PaperlayerView exact revision on arXiv

Source-linked AI summary

Sentiment Analysis: Automatically Detecting Valence, Emotions, and Other Affectual States from Text

Saif M. Mohammad

arXiv:2005.11882v2cs.CL

TL;DR

Sentiment analysis seeks to automatically detect valence, emotions, and other affectual states despite contextual, linguistic, data, and fairness challenges. This article provides a comprehensive survey of the field’s tasks, methods, resources, applications, and ethical concerns, concluding that substantial progress remains necessary while highlighting risks of bias and misuse.

  • Problem

    Sentiment analysis must handle subtle, contextual, figurative, and unstated emotion, while supervised systems face expensive annotation, subjective judgments, and incomplete para-linguistic information.

  • Method

    The article synthesizes research on sentiment-analysis tasks, machine-learning methods, emotion resources, modifiers, figurative language, applications, and fairness.

  • Results

    The survey identifies broad progress across sentiment-analysis approaches and resources while documenting persistent challenges in language interpretation, annotation, stance detection, and sarcasm.

  • Takeaways & Limitations

    Sentiment analysis supports applications across commerce, public health, policy, social science, and art, but its use requires attention to fairness and harmful misuse.

  • Takeaways & Limitations

    Sentiment analysis can expose individuals and populations to manipulation, deception, and indoctrination, while systems may perpetuate human biases.

Abstract

from arXiv · show

Recent advances in machine learning have led to computer systems that are human-like in behaviour. Sentiment analysis, the automatic determination of emotions in text, is allowing us to capitalize on substantial previously unattainable opportunities in commerce, public health, government policy, social sciences, and art. Further, analysis of emotions in text, from news to social media posts, is improving our understanding of not just how people convey emotions through language but also how emotions shape our behaviour. This article presents a sweeping overview of sentiment analysis research that includes: the origins of the field, the rich landscape of tasks, challenges, a survey of the methods and resources used, and applications. We also discuss discuss how, without careful fore-thought, sentiment analysis has the potential for harmful outcomes. We outline the latest lines of research in pursuit of fairness in sentiment analysis.

1. Introduction

Sentiment analysis is a broad field for automatically detecting valence, emotions, and other affectual states, emerging from early work on reviews and subjective language. This overview surveys its history, tasks, challenges, methods, resources, applications, and fairness concerns.

  • Scope and origins: Sentiment analysis automatically determines valence, emotions, and other affectual states from text or speech.It can assess polarity, ratings, attitudes toward targets, or a speaker’s emotional state.
  • Scope and origins: Early research focused on valence in customer reviews and separating affective or subjective text from factual text.The field’s platform included growing digital text availability and earlier psychological research on emotion and language.
  • Tasks and challenges: The overview organizes sentiment analysis around diverse tasks, including entities, textual units, stance, product aspects, and semantic roles of feelings.It also discusses challenges such as subtle expression, creative language, annotated-data requirements, and missing para-linguistic context.
  • Methods and resources: The survey covers machine-learning approaches, emotion lexicons, sentiment modifiers, figurative language, and sentiment analysis beyond English.Topics include transfer learning, deep neural networks, dense representations, influential resources, and resource-poor languages.
  • Applications and fairness: Applications span commerce, intelligence gathering, policymaking, public health, and art, but sentiment systems can perpetuate inappropriate human biases.The overview therefore includes fairness and bias in sentiment analysis.

2. Challenges in Sentiment Analysis

Sentiment analysis is difficult because emotional meaning depends on context, composition, modifiers, figurative language, and unstated or para-linguistic information. Limited labeled data, subjective judgments, and cultural variation further constrain reliable detection.

