---
title: "Sentiment & emotion — papers — Elamur Research"
description: "Sentiment & emotion: 609 papers of the Elamur Research library, the most cited first. Among them: An Argument for Basic Emotions."
canonical: "https://research.elamur.ai/free/en/topics/sentiment-emotion/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/topics/sentiment-emotion/)

[home](https://md.research.elamur.ai/free/en/) / [topics](https://md.research.elamur.ai/free/en/topics/) / Sentiment & emotion

# 😊 Sentiment & emotion

609 papers · page 1 of 7 · 1–25 · the most cited first · [all topics](https://md.research.elamur.ai/free/en/topics/)

Opinions, feelings and attitudes in text: what people like, how they feel, what they mean by it.

A survey · about 19 minutes to read · every paper named here has its own page in this library

**In this article**

1. [What it is](#a-what-it-is)
2. [History and milestones](#a-history-and-milestones)
3. [Main problems and tasks](#a-main-problems-and-tasks)
4. [Methods and schools](#a-methods-and-schools)
5. [Datasets, resources and benchmarks](#a-datasets-resources-and-benchmarks)
6. [Landmark papers in this library](#a-landmark-papers-in-this-library)
7. [People](#a-people)
8. [Institutions and places](#a-institutions-and-places)
9. [Open questions](#a-open-questions)
10. [Neighbouring topics](#a-neighbouring-topics)
11. [Key works in this library](#a-key-works-in-this-library)

Sentiment and emotion analysis is the part of natural language processing that deals with the evaluative and affective side of text: whether a writer likes or dislikes something, how strongly, towards which target, and what feeling the words express or evoke. The field is also known as opinion mining, subjectivity analysis or affective computing for text. Its questions range from the simple (is this movie review positive or negative?) to the subtle (is this tweet sarcastic, and if so, what does its author really think?).

The topic grew out of two older traditions. One is the study of subjectivity in computational linguistics, which asked how private states such as beliefs, judgements and feelings are expressed in language. The other is psychology, where theories of emotion supplied the categories that annotators still use. Around 2002 the two met the practical demand of the web: millions of product reviews, blog posts and later tweets that companies, political scientists and platform moderators wanted to read at scale. Since then the area has become one of the busiest in the field, with its own shared tasks, lexicons, benchmarks and a steady stream of new phenomena to model, from aspects and stance to irony, hate speech and empathy in dialogue.

The papers here cover that whole arc: early statistical classifiers of movie reviews, sentiment lexicons built from WordNet, the SemEval series on Twitter and on aspect-based analysis, neural models with attention and graph networks, multimodal datasets drawn from video, and the more recent turn towards abusive language, bias and large language models.

## What it is

At its core, sentiment analysis assigns a **polarity** (positive, negative, sometimes neutral or mixed) or a **score** on a scale to a unit of text. Emotion analysis replaces or extends polarity with a set of categories such as joy, sadness, anger and fear, or with continuous dimensions such as valence (pleasantness) and arousal (intensity). Both rest on the notion of an **opinion**, which most researchers model as a structured object: an opinion holder, a target (an entity or one of its aspects), a sentiment or emotion, and often a time. A review sentence like "the screen is gorgeous but the battery dies by noon" holds two opinions by the same holder about two aspects of one product, with opposite polarities.

The field works at several levels of granularity:

- **Document level** — the overall verdict of a review or post.
- **Sentence level** — first whether a sentence is subjective at all, then its polarity.
- **Phrase or expression level** — the contextual polarity of a word or phrase in its sentence.
- **Aspect or target level** — the sentiment towards a particular entity or attribute mentioned in the text.

Several neighbouring tasks share its methods and its data. **Stance detection** asks which side an author takes on a proposition or target, which need not be mentioned directly. **Sarcasm and irony detection** looks for cases where the literal sentiment of the words contradicts the intended one. **Hate speech and offensive language detection** classifies abusive content and its targets; it borrows features and annotation practice from sentiment work, although its object is harm rather than evaluation. **Emotion recognition in conversation** follows how feelings shift across the turns of a dialogue. Each of these is covered in the papers here, and together they explain why the topic overlaps so much with [Bias, fairness & safety](https://md.research.elamur.ai/free/en/topics/bias-fairness-safety/), [Dialogue systems](https://md.research.elamur.ai/free/en/topics/dialogue-systems/) and [Vision & language](https://md.research.elamur.ai/free/en/topics/multimodal/).

What sets the area apart from ordinary text classification is that its signal is diffuse, contextual and easily reversed. A single negation, a contrastive "but", a domain shift ("unpredictable" is good for a film plot and bad for a car's steering) or an ironic tone can flip the answer. Much of the history of the field is the story of finding ways to model these effects.

## History and milestones

### Foundations: emotion theory and subjectivity (before 2002)

The oldest work in this library is not computational at all. [Paul Ekman's argument for basic emotions](https://md.research.elamur.ai/free/en/papers/ekman-1992-argument-basic-emotions/) defended the view that emotions form a small set of discrete families with distinctive signals and physiology, rather than points in a continuous space. His six categories (happiness, sadness, fear, anger, surprise, disgust) became, with Robert Plutchik's eight-emotion wheel, the default label sets for emotion datasets decades later. In computational linguistics, the 1990s brought work on subjectivity by Janyce Wiebe and colleagues, and on the semantic orientation of adjectives, which established that evaluative meaning could be learned from corpora.

### The statistical turn (2002–2009)

The year 2002 is usually taken as the start of modern sentiment analysis. [Pang, Lee and Vaithyanathan's "Thumbs up?"](https://md.research.elamur.ai/free/en/papers/pang-2002-thumbs-up-sentiment-classification-machine/) treated review polarity as a supervised learning problem and showed that standard classifiers beat word lists chosen by people, yet fell short of topic-classification accuracy. The following years filled in the picture: [subjectivity extraction with minimum cuts](https://md.research.elamur.ai/free/en/papers/pang-2004-sentimental-education-sentiment-analysis-subjectivity/), [rating-scale inference](https://md.research.elamur.ai/free/en/papers/pang-2005-seeing-stars-exploiting-class-relationships/), [opinion holders and their sentiment](https://md.research.elamur.ai/free/en/papers/kim-2004-determining-sentiment-opinions/), [contextual polarity at the phrase level](https://md.research.elamur.ai/free/en/papers/wilson-2005-recognizing-contextual-polarity-phrase-level/), and [domain adaptation across product categories](https://md.research.elamur.ai/free/en/papers/blitzer-2007-biographies-bollywood-boom-boxes-blenders/). Lexical resources appeared in parallel: [WordNet-Affect](https://md.research.elamur.ai/free/en/papers/strapparava-2004-wordnet-affect-affective-extension-wordnet/) and [SentiWordNet](https://md.research.elamur.ai/free/en/papers/esuli-2006-sentiwordnet-publicly-available-lexical-resource/).

Landmark papers on this topic: the 12 most cited, in time The 12 most cited papers on this topic span 1992–2016; the dot grows with the citations.

