinion representation: An opinion can be represented by an object, feature, polarity, holder, and expression time.The object may be a person, event, product, organization, or topic, while the feature is the evaluated attribute.
- Prior Twitter research: Twitter sentiment research progressed from binary positive-versus-negative classification toward objective, positive, and negative classes.Several surveyed studies formulate three-way classification using Twitter corpora and different feature representations.
- Prior Twitter research: Prior studies used Naive Bayes, Maximum Entropy, SVM, stochastic gradient descent, Hoeffding trees, nearest neighbors, and tree-kernel models.Reported findings favored SVM, appropriately tuned SGD, tree kernels, or Naive Bayes depending on the study and setup.
- Lexicon-oriented methods: Lexicon-oriented studies include bag-of-words aggregation and WordNet-based semantic polarity or emotional-content measures.These approaches determine sentiment from word-level scores or lexical relationships rather than only trained classifiers.
3. APPROACHES FOR SENTIMENT ANALYSIS
Twitter sentiment analysis approaches primarily use supervised or unsupervised machine learning, alongside lexicon-based methods that infer polarity from sentiment dictionaries. The survey compares these techniques using evaluation measures such as precision and recall.
- Machine Learning Approaches: Machine-learning approaches classify text into sentiment classes using training data and selected textual features.The workflow collects data, trains a classifier, and represents documents through features such as term frequency, part of speech, and opinion phrases.
- Machine Learning Approaches: Supervised learning uses labeled datasets, whereas unsupervised learning lacks target categories and relies on clustering.Semi-supervised and unsupervised methods are relevant when labeled opinions are unavailable.
- Machine Learning Approaches: SVM, Naive Bayes, and Maximum Entropy are among the most common supervised techniques for sentiment classification.The surveyed literature reports substantial success for these models in sentiment analysis.
- Lexicon-Based Approaches: Lexicon-based methods match texts against sentiment dictionaries and assign polarity scores to opinion words.Lexicons may contain sentiment terms, phrases, and idioms developed for particular communication genres.
- Lexicon-Based Approaches: Dictionary-based lexicons expand manually annotated seed terms through synonyms and antonyms, but may not handle domain- or context-specific orientations.WordNet and SentiWordNet are cited as examples of this approach.
- Lexicon-Based Approaches: Corpus-based lexicons grow domain-specific dictionaries from seed opinion terms using statistical or semantic techniques such as LSA or thesaurus relations.The survey compares lexicon-based, machine-learning, cross-domain, and cross-lingual approaches using precision and recall.
4. SENTIMENT ANALYSIS TASKS
Sentiment analysis is presented as a challenging interdisciplinary task involving natural language processing, web mining, and machine learning. Its main tasks include subjectivity classification, sentiment classification, and object or feature extraction.
- Sentiment analysis combines natural language processing, web mining, and machine learning.
- The section identifies subjectivity classification, sentiment classification, and object or feature extraction as core tasks.
A. Subjectivity classification
Subjectivity classification distinguishes opinionated sentences from sentences that present factual information.
- Subjectivity classification identifies sentences that express opinions or other forms of subjectivity.
- The task separates subjective sentences from objective sentences in a document.
B. Sentiment Classification
Sentiment classification determines the polarity of opinionated text after subjectivity has been identified. It can use binary, multi-class, regression, or ranking formulations.
- Sentiment classification determines whether a sentence expresses a positive or negative opinion.
- The task may be formulated as binary classification, multi-class classification, regression, or ranking.
- Multi-class classification can distinguish extremely negative, negative, neutral, positive, and extremely positive categories.
- Opinion-holder and object-feature extraction may be treated as optional subtasks depending on the application.
C. Complimentary Tasks
Complementary tasks and analysis levels extend sentiment analysis beyond polarity classification. Document-, sentence-, and aspect-level approaches differ in their targets, procedures, and assumptions.
- Complementary Tasks: Opinion-holder extraction discovers the direct or indirect sources of opinions.
- Complementary Tasks: Object or feature extraction discovers the target entity toward which sentiment is directed.
- Document Level: Document-level analysis assigns sentiment to an entire document, typically by combining the polarities of individual sentences or words.
- Document Level: Document-level analysis commonly uses positive, negative, and neutral classes but assumes one object and one opinion holder per document.
- Sentence Level: Sentence-level analysis identifies subjectivity and assigns positive, negative, or neutral polarity to individual sentences.
- Aspect or Feature Level: Aspect-level analysis extracts commented-on features, determines their polarity, and finds feature synonyms using dependency parsing and discourse structures.
6. EVALUATION OF SENTIMENT CLASSIFICATION
The paper evaluates sentiment classification using four indexes derived from confusion-matrix outcomes. TP, FN, FP, and TN denote the corresponding instance counts.
- Sentiment classification performance is evaluated using four indexes calculated from confusion-matrix outcomes.
- TP, FN, FP, and TN represent true positive, false negative, false positive, and true negative instances, respectively.