  • Complexity and subtlety of language use: Emotional meaning is not simply the sum of word associations, and emotions may be conveyed without explicit affective words.For example, a sentence about working one’s tail off can imply frustration without overtly negative vocabulary.
  • Complexity and subtlety of language use: Negations, modals, degree adverbs, and other modifiers can change a phrase’s sentiment despite having weak sentiment associations themselves.Systems must distinguish expressions such as “may be good,” “was good,” and “was not good.”
  • Complexity and subtlety of language use: Words can express different emotions across contexts and senses, making word-sense disambiguation a continuing challenge.The word “hug” can signal joy and affection in one sense but be unemotional in another context.
  • Complexity and subtlety of language use: Utterances may express multiple emotions, contrastive evaluations, or emotional events without revealing the speaker’s own feelings.These cases require distinguishing event sentiment from the sentiment of the person producing the utterance.
  • Creative and non-standard language: Sarcasm, irony, humour, metaphor, misspellings, creative spellings, hashtags, emoticons, and abbreviations complicate automatic interpretation.Social-media lexicon methods address some non-standard terms by learning sentiment associations from social-media data.
  • Lack of para-linguistic information: Written text usually lacks tone, pitch, emphasis, and facial-expression information that can communicate affect.Explicit emphasis markers, sentiment markers, emoticons, and emoji compensate for this absence to some degree.
  • Labeled data and subjectivity: Large labeled datasets are expensive, so research often addresses only a small subset of the many emotions humans can perceive and express.Annotation quality and agreement also vary with instructions, task difficulty, respondent training, annotation scheme, and culture.
  • Open challenges: Despite approaches that address these issues, significant progress remains to be made.

3. Sentiment Analysis Tasks

Sentiment analysis spans multiple targets, text levels, and semantic tasks, so systems must specify what sentiment is being inferred and toward whom. The section also covers lexicons, aspect sentiment, stance, and emotion roles, while highlighting challenges from biased streams and incomplete intent information.

  • 3.1 Detecting Sentiment of the Writer, Reader, and other Entities: Sentiment systems must distinguish whether they infer the writer’s, reader’s, or a mentioned entity’s sentiment.The same utterance may describe a negative event without showing that the writer personally feels sadness, while readers can react differently to identical text.
  • 3.2 Detecting Sentiment from Different Textual Chunks: Sentiment can be analyzed from words and phrases through sentences, reviews, documents, and streams over time.The appropriate textual unit depends on the intended task, from individual lexical associations to aggregate emotional trends.
  • 3.2 Detecting Sentiment from Different Textual Chunks: Valence and emotion lexicons record words’ polarity or associations with selected emotions, using manual or automatic construction.Automatically generated lexicons can cover hundreds of thousands of unigrams and larger expressions, often with real-valued association strengths.
  • 3.2 Detecting Sentiment from Different Textual Chunks: Aggregate sentiment from social streams is vulnerable to topic drift and posting biases that favor some experiences and high-arousal content.Tracking changing topics requires continually updating search terms, while positive self-presentation, negative product reporting, and virality can distort observed emotion.
  • 3.3.1 Detecting Sentiment Towards Aspects of an Entity or Aspect Based Sentiment Analysis (ABSA): Aspect-based sentiment analysis identifies sentiment toward particular product or service aspects rather than treating the review as a single sentiment.Strong systems combined sentence-level sentiment models with localization features that emphasized sentiment near aspect mentions.
  • 3.3.2 Detecting Stance: Stance detection infers a tweeter’s position toward a target and may require information absent from the focus text.In the first shared task, none of about twenty teams surpassed the organizers’ simple support vector machine baseline using word and character n-grams.
  • 3.4 Detecting Semantic Roles of Emotion: Emotion analysis can be enriched by semantic roles, while emotion labels alone may fail to distinguish different communicative intents.Frame-based analysis identifies roles such as the emotion experiencer and the person or event evoking the emotion; disgust may signal criticism, venting, or ridicule.

4. Detecting Valence and Emotions in Sentences and Tweets

Valence and emotion detection spans diverse text sources and targets, using supervised learning alongside evolving representations and resources. The section also highlights annotation costs, domain dependence, distant supervision, and unresolved questions about how emotions are conveyed.