**The numbers**

| Paper | Citations |
| --- | --- |
| [1992 — An Argument for Basic Emotions](https://md.research.elamur.ai/free/en/papers/ekman-1992-argument-basic-emotions/) | 9,903 |
| [2002 — Thumbs up? Sentiment Classification using Machine Learning Techniques](https://md.research.elamur.ai/free/en/papers/pang-2002-thumbs-up-sentiment-classification-machine/) | 7,762 |
| [2004 — A Sentimental Education: Sentiment Analysis Using Subjectivity Summarization Based on Minimum Cuts](https://md.research.elamur.ai/free/en/papers/pang-2004-sentimental-education-sentiment-analysis-subjectivity/) | 3,849 |
| [2005 — Recognizing Contextual Polarity in Phrase-Level Sentiment Analysis](https://md.research.elamur.ai/free/en/papers/wilson-2005-recognizing-contextual-polarity-phrase-level/) | 3,991 |
| [2005 — Seeing Stars: Exploiting Class Relationships for Sentiment Categorization with Respect to Rating Scales](https://md.research.elamur.ai/free/en/papers/pang-2005-seeing-stars-exploiting-class-relationships/) | 2,733 |
| [2006 — SENTIWORDNET: A Publicly Available Lexical Resource for Opinion Mining](https://md.research.elamur.ai/free/en/papers/esuli-2006-sentiwordnet-publicly-available-lexical-resource/) | 2,722 |
| [2010 — SentiWordNet 3.0: An Enhanced Lexical Resource for Sentiment Analysis and Opinion Mining](https://md.research.elamur.ai/free/en/papers/baccianella-2010-sentiwordnet-3-0-enhanced-lexical/) | 3,076 |
| [2010 — Twitter as a Corpus for Sentiment Analysis and Opinion Mining](https://md.research.elamur.ai/free/en/papers/pak-2010-twitter-as-corpus-sentiment-analysis/) | 2,915 |
| [2011 — Learning Word Vectors for Sentiment Analysis](https://md.research.elamur.ai/free/en/papers/maas-2011-learning-word-vectors-sentiment-analysis/) | 4,290 |
| [2011 — Lexicon-Based Methods for Sentiment Analysis](https://md.research.elamur.ai/free/en/papers/taboada-2011-lexicon-based-methods-sentiment-analysis/) | 3,563 |
| [2014 — SemEval-2014 Task 4: Aspect Based Sentiment Analysis](https://md.research.elamur.ai/free/en/papers/pontiki-2014-semeval-2014-task-4-aspect/) | 2,409 |
| [2016 — Attention-based LSTM for Aspect-level Sentiment Classification](https://md.research.elamur.ai/free/en/papers/wang-2016-attention-based-lstm-aspect-level/) | 2,619 |

### Social media and shared tasks (2010–2014)

Twitter changed the data. Short, noisy, emoticon-rich messages arrived in huge volumes, and researchers found they could label them for free by treating emoticons and hashtags as noisy labels. [Pak and Paroubek](https://md.research.elamur.ai/free/en/papers/pak-2010-twitter-as-corpus-sentiment-analysis/), [Barbosa and Feng](https://md.research.elamur.ai/free/en/papers/barbosa-2010-robust-sentiment-detection-twitter-biased/) and [Davidov, Tsur and Rappoport](https://md.research.elamur.ai/free/en/papers/davidov-2010-enhanced-sentiment-learning-twitter-hashtags/) all built on this distant supervision, while [Agarwal and colleagues](https://md.research.elamur.ai/free/en/papers/agarwal-2011-sentiment-analysis-twitter-data/) worked with hand-labelled tweets. Lexicons were scaled up through crowdsourcing ([EmoLex](https://md.research.elamur.ai/free/en/papers/mohammad-2010-emotions-evoked-common-words-phrases/)) and semi-supervised refinement ([SentiWordNet 3.0](https://md.research.elamur.ai/free/en/papers/baccianella-2010-sentiwordnet-3-0-enhanced-lexical/)). The SemEval workshops turned Twitter sentiment into a yearly competition from 2013, starting with [Task 2](https://md.research.elamur.ai/free/en/papers/nakov-2013-semeval-2013-task-2-sentiment/), and launched aspect-based sentiment analysis as a benchmark in [2014](https://md.research.elamur.ai/free/en/papers/pontiki-2014-semeval-2014-task-4-aspect/). Winning systems of this period, such as [NRC-Canada](https://md.research.elamur.ai/free/en/papers/mohammad-2013-nrc-canada-building-state-art/), were linear classifiers fed with very large automatically induced lexicons.

### The neural era (2011–2019)

Neural methods arrived early in this area. [Maas and colleagues](https://md.research.elamur.ai/free/en/papers/maas-2011-learning-word-vectors-sentiment-analysis/) learned word vectors that encoded polarity and released the IMDB dataset; [Socher's recursive autoencoders](https://md.research.elamur.ai/free/en/papers/socher-2011-semi-supervised-recursive-autoencoders-predicting/) composed phrase meanings along trees, a line that led in 2013 to the Stanford Sentiment Treebank. By 2014–2016 the standard toolkit was word embeddings plus convolutional or recurrent networks: [character-level CNNs for short texts](https://md.research.elamur.ai/free/en/papers/santos-2014-deep-convolutional-neural-networks-sentiment/), [sentiment-specific embeddings](https://md.research.elamur.ai/free/en/papers/tang-2014-learning-sentiment-specific-word-embedding/), [hierarchical document models](https://md.research.elamur.ai/free/en/papers/tang-2015-document-modeling-gated-recurrent-neural/), and a series of target-aware models for aspects: [TD-LSTM](https://md.research.elamur.ai/free/en/papers/tang-2016-effective-lstms-target-dependent-sentiment/), [attention-based LSTMs](https://md.research.elamur.ai/free/en/papers/wang-2016-attention-based-lstm-aspect-level/), [memory networks](https://md.research.elamur.ai/free/en/papers/tang-2016-aspect-level-sentiment-classification-deep/) and [recurrent attention on memory](https://md.research.elamur.ai/free/en/papers/chen-2017-recurrent-attention-network-memory-aspect/). The same years saw the rise of multimodal sentiment from video ([Tensor Fusion Network](https://md.research.elamur.ai/free/en/papers/zadeh-2017-tensor-fusion-network-multimodal-sentiment/), [CMU-MOSEI](https://md.research.elamur.ai/free/en/papers/zadeh-2018-multimodal-language-analysis-wild-cmu/)) and of emotion recognition in conversation ([MELD](https://md.research.elamur.ai/free/en/papers/poria-2018-meld-multimodal-multi-party-dataset/), [DialogueGCN](https://md.research.elamur.ai/free/en/papers/ghosal-2019-dialoguegcn-graph-convolutional-neural-network/)).

Papers on this topic by year 609 dated papers from 1992 to 2026; the busiest year is 2020 with 82, and 75% of them appeared in 2018 or later.