7. RESULTS AND DISCUSSION
The experiments compare Naïve Bayes configurations using preprocessing and n-gram features, with stopword removal and bigrams improving accuracy while trigrams add little benefit. Results are reported across dataset sizes and feature settings.
- The experiments use a publicly available Stanford Twitter dataset, preprocessing raw sentences before feature-vector training and semantic polarity analysis.
- Baseline and preprocessing: 73.65 percent accuracy was obtained by the Naïve Bayes baseline without preprocessing and using unigrams.
- Baseline and preprocessing: 74.56 percent accuracy followed stopword removal with preprocessed data and a unigram model.
- Preprocessing effects: Stopword removal changed predictions because stopwords are frequently used in short tweets.
- N-gram features: 76.44 percent accuracy was achieved with Naïve Bayes using stopword removal and bigram features, compared with 74.56 percent for unigrams.
- N-gram features: 75.41 percent accuracy from combining trigrams, bigrams, and unigrams was lower than the bigram result, while the larger feature space slowed execution.
C. Support Vector Machine (SVM):
The SVM experiments compare unigram and bigram configurations with other classifiers and feature settings. SVM achieved the highest reported accuracy, while Maximum Entropy performed between Naïve Bayes and SVM but required longer execution.
- SVM evaluation: The SVM experiments evaluate unigram features and summarize accuracy results for unigram and bigram configurations.
- Classifier comparison: Maximum Entropy achieved 74.93 percent accuracy with a unigram model.
- Classifier comparison: 77.73 accuracy was the maximum reported result, obtained by SVM with slang stopword removal and bigram features.
- Classifier comparison: SVM outperformed Naïve Bayes and Maximum Entropy across the considered feature settings.
- Classifier comparison: Maximum Entropy performed between Naïve Bayes and SVM and was not used for all feature combinations because its iterative execution took substantial time.
- Dataset-size comparison: On the bigger dataset, Naïve Bayes reached 76.44 percent and SVM reached 77.73percent.
8. CHALLENGES IN SENTIMENT ANALYSIS
Sentiment analysis on Twitter is challenged by linguistic ambiguity, context dependence, discourse structure, entity-specific opinions, limited language coverage, and restricted access to some social-media data.
- Subjectivity and context: Subjective content is difficult to identify because the same word may express sentiment in one context but factual information in another.“Crude” is evaluative in a description of a person but objective when referring to crude oil.
- Subjectivity and context: Domain dependence changes polarity: “unpredictable” can be positive for movies but negative for vehicle steering.
- Linguistic phenomena: Sarcasm can express negative sentiment through positive words, making surface-level polarity unreliable.The example “Nice perfume. You must shower in it.” contains positive words but conveys a negative sentiment.
- Linguistic phenomena: Negation and thwarted expressions can reverse or complicate the sentiment suggested by individual words or clauses.“Avoids” negates negative terms, while positive-looking movie descriptions can culminate in an overall negative sentiment.
- Structure and relations: Discourse order and comparisons require relational interpretation that bag-of-words models do not capture reliably.Reversing “A is better than B” changes the opinion, while “better” can incorrectly assign positive sentiment to both entities.
- Structure and relations: Entity recognition is necessary because a single tweet may express different sentiments toward different entities, unlike a single neutral bag-of-words label.The example praises Linux while criticizing Microsoft.
- Scope and data: Research needs to distinguish sentiment-bearing tweets from non-sentimental tweets, while broader deployment is constrained by limited Facebook access and English-focused work.
9. APPLICATIONS OF SENTIMENT ANALYSIS
Sentiment analysis supports review mining, recommendations, moderation, business feedback, trend prediction, and applications across research and emerging technology domains. The survey compares machine-learning and lexicon-based approaches and reports that SVM and Naive Bayes often provide high accuracy, while cleaner data and bigrams improve results.
- Review and business applications: Review-mining systems can extract opinions about products, services, and political issues to automate feedback or ratings for users and vendors.
- Review and business applications: Sentiment predictors can support recommender systems by avoiding items with substantial negative feedback or few ratings.
- Moderation and prediction: Highly negative sentiment can help identify abusive language and other negative elements in online communication for corresponding action.
- Review and business applications: Businesses use online sentiment to assess product success, improve products and reputation, and support customer satisfaction.
- Cross-domain applications: Sentiment analysis can reveal trends in human emotions across sociology, medicine, sports, and social-media research.
- Cross-domain applications: In IoT smart homes, detected user sentiment could guide changes to home ambiance toward a soothing and peaceful environment.
- Moderation and prediction: Tracking public views can support trend prediction by extracting information about sales trends and customer satisfaction.
- Survey findings: The survey compares machine-learning, lexicon-based, cross-domain, and cross-lingual methods using evaluation metrics.It reports SVM and Naive Bayes as high-accuracy baseline methods, lexicon approaches as effective in some cases, and cleaner data and bigrams as beneficial.