  • Sentiment analysis systems address diverse texts and targets, including writer, reader, entities, words, sentences, documents, and tweets.
  • Supervised systems learn emotion labels from annotated training data and evaluate predictions on held-out labeled test sets.
  • Since the 2010s, deep learning has favored low-dimensional continuous embeddings, while transfer learning helps address their need for very large training datasets.
  • Traditional models and features remain relevant because of their simplicity, interpretability, ease of use, and potential complementarity with neural representations.
  • Large emotion datasets are expensive to annotate, and important questions remain about implicit emotion expression, emotion taxonomies, basic emotions, and transfer across emotion categories.
  • Domain-specific systems are more accurate when trained on target-domain data, while distant supervision can produce emotion-labeled tweets comparable to explicitly annotated data.

5. Capturing Term–Sentiment Associations

Term–sentiment associations range from contextual majority tendencies to graded, crowdsourced, and automatically induced lexicons. These resources support large-scale analysis while capturing domain-specific vocabulary and varying association strength.

  • A word’s sentiment can vary by context, although many words show a dominant sentiment association across most uses.
  • Crowdsourced lexicons include binary emotion associations, real-valued valence-arousal-dominance scores, and emotion-intensity scores across large English vocabularies.
  • Real-valued sentiment strength is useful downstream, but direct rating scales are difficult for annotators because scores vary across people and over time.
  • Best–Worst Scaling asks annotators to select the most and least positive item among four or five items, producing rankings and real-valued scores.
  • BWS produces more reliable and discriminating scores than rating scales and supports resources such as NRC-VAD and NRC-EIL.
  • Emotion lexicons provide simple, effective tools for analyzing large text collections and can improve machine-learning prediction when training data is limited.
  • Automatic methods learn associations for hundreds of thousands of words and sequences, often improving coverage and adapting to domain-specific language.

6. Modeling the impact of sentiment modifiers

Sentiment modifiers alter the meaning or intensity of associated expressions, making compositional handling necessary for accurate analysis. Research examines negation, modality, degree adverbs, and intensifiers through annotation and modeling approaches.

  • Best–Worst Scaling annotations show that modifier effects vary substantially within groups and across individual modified words.
  • 6.1 Negation: Negation handling requires identifying negators, determining their scope, and modeling their effect on sentiment.
  • 6.1 Negation: Simple polarity-flipping and fixed-constant heuristics fail to capture negators’ true effects on words within their scope.
  • 6.1 Negation: Negators often make positive words negative with lower intensity and make negative words less negative rather than positive.
  • Modals can change the sentiment of an expression, so distinguishing combinations such as cannot work from work or will work matters for analysis.

7. Sentiment in Sarcasm, Metaphor, and other Figurative Language

Figurative language, including sarcasm, irony, metaphors, and idioms, creates distinctive challenges for sentiment analysis because literal wording may not express affect directly. Studies show that sarcasm can sharply reduce system performance, while metaphorical meaning derives from interaction between domains.

  • Sarcasm and irony: The 2015 SemEval task measured sentiment in metaphor- and irony-rich tweets on a -5 to +5 scale, whose data were mostly negative.The task targeted degree of sentiment rather than only positive, negative, or neutral classes.
  • Metaphor: Metaphorical expressions were significantly more emotional on average than their literal counterparts.The emotional content resulted from meaning composition and interaction between the metaphor’s source and target domains, rather than simple transfer from source to target.
  • Idioms: Automatic sentiment systems often struggle with idioms, although idiom-related features can improve sentiment analysis in English texts.The cited studies connect the difficulty to the need for language-specific handling of idiomatic expressions.

8. Multilingual and Crosslingual Sentiment Analysis

Multilingual and crosslingual sentiment analysis addresses the scarcity and smaller scale of affect-related resources outside English. Research transfers information across languages through annotated corpora, lexicons, translation, multilingual representations, and parallel data, while several preservation and adaptation problems remain.