**The numbers**

| Year | Papers |
| --- | --- |
| 2026 | 36 |
| 2025 | 16 |
| 2024 | 31 |
| 2023 | 31 |
| 2022 | 65 |
| 2021 | 79 |
| 2020 | 82 |
| 2019 | 48 |
| 2018 | 66 |
| 2017 | 30 |
| 2016 | 35 |
| 2015 | 17 |
| 2014 | 20 |
| 2013 | 9 |
| 2012 | 7 |
| 2011 | 8 |
| 2010 | 9 |
| 2009 | 5 |
| 2008 | 2 |
| 2007 | 3 |
| 2006 | 1 |
| 2005 | 3 |
| 2004 | 3 |
| 2002 | 1 |
| 1996 | 1 |
| 1992 | 1 |

### Pretrained transformers, abuse and bias (2018 onward)

From 2018 pretrained language models such as BERT became the default encoder, and leaderboard gains on classic polarity benchmarks flattened. Attention moved to harder or socially pressing tasks: hate speech and offensive language ([OffensEval](https://md.research.elamur.ai/free/en/papers/zampieri-2019-semeval-2019-task-6-identifying/), [HatEval](https://md.research.elamur.ai/free/en/papers/basile-2019-semeval-2019-task-5-multilingual/)), stance, sarcasm in context, structured aspect extraction ([triplets](https://md.research.elamur.ai/free/en/papers/xu-2020-position-aware-tagging-aspect-sentiment/), [quadruples](https://md.research.elamur.ai/free/en/papers/zhang-2021-aspect-sentiment-quad-prediction-as/)), and the fairness of sentiment systems themselves ([racial bias in hate speech detection](https://md.research.elamur.ai/free/en/papers/sap-2019-risk-racial-bias-hate-speech/), [gender and race bias across two hundred systems](https://md.research.elamur.ai/free/en/papers/kiritchenko-2018-examining-gender-race-bias-two/)). Most papers on this topic in the library date from this period. Large language models posed a new question, taken up in [a reality check of sentiment analysis in the LLM era](https://md.research.elamur.ai/free/en/papers/zhang-2023-sentiment-analysis-era-large-language/): how far do prompted general models get on the classic tasks without task-specific training?

## Main problems and tasks

### Polarity classification

The founding task, deciding whether a text is positive or negative, remains the entry point. Its difficulties were catalogued early: reviews spend much of their length on plot summary; writers build up expectations and then overturn them; a word's sentiment depends on the domain. [Pang and Lee's 2004 work](https://md.research.elamur.ai/free/en/papers/pang-2004-sentimental-education-sentiment-analysis-subjectivity/) responded by filtering out objective sentences first, and [Barbosa and Feng](https://md.research.elamur.ai/free/en/papers/barbosa-2010-robust-sentiment-detection-twitter-biased/) used the same two-step idea on tweets.

### Rating and intensity prediction

Many applications need more than a binary label. [Seeing Stars](https://md.research.elamur.ai/free/en/papers/pang-2005-seeing-stars-exploiting-class-relationships/) framed rating prediction as ordinal classification in which neighbouring labels are closer than distant ones. For emotions, the [WASSA-2017 emotion-intensity task](https://md.research.elamur.ai/free/en/papers/mohammad-2017-wassa-2017-shared-task-emotion/), [Emotion Intensities in Tweets](https://md.research.elamur.ai/free/en/papers/mohammad-2017-emotion-intensities-tweets/) and [SemEval-2018 Affect in Tweets](https://md.research.elamur.ai/free/en/papers/mohammad-2018-semeval-2018-task-1-affect/) asked systems to estimate how strongly a feeling is expressed, using intensity scores obtained by comparative annotation.

### Phrase-level and compositional sentiment

Negators, intensifiers, modals and contrast all change polarity. [Wilson, Wiebe and Hoffmann](https://md.research.elamur.ai/free/en/papers/wilson-2005-recognizing-contextual-polarity-phrase-level/) separated a word's prior polarity from its polarity in context and later [explored the features behind it in depth](https://md.research.elamur.ai/free/en/papers/wilson-2009-recognizing-contextual-polarity-exploration-features/). [Implicature-based sentiment propagation](https://md.research.elamur.ai/free/en/papers/deng-2014-sentiment-propagation-implicature-constraints/) went further, inferring attitudes towards entities that are not explicitly evaluated, a link to [Implicature](https://md.research.elamur.ai/free/en/topics/implicature/).

### Aspect-based sentiment analysis

Aspect-based analysis (ABSA) asks which features of a product or service are discussed and how each is judged. The task family grew from simple polarity of a given target to full extraction of (aspect, opinion, sentiment) triplets and (aspect, category, opinion, sentiment) quadruples. Unsupervised and weakly supervised variants learn aspects from review ratings, as in the [joint model of text and aspect ratings](https://md.research.elamur.ai/free/en/papers/titov-2008-joint-model-text-aspect-ratings/) and the [unsupervised aspect-sentiment model](https://md.research.elamur.ai/free/en/papers/brody-2010-unsupervised-aspect-sentiment-model-online/).

### Emotion detection

Emotion classification uses category sets from psychology (Ekman, Plutchik) or richer custom inventories. Early work such as [Alm and colleagues' emotion prediction from text](https://md.research.elamur.ai/free/en/papers/alm-2005-emotions-text-machine-learning-text/) and [SemEval-2007 Affective Text](https://md.research.elamur.ai/free/en/papers/strapparava-2007-semeval-2007-task-14-affective/) was followed by tweet-scale models like [EmoNet](https://md.research.elamur.ai/free/en/papers/mageed-2017-emonet-fine-grained-emotion-detection/) and [CARER](https://md.research.elamur.ai/free/en/papers/saravia-2018-carer-contextualized-affect-representations-emotion/), and by context-aware tasks such as [EmoContext](https://md.research.elamur.ai/free/en/papers/chatterjee-2019-semeval-2019-task-3-emocontext/). In dialogue, emotion feeds generation: [MoEL](https://md.research.elamur.ai/free/en/papers/lin-2019-moel-mixture-empathetic-listeners/), [MIME](https://md.research.elamur.ai/free/en/papers/majumder-2020-mime-mimicking-emotions-empathetic-response/) and [emotional support dialogue](https://md.research.elamur.ai/free/en/papers/liu-2021-emotional-support-dialog-systems/) aim at responses that acknowledge how the user feels.

### Sarcasm and irony

Sarcasm defeats any system that reads words at face value. The work here ranges from [authors' own #sarcasm labels](https://md.research.elamur.ai/free/en/papers/ibanez-2011-identifying-sarcasm-twitter-closer-look/) and the [positive-sentiment-meets-negative-situation pattern](https://md.research.elamur.ai/free/en/papers/riloff-2013-sarcasm-as-contrast-between-positive/) to [context incongruity](https://md.research.elamur.ai/free/en/papers/joshi-2015-harnessing-context-incongruity-sarcasm-detection/), [conversation context](https://md.research.elamur.ai/free/en/papers/ghosh-2018-sarcasm-analysis-conversation-context/), [user embeddings](https://md.research.elamur.ai/free/en/papers/amir-2016-modelling-context-user-embeddings-sarcasm/) and [multimodal sarcasm in video](https://md.research.elamur.ai/free/en/papers/castro-2019-towards-multimodal-sarcasm-detection-obviously/). Shared tasks include [SemEval-2018 irony detection](https://md.research.elamur.ai/free/en/papers/hee-2018-semeval-2018-task-3-irony/) and [iSarcasmEval](https://md.research.elamur.ai/free/en/papers/farha-2022-semeval-2022-task-6-isarcasmeval/), which asked authors themselves to mark intended sarcasm.

### Stance detection

Stance is an attitude towards a target that may be absent from the text. Work here goes from [ideological online debates](https://md.research.elamur.ai/free/en/papers/somasundaran-2010-recognizing-stances-ideological-line-debates/) and [bidirectional conditional encoding](https://md.research.elamur.ai/free/en/papers/augenstein-2016-stance-detection-bidirectional-conditional-encoding/) to [zero-shot stance](https://md.research.elamur.ai/free/en/papers/allaway-2020-zero-shot-stance-detection-dataset/), [P-Stance](https://md.research.elamur.ai/free/en/papers/li-2021-p-stance-large-dataset-stance/) and [COVID-19 stance](https://md.research.elamur.ai/free/en/papers/glandt-2021-stance-detection-covid-19-tweets/), and to stance as a step in fact checking ([survey](https://md.research.elamur.ai/free/en/papers/hardalov-2022-survey-stance-detection-mis-disinformation/)).

### Abusive language and hate speech

A large body of work treats hate speech and offensive language as a sentiment-adjacent classification task. Its central issues are definitions and annotation: what counts as hateful, who decides, and whether merely offensive talk is being confused with hate. These questions run through [Warner and Hirschberg](https://md.research.elamur.ai/free/en/papers/warner-2012-detecting-hate-speech-world-wide/), [Waseem and Hovy](https://md.research.elamur.ai/free/en/papers/waseem-2016-hateful-symbols-or-hateful-people/), [Davidson and colleagues](https://md.research.elamur.ai/free/en/papers/davidson-2017-automated-hate-speech-detection-problem/) and [the analysis of what hate speech datasets actually classify](https://md.research.elamur.ai/free/en/papers/fortuna-2020-toxic-hateful-offensive-or-abusive/).

## Methods and schools

### Lexicon-based methods

The lexicon school scores words with a prior polarity or emotion and combines the scores with rules for negation, intensification and modality. Resources were built by hand, propagated through WordNet relations from small seed sets ([SentiWordNet](https://md.research.elamur.ai/free/en/papers/esuli-2006-sentiwordnet-publicly-available-lexical-resource/), [Kim and Hovy](https://md.research.elamur.ai/free/en/papers/kim-2004-determining-sentiment-opinions/)), crowdsourced ([EmoLex](https://md.research.elamur.ai/free/en/papers/mohammad-2010-emotions-evoked-common-words-phrases/), [Word Affect Intensities](https://md.research.elamur.ai/free/en/papers/mohammad-2018-word-affect-intensities/)) or induced from corpora ([domain-specific lexicons from unlabeled text](https://md.research.elamur.ai/free/en/papers/hamilton-2016-inducing-domain-specific-sentiment-lexicons/)). The strongest case for this school is [SO-CAL](https://md.research.elamur.ai/free/en/papers/taboada-2011-lexicon-based-methods-sentiment-analysis/), which argued that a carefully curated dictionary transfers across domains better than a classifier trained on one of them. Lexicons never disappeared: they were among the most useful features in SemEval winners and are still used for analysis and as priors.

### Feature-based machine learning

The supervised school treated sentiment as text classification with Naive Bayes, maximum entropy and SVMs over n-grams, part-of-speech tags, negation-scope markers and lexicon scores. [Wang and Manning's NBSVM](https://md.research.elamur.ai/free/en/papers/wang-2012-baselines-bigrams-simple-good-sentiment/) is the best-known demonstration that a well-tuned linear baseline is hard to beat. Distant supervision from emoticons, hashtags and star ratings was the key to scaling this approach to social media, and domain adaptation ([structural correspondence learning](https://md.research.elamur.ai/free/en/papers/blitzer-2007-biographies-bollywood-boom-boxes-blenders/)) and cross-lingual transfer ([co-training through machine translation](https://md.research.elamur.ai/free/en/papers/wan-2009-co-training-cross-lingual-sentiment/)) addressed the cost of labels in new domains and languages.

### Neural composition and representation learning

Neural models promised to learn negation and composition rather than encode them by hand. Recursive networks composed vectors along parse trees; convolutional and recurrent networks did so along sequences; attention let the model choose which words matter for a target. A recurring debate set syntax-driven models against syntax-free ones. [AdaRNN](https://md.research.elamur.ai/free/en/papers/dong-2014-adaptive-recursive-neural-network-target/), [ASGCN](https://md.research.elamur.ai/free/en/papers/zhang-2019-aspect-based-sentiment-classification-aspect/), [convolution over dependency trees](https://md.research.elamur.ai/free/en/papers/sun-2019-aspect-level-sentiment-analysis-convolution/) and [R-GAT](https://md.research.elamur.ai/free/en/papers/wang-2020-relational-graph-attention-network-aspect/) argued that dependency structure helps link an aspect to the right opinion word, while attention and memory models argued that parsers were too unreliable on noisy text. Pretraining on noisy affective signals also proved powerful: [DeepMoji](https://md.research.elamur.ai/free/en/papers/felbo-2017-millions-emoji-occurrences-learn-any/) learned from emoji in about a billion tweets, and [a character-level language model trained on reviews](https://md.research.elamur.ai/free/en/papers/radford-2017-learning-generate-reviews-discovering-sentiment/) was found to contain a single unit that tracked sentiment.

### Pretrained transformers and generation

With BERT, aspect-based analysis was recast as [sentence-pair classification with auxiliary sentences](https://md.research.elamur.ai/free/en/papers/sun-2019-utilizing-bert-aspect-based-sentiment/) or improved through [post-training on review text](https://md.research.elamur.ai/free/en/papers/xu-2019-bert-post-training-review-reading/). Generative formulations followed, in which all subtasks become sequence generation ([a unified generative framework](https://md.research.elamur.ai/free/en/papers/yan-2021-unified-generative-framework-aspect-based/), [quad prediction as paraphrase generation](https://md.research.elamur.ai/free/en/papers/zhang-2021-aspect-sentiment-quad-prediction-as/)). Multilingual models adapted to social media, such as [XLM-T](https://md.research.elamur.ai/free/en/papers/barbieri-2022-xlm-t-multilingual-language-models/), extended the reach to many languages. Sentiment also became something to control in generation: [Delete, Retrieve, Generate](https://md.research.elamur.ai/free/en/papers/li-2018-delete-retrieve-generate-simple-approach/) and [cycled reinforcement learning](https://md.research.elamur.ai/free/en/papers/xu-2018-unpaired-sentiment-sentiment-translation-cycled/) rewrite a sentence's polarity while keeping its content.