  • Resource imbalance: Non-English sentiment systems tend to be less accurate because languages other than English have fewer and smaller affect-related resources.These resources include sentiment lexicons and emotion-annotated corpora.
  • Crosslingual methods: Crosslingual methods leverage source-language annotated corpora, bilingual parallel data, sentiment lexicons, multilingual WordNet, and multilingual embeddings.These resources support transferring sentiment information between languages.
  • Translation-based transfer: English sentiment analysis achieved only slightly lower accuracy on English translations of Arabic text than Arabic sentiment analysis on the original text.The comparison used manually determined sentiment of the Arabic text as the benchmark.
  • Translation-based transfer: Automatically translated English valence lexicons improved the accuracies of an Arabic sentiment analysis system.The result supports translation as one way to expand Arabic sentiment resources.
  • Open challenges: Less-explored problems include preserving sentiment degree, handling cross-language modifiers, comparing manual and automatic translation, and translating figurative language without losing affect.These issues define important boundaries for translation-based sentiment analysis.

9. Applications

Sentiment and emotion detection is applied across commerce, politics, public health, education, social media, art, digital humanities, personality research, social-group analysis, and interactive visualization. These applications use affective text analysis to monitor trends, support decisions, study people and texts, and enable emotionally responsive systems.

  • Commerce and finance: Sentiment analysis supports brand management, customer response tracking, and automated dialogue systems for customer queries and complaints.The cited applications use blogs, tweets, and Facebook posts.
  • Art: TransProse uses the NRC Emotion Lexicon to generate music that captures emotions in literature.The resulting music was performed by a symphony orchestra and a human composer at the Louvre museum in Paris.
  • Education, culture, and society: Applications also include emotion-aware tutoring, disaster-management analysis, literary and film analysis, personality-trait detection, and comparisons across social groups.These systems examine emotional states, emotional diffusion, narrative patterns, personality indicators, and differences in language across groups.
  • Politics: Political applications track sentiment toward politicians, electoral issues, and events while identifying contentious issues and voter polarization.Social media, especially Twitter, is a major source for this work.
  • Public health and psychology: Emotion detection is used to detect depression, identify cyber-bullying, predict community-level health attributes, and track well-being.Related work also examines gender differences in connotative word meaning.
  • Assistive systems: Emotion-sensitive robots are being explored for elderly, disabled, and sick patients.The systems are intended to respond to the patient’s emotional state.
  • Visualization: Interactive emotion visualizations let users select entities, emotions, or time frames and inspect relevant information in greater detail.Interactivity is presented as supporting user-driven exploration of emotional data.
  • Future adoption: As automatic affect-category detection becomes more accurate, its use in natural-language applications is expected to become more ubiquitous.This is presented as a broad direction for the field’s applications.

10. Ethics and Fairness

Sentiment analysis can create substantial opportunities, but machine-learning systems may reproduce human biases and enable manipulation. The paper therefore frames fairness and responsible use as necessary considerations for applications affecting users and populations.

  • Bias and fairness: Machine-learning systems can perpetuate human biases, producing inappropriate outcomes and negative experiences for some users.The passage notes that some learned biases may benefit a downstream application, while others may be harmful.
  • Misuse and manipulation: Sentiment analysis can expose people’s opinions, preferences, personality traits, and emotional vulnerability to manipulation, deception, and indoctrination.The paper connects this risk to commercial persuasion and political agendas.

11. Summary and Future Directions

The chapter surveys sentiment analysis problems, applications, challenges, methods, and resources, then identifies several directions for future research. These include richer affect modeling, multimodal analysis, broader applications, interdisciplinary work, and attention to ethics, fairness, and bias.

  • The chapter surveys sentiment analysis problems, applications, challenges, algorithms, features, datasets, and lexicon-construction approaches.
  • Sentiment analysis is expanding across psychology, digital humanities, history, art, and social sciences by enabling analysis of massive amounts of text.
  • The chapter anticipates continued interdisciplinary convergence alongside increased research on ethics, fairness, bias, and identifying biases in sentiment analysis systems.
Loading 2005.11882v2…