### Multimodal fusion

In video, sentiment comes from words, voice and face together. The main question is how to fuse them: early concatenation, late voting, or explicit modelling of cross-modal interactions. [Tensor Fusion](https://md.research.elamur.ai/free/en/papers/zadeh-2017-tensor-fusion-network-multimodal-sentiment/) took the outer product of modality embeddings; the [Dynamic Fusion Graph](https://md.research.elamur.ai/free/en/papers/zadeh-2018-multimodal-language-analysis-wild-cmu/) learned which combinations to use; [contextual LSTMs](https://md.research.elamur.ai/free/en/papers/poria-2017-context-dependent-sentiment-analysis-user/) added neighbouring utterances. Multimodal sarcasm and hateful memes ([Hateful Memes Challenge](https://md.research.elamur.ai/free/en/papers/kiela-2020-hateful-memes-challenge-detecting-hate/)) carry the same problem into images.

### Annotation as method

A distinctive feature of this area is how much of its methodology concerns annotation. Questions about who labels, with what guidelines and with what scale recur across the papers: [Best–Worst Scaling](https://md.research.elamur.ai/free/en/papers/kiritchenko-2017-best-worst-scaling-more-reliable/) instead of rating scales; expert against crowd labels for [hate speech](https://md.research.elamur.ai/free/en/papers/waseem-2016-are-you-racist-or-am/); priming annotators with dialect information to reduce [racial bias](https://md.research.elamur.ai/free/en/papers/sap-2019-risk-racial-bias-hate-speech/); and the perspectivist position that disagreement is data, not noise ([Agreeing to Disagree](https://md.research.elamur.ai/free/en/papers/leonardelli-2021-agreeing-disagree-annotating-offensive-language/), [EPIC](https://md.research.elamur.ai/free/en/papers/frenda-2023-epic-multi-perspective-annotation-corpus/), [MultiPICo](https://md.research.elamur.ai/free/en/papers/casola-2024-multipico-multilingual-perspectivist-irony-corpus/)).

## Datasets, resources and benchmarks

| Resource | Kind | Notes |
| --- | --- | --- |
| Movie review polarity data | Documents | Pang and Lee's review corpora, the first common benchmark |
| [IMDB Large Movie Review Dataset](https://md.research.elamur.ai/free/en/papers/maas-2011-learning-word-vectors-sentiment-analysis/) | Documents | 50,000 polar reviews, train and test split by film |
| [Amazon multi-domain reviews](https://md.research.elamur.ai/free/en/papers/blitzer-2007-biographies-bollywood-boom-boxes-blenders/) | Documents | Books, DVDs, electronics, kitchen; the standard domain-adaptation testbed |
| MPQA Opinion Corpus | Expressions | News annotated for private states; extended with [contextual polarity](https://md.research.elamur.ai/free/en/papers/wilson-2005-recognizing-contextual-polarity-phrase-level/) |
| Stanford Sentiment Treebank | Phrases | Sentiment labels for every phrase of parsed movie-review sentences |
| [SemEval Twitter tasks](https://md.research.elamur.ai/free/en/papers/rosenthal-2017-semeval-2017-task-4-sentiment/) | Tweets | Yearly editions 2013–2017, later with Arabic |
| [SemEval ABSA](https://md.research.elamur.ai/free/en/papers/pontiki-2016-semeval-2016-task-5-aspect/) | Review sentences | Restaurants and laptops, extended to eight languages in 2016 |
| [SentiWordNet](https://md.research.elamur.ai/free/en/papers/baccianella-2010-sentiwordnet-3-0-enhanced-lexical/), [WordNet-Affect](https://md.research.elamur.ai/free/en/papers/strapparava-2004-wordnet-affect-affective-extension-wordnet/), [EmoLex](https://md.research.elamur.ai/free/en/papers/mohammad-2010-emotions-evoked-common-words-phrases/) | Lexicons | Sense-level polarity, affective labels, crowd-sourced emotions |
| [CMU-MOSEI](https://md.research.elamur.ai/free/en/papers/zadeh-2018-multimodal-language-analysis-wild-cmu/), CMU-MOSI | Video | Sentiment and emotion from text, audio and vision |
| [MELD](https://md.research.elamur.ai/free/en/papers/poria-2018-meld-multimodal-multi-party-dataset/), IEMOCAP | Dialogue | Emotion in multi-party and dyadic conversation |
| [OLID / OffensEval](https://md.research.elamur.ai/free/en/papers/zampieri-2019-semeval-2019-task-6-identifying/), [HatEval](https://md.research.elamur.ai/free/en/papers/basile-2019-semeval-2019-task-5-multilingual/) | Tweets | Offensive language and hate speech with targets |
| [ToxiGen](https://md.research.elamur.ai/free/en/papers/hartvigsen-2022-toxigen-large-scale-machine-generated/), [HateCheck](https://md.research.elamur.ai/free/en/papers/rottger-2021-hatecheck-functional-tests-hate-speech/) | Test suites | Machine-generated implicit hate; functional tests for classifiers |
| [SARC](https://md.research.elamur.ai/free/en/papers/khodak-2018-large-self-annotated-corpus-sarcasm/) | Reddit comments | A large corpus of sarcasm labelled by the writers themselves |

Beyond English, the library holds resources such as the [Arabic Sentiment Tweets Dataset](https://md.research.elamur.ai/free/en/papers/nabil-2015-astd-arabic-sentiment-tweets-dataset/), [ArSarcasm](https://md.research.elamur.ai/free/en/papers/farha-2020-arabic-sentiment-analysis-sarcasm-detection/), the [Italian hate speech corpus](https://md.research.elamur.ai/free/en/papers/sanguinetti-2018-italian-twitter-corpus-hate-speech/), [code-mixed Tamil–English](https://md.research.elamur.ai/free/en/papers/chakravarthi-2020-corpus-creation-sentiment-analysis-code/) and [Malayalam–English](https://md.research.elamur.ai/free/en/papers/chakravarthi-2020-sentiment-analysis-dataset-code-mixed/) data, [Turkish](https://md.research.elamur.ai/free/en/papers/coltekin-2020-corpus-turkish-offensive-language-social/), [Greek](https://md.research.elamur.ai/free/en/papers/pitenis-2020-offensive-language-identification-greek/) and [Danish](https://md.research.elamur.ai/free/en/papers/sigurbergsson-2020-offensive-language-hate-speech-detection/) offensive-language corpora, [NaijaSenti](https://md.research.elamur.ai/free/en/papers/muhammad-2022-naijasenti-nigerian-twitter-sentiment-corpus/) and [AfriSenti](https://md.research.elamur.ai/free/en/papers/muhammad-2023-afrisenti-twitter-sentiment-analysis-benchmark/) for African languages.

## Landmark papers in this library

### An Argument for Basic Emotions (1992)

[Ekman's essay](https://md.research.elamur.ai/free/en/papers/ekman-1992-argument-basic-emotions/) lists the properties that, in his view, mark an emotion as basic: universal signals such as facial expressions, specific physiology, automatic appraisal, fast onset and short duration. It is a psychology paper, but its categories became the label set of countless emotion corpora, which is why it is the most cited work on this topic.

### Thumbs up? (2002)

[Pang, Lee and Vaithyanathan](https://md.research.elamur.ai/free/en/papers/pang-2002-thumbs-up-sentiment-classification-machine/) built a movie-review corpus whose labels came from star ratings and compared Naive Bayes, maximum entropy and SVMs. Machine learning clearly beat keyword lists, binary word presence beat frequency counts, and the remaining errors came from reviews whose wording pointed one way while the verdict pointed the other. The paper set the template for supervised sentiment classification.

### A Sentimental Education (2004) and Seeing Stars (2005)

The same group then [removed objective sentences with a graph minimum cut](https://md.research.elamur.ai/free/en/papers/pang-2004-sentimental-education-sentiment-analysis-subjectivity/) before classifying, keeping accuracy while shrinking reviews to their opinionated core, and [predicted ratings on a multi-point scale](https://md.research.elamur.ai/free/en/papers/pang-2005-seeing-stars-exploiting-class-relationships/) by exploiting the fact that nearby labels are similar. Both released datasets that became standard.

### Recognizing Contextual Polarity (2005)

[Wilson, Wiebe and Hoffmann](https://md.research.elamur.ai/free/en/papers/wilson-2005-recognizing-contextual-polarity-phrase-level/) showed that most occurrences of words from a sentiment lexicon are neutral in context and built a two-stage classifier that first filters these out and then decides direction. The accompanying subjectivity lexicon (the MPQA lexicon) is still one of the most used resources in the area.

### SentiWordNet (2006, 2010)

[Esuli and Sebastiani](https://md.research.elamur.ai/free/en/papers/esuli-2006-sentiwordnet-publicly-available-lexical-resource/) gave every WordNet synset three graded scores for positivity, negativity and objectivity, derived from a committee of gloss classifiers trained on expanded seed sets. [Version 3.0](https://md.research.elamur.ai/free/en/papers/baccianella-2010-sentiwordnet-3-0-enhanced-lexical/) added a random-walk refinement and an evaluation against hand-labelled synsets. Sense-level, wide-coverage scores made it a default source of priors.

### Biographies, Bollywood, Boom-boxes and Blenders (2007)

[Blitzer, Dredze and Pereira](https://md.research.elamur.ai/free/en/papers/blitzer-2007-biographies-bollywood-boom-boxes-blenders/) introduced the four-domain Amazon review corpus and adapted structural correspondence learning by choosing pivot features that predict the label. The dataset became the canonical benchmark for domain adaptation in NLP generally, not only in sentiment.

### Learning Word Vectors for Sentiment Analysis (2011)

[Maas and colleagues](https://md.research.elamur.ai/free/en/papers/maas-2011-learning-word-vectors-sentiment-analysis/) combined an unsupervised document model with a supervised sentiment objective so that word vectors would capture polarity as well as topic. The IMDB dataset they released remains one of the most used text classification benchmarks.

### Lexicon-Based Methods for Sentiment Analysis (2011)

[SO-CAL](https://md.research.elamur.ai/free/en/papers/taboada-2011-lexicon-based-methods-sentiment-analysis/) is the fullest statement of the lexicon approach: hand-ranked dictionaries for several parts of speech, with negation and intensification treated as general operations. Its argument for robustness across domains and transparency is still cited whenever trained classifiers fail outside their training data.

### SemEval aspect-based sentiment tasks (2014–2016)

The [2014](https://md.research.elamur.ai/free/en/papers/pontiki-2014-semeval-2014-task-4-aspect/), [2015](https://md.research.elamur.ai/free/en/papers/pontiki-2015-semeval-2015-task-12-aspect/) and [2016](https://md.research.elamur.ai/free/en/papers/pontiki-2016-semeval-2016-task-5-aspect/) tasks defined ABSA as a benchmark: aspect terms, aspect categories as entity–attribute pairs, opinion targets and their polarity. The restaurant and laptop data from 2014 is the dataset on which almost every neural aspect model of the following decade reports results.

### Attention-based LSTM for Aspect-level Sentiment (2016)

[Wang, Huang and colleagues](https://md.research.elamur.ai/free/en/papers/wang-2016-attention-based-lstm-aspect-level/) gave each aspect its own embedding and used it to steer attention over the sentence, so the same sentence could be read differently for different aspects. It is one of the most cited neural ABSA papers and set the pattern for a long line of attention, memory and graph models.

### Hateful Symbols or Hateful People? (2016)

[Waseem and Hovy](https://md.research.elamur.ai/free/en/papers/waseem-2016-hateful-symbols-or-hateful-people/) released an annotated Twitter corpus of racist and sexist tweets with explicit guidelines and showed that character n-grams carried most of the signal. Together with [Davidson and colleagues' work separating hate from offensive language](https://md.research.elamur.ai/free/en/papers/davidson-2017-automated-hate-speech-detection-problem/) and the [survey by Schmidt and Wiegand](https://md.research.elamur.ai/free/en/papers/schmidt-2017-survey-hate-speech-detection-natural/), it shaped the abusive-language subfield.

### Tensor Fusion Network (2017) and CMU-MOSEI (2018)

[Zadeh and colleagues](https://md.research.elamur.ai/free/en/papers/zadeh-2017-tensor-fusion-network-multimodal-sentiment/) modelled unimodal, bimodal and trimodal interactions explicitly through an outer product of embeddings. The follow-up [CMU-MOSEI](https://md.research.elamur.ai/free/en/papers/zadeh-2018-multimodal-language-analysis-wild-cmu/) supplied a large set of YouTube monologues labelled for sentiment and emotion. Together they defined multimodal sentiment analysis as a research area.

### MELD (2018)

[MELD](https://md.research.elamur.ai/free/en/papers/poria-2018-meld-multimodal-multi-party-dataset/) aligned the dialogue of the sitcom Friends with video and audio and relabelled every utterance for emotion and sentiment while watching the clips. It became the standard multi-party benchmark for emotion recognition in conversation.

### SemEval-2018 Affect in Tweets (2018)

[Mohammad and colleagues](https://md.research.elamur.ai/free/en/papers/mohammad-2018-semeval-2018-task-1-affect/) moved emotion analysis from presence to intensity, in English, Arabic and Spanish, using Best–Worst Scaling. A bias audit run alongside the task found that most submitted systems scored sentences differently depending on the gender or race implied by names, an early warning that sentiment systems inherit social biases.

### Racial bias in hate speech detection (2019)

[Sap and colleagues](https://md.research.elamur.ai/free/en/papers/sap-2019-risk-racial-bias-hate-speech/) showed that tweets in African American English were much more likely to be labelled offensive in popular datasets, and that classifiers trained on them reproduced the skew. Telling annotators about dialect reduced it. The paper made annotation bias a central concern for the field.

Citations of the papers on this topic, by when they were published Papers published in 2015–2019 gather 40% of the 159,171 citations of the papers on this topic.

**The numbers**

| Published | Papers and their citations |
| --- | --- |
| 2025–2029 | 52 papers, 55 citations |
| 2020–2024 | 288 papers, 20,241 citations |
| 2015–2019 | 196 papers, 62,990 citations |
| 2010–2014 | 53 papers, 35,384 citations |
| 2005–2009 | 14 papers, 15,661 citations |
| 2000–2004 | 4 papers, 14,867 citations |
| 1995–1999 | 1 papers, 70 citations |
| 1990–1994 | 1 papers, 9,903 citations |

## People

- [Pushpak Bhattacharyya](https://md.research.elamur.ai/free/en/authors/pushpak-bhattacharyya/) (Indian Institute of Technology Bombay) leads the topic by number of papers. His group's work centres on sarcasm: [context incongruity](https://md.research.elamur.ai/free/en/papers/joshi-2015-harnessing-context-incongruity-sarcasm-detection/), [a survey of automatic sarcasm detection](https://md.research.elamur.ai/free/en/papers/joshi-2016-automatic-sarcasm-detection-survey/), and [multi-task learning of sarcasm with sentiment and emotion](https://md.research.elamur.ai/free/en/papers/chauhan-2020-sentiment-emotion-help-sarcasm-multi/).
- [Aditya Joshi](https://md.research.elamur.ai/free/en/authors/aditya-joshi/), first author of much of that sarcasm work, also studied [whether embedding features help](https://md.research.elamur.ai/free/en/papers/joshi-2016-are-word-embedding-based-features/) and [using an author's past tweets](https://md.research.elamur.ai/free/en/papers/khattri-2015-your-sentiment-precedes-you-author/); his co-author [Mark James Carman](https://md.research.elamur.ai/free/en/authors/mark-carman/) is now at Politecnico di Milano.
- [Saif Mohammad](https://md.research.elamur.ai/free/en/authors/saif-mohammad/) (National Research Council Canada) built some of the field's most used affect resources, from [EmoLex](https://md.research.elamur.ai/free/en/papers/mohammad-2010-emotions-evoked-common-words-phrases/) to [Word Affect Intensities](https://md.research.elamur.ai/free/en/papers/mohammad-2018-word-affect-intensities/), and led the [Affect in Tweets](https://md.research.elamur.ai/free/en/papers/mohammad-2018-semeval-2018-task-1-affect/) task. With [Svetlana Kiritchenko](https://md.research.elamur.ai/free/en/authors/svetlana-kiritchenko/), also at NRC, he built the [NRC-Canada systems](https://md.research.elamur.ai/free/en/papers/kiritchenko-2014-nrc-canada-2014-detecting-aspects/) and audited [bias in sentiment systems](https://md.research.elamur.ai/free/en/papers/kiritchenko-2018-examining-gender-race-bias-two/).
- [Soujanya Poria](https://md.research.elamur.ai/free/en/authors/soujanya-poria/) and [Erik Cambria](https://md.research.elamur.ai/free/en/authors/erik-cambria/), both linked to Nanyang Technological University among other institutions, are central to multimodal sentiment and emotion in conversation: [Tensor Fusion](https://md.research.elamur.ai/free/en/papers/zadeh-2017-tensor-fusion-network-multimodal-sentiment/), [contextual video sentiment](https://md.research.elamur.ai/free/en/papers/poria-2017-context-dependent-sentiment-analysis-user/), [MELD](https://md.research.elamur.ai/free/en/papers/poria-2018-meld-multimodal-multi-party-dataset/) and [DialogueGCN](https://md.research.elamur.ai/free/en/papers/ghosal-2019-dialoguegcn-graph-convolutional-neural-network/). [Devamanyu Hazarika](https://md.research.elamur.ai/free/en/authors/devamanyu-hazarika/) (now at Meta) contributed the [conversational memory networks](https://md.research.elamur.ai/free/en/papers/hazarika-2018-conversational-memory-network-emotion-recognition/) and [ICON](https://md.research.elamur.ai/free/en/papers/hazarika-2018-icon-interactive-conversational-memory-network/).
- [Preslav Nakov](https://md.research.elamur.ai/free/en/authors/preslav-nakov/) (Mohamed bin Zayed University of Artificial Intelligence) and [Sara Rosenthal](https://md.research.elamur.ai/free/en/authors/sara-rosenthal/) (IBM) organised the SemEval Twitter sentiment series from [2014](https://md.research.elamur.ai/free/en/papers/rosenthal-2014-semeval-2014-task-9-sentiment/) through [2016](https://md.research.elamur.ai/free/en/papers/nakov-2016-semeval-2016-task-4-sentiment/) and [2017](https://md.research.elamur.ai/free/en/papers/rosenthal-2017-semeval-2017-task-4-sentiment/), and then [OffensEval](https://md.research.elamur.ai/free/en/papers/zampieri-2019-semeval-2019-task-6-identifying/).
- [Janyce Wiebe](https://md.research.elamur.ai/free/en/authors/janyce-wiebe/), a pioneer of subjectivity analysis, co-authored the [contextual polarity](https://md.research.elamur.ai/free/en/papers/wilson-2005-recognizing-contextual-polarity-phrase-level/) papers with [Theresa Wilson](https://md.research.elamur.ai/free/en/authors/theresa-wilson/) and early stance work on [ideological debates](https://md.research.elamur.ai/free/en/papers/somasundaran-2010-recognizing-stances-ideological-line-debates/).
- [Bo Pang](https://md.research.elamur.ai/free/en/authors/bo-pang/) and [Lillian Lee](https://md.research.elamur.ai/free/en/authors/lillian-lee/) wrote the three papers that founded statistical sentiment classification: [Thumbs up?](https://md.research.elamur.ai/free/en/papers/pang-2002-thumbs-up-sentiment-classification-machine/), [subjectivity extracts](https://md.research.elamur.ai/free/en/papers/pang-2004-sentimental-education-sentiment-analysis-subjectivity/) and [rating scales](https://md.research.elamur.ai/free/en/papers/pang-2005-seeing-stars-exploiting-class-relationships/).
- [Duyu Tang](https://md.research.elamur.ai/free/en/authors/duyu-tang/) (now at Huawei) and [Bing Qin](https://md.research.elamur.ai/free/en/authors/bing-qin/) (Harbin Institute of Technology) produced a run of influential neural models: [sentiment-specific embeddings](https://md.research.elamur.ai/free/en/papers/tang-2014-learning-sentiment-specific-word-embedding/), [gated document models](https://md.research.elamur.ai/free/en/papers/tang-2015-document-modeling-gated-recurrent-neural/), [TD-LSTM](https://md.research.elamur.ai/free/en/papers/tang-2016-effective-lstms-target-dependent-sentiment/) and [deep memory networks](https://md.research.elamur.ai/free/en/papers/tang-2016-aspect-level-sentiment-classification-deep/).
- [Cristina Bosco](https://md.research.elamur.ai/free/en/authors/cristina-bosco/), [Viviana Patti](https://md.research.elamur.ai/free/en/authors/viviana-patti/) and [Valerio Basile](https://md.research.elamur.ai/free/en/authors/valerio-basile/) (University of Turin) built Italian and multilingual resources for hate speech and irony, including [HatEval](https://md.research.elamur.ai/free/en/papers/basile-2019-semeval-2019-task-5-multilingual/), [EVALITA hate speech](https://md.research.elamur.ai/free/en/papers/bosco-2018-overview-evalita-2018-hate-speech/) and [IronITA](https://md.research.elamur.ai/free/en/papers/cignarella-2018-overview-evalita-2018-task-irony/), and perspectivist irony corpora.
- [Cornelia Caragea](https://md.research.elamur.ai/free/en/authors/cornelia-caragea/) (University of Illinois Chicago) and [Yingjie Li](https://md.research.elamur.ai/free/en/authors/yingjie-li/) work on stance: [P-Stance](https://md.research.elamur.ai/free/en/papers/li-2021-p-stance-large-dataset-stance/), [COVID-19 stance](https://md.research.elamur.ai/free/en/papers/glandt-2021-stance-detection-covid-19-tweets/) and [stance with sentiment lexicons](https://md.research.elamur.ai/free/en/papers/li-2019-multi-task-stance-detection-sentiment/).
- [Walid Magdy](https://md.research.elamur.ai/free/en/authors/walid-magdy/) (University of Edinburgh) and [Ibrahim Abu Farha](https://md.research.elamur.ai/free/en/authors/ibrahim-abu-farha/) built [ArSarcasm](https://md.research.elamur.ai/free/en/papers/farha-2020-arabic-sentiment-analysis-sarcasm-detection/), organised [iSarcasmEval](https://md.research.elamur.ai/free/en/papers/farha-2022-semeval-2022-task-6-isarcasmeval/) and [benchmarked transformers for Arabic sentiment](https://md.research.elamur.ai/free/en/papers/farha-2021-benchmarking-transformer-based-language-models/).
- [Smaranda Muresan](https://md.research.elamur.ai/free/en/authors/smaranda-muresan/) (Columbia University) and [Debanjan Ghosh](https://md.research.elamur.ai/free/en/authors/debanjan-ghosh/) studied sarcasm from [literal versus sarcastic word senses](https://md.research.elamur.ai/free/en/papers/ghosh-2015-sarcastic-or-not-word-embeddings/) to [the role of conversation context](https://md.research.elamur.ai/free/en/papers/ghosh-2017-role-conversation-context-sarcasm-detection/).
- [Véronique Hoste](https://md.research.elamur.ai/free/en/authors/veronique-hoste/) (Ghent University) co-organised [SemEval-2018 irony detection](https://md.research.elamur.ai/free/en/papers/hee-2018-semeval-2018-task-3-irony/) and studied [how irony is realised on Twitter](https://md.research.elamur.ai/free/en/papers/hee-2016-exploring-realization-irony-twitter-data/).
- [Paolo Rosso](https://md.research.elamur.ai/free/en/authors/paolo-rosso/) (Universitat Politècnica de València) contributed to hate speech, irony and [stance in fake news](https://md.research.elamur.ai/free/en/papers/ghanem-2018-stance-detection-fake-news-combined/).
- [Rada Mihalcea](https://md.research.elamur.ai/free/en/authors/rada-mihalcea/) co-authored multimodal work from [MELD](https://md.research.elamur.ai/free/en/papers/poria-2018-meld-multimodal-multi-party-dataset/) to [multimodal sarcasm](https://md.research.elamur.ai/free/en/papers/castro-2019-towards-multimodal-sarcasm-detection-obviously/) and [empathetic response generation](https://md.research.elamur.ai/free/en/papers/majumder-2020-mime-mimicking-emotions-empathetic-response/).
- [Lidong Bing](https://md.research.elamur.ai/free/en/authors/lidong-bing/) appears on [RAM](https://md.research.elamur.ai/free/en/papers/chen-2017-recurrent-attention-network-memory-aspect/), [TNet](https://md.research.elamur.ai/free/en/papers/li-2018-transformation-networks-target-oriented-sentiment/) and the [LLM-era reality check](https://md.research.elamur.ai/free/en/papers/zhang-2023-sentiment-analysis-era-large-language/); [Roman Klinger](https://md.research.elamur.ai/free/en/authors/roman-klinger/) (University of Bamberg) on [fine-grained emotion annotation](https://md.research.elamur.ai/free/en/papers/schuff-2017-annotation-modelling-analysis-fine-grained/) and irony features.

Authors with the most papers on this topic Pushpak Bhattacharyya leads with 21 papers; the 12 authors shown wrote 148 of the papers on this topic between them, counting each paper once per author.

**The numbers**

| Author | Papers here · their citations |
| --- | --- |
| [Pushpak Bhattacharyya](https://md.research.elamur.ai/free/en/authors/pushpak-bhattacharyya/) | 21 papers · 1,780 citations |
| [Saif Mohammad](https://md.research.elamur.ai/free/en/authors/saif-mohammad/) | 16 papers · 6,586 citations |
| [Aditya Joshi](https://md.research.elamur.ai/free/en/authors/aditya-joshi/) | 15 papers · 1,167 citations |
| [Soujanya Poria](https://md.research.elamur.ai/free/en/authors/soujanya-poria/) | 13 papers · 10,047 citations |
| [Walid Magdy](https://md.research.elamur.ai/free/en/authors/walid-magdy/) | 12 papers · 914 citations |
| [Preslav Nakov](https://md.research.elamur.ai/free/en/authors/preslav-nakov/) | 11 papers · 5,456 citations |
| [Cristina Bosco](https://md.research.elamur.ai/free/en/authors/cristina-bosco/) | 11 papers · 2,159 citations |
| [Viviana Patti](https://md.research.elamur.ai/free/en/authors/viviana-patti/) | 11 papers · 1,917 citations |
| [Erik Cambria](https://md.research.elamur.ai/free/en/authors/erik-cambria/) | 10 papers · 7,821 citations |
| [Paolo Rosso](https://md.research.elamur.ai/free/en/authors/paolo-rosso/) | 10 papers · 1,841 citations |
| [Smaranda Muresan](https://md.research.elamur.ai/free/en/authors/smaranda-muresan/) | 9 papers · 1,329 citations |
| [Mark James Carman](https://md.research.elamur.ai/free/en/authors/mark-carman/) | 9 papers · 712 citations |

## Institutions and places

Work on this topic is spread widely. Among the current homes of its authors, the [University of Edinburgh](https://md.research.elamur.ai/free/en/universities/university-of-edinburgh/) comes first, followed by the [National University of Singapore](https://md.research.elamur.ai/free/en/universities/national-university-of-singapore/) and [Nanyang Technological University](https://md.research.elamur.ai/free/en/universities/nanyang-technological-university/), the two Singapore institutions reflecting the multimodal and conversational emotion work of the Poria–Cambria network. China is strongly represented through the [Chinese Academy of Sciences](https://md.research.elamur.ai/free/en/government/chinese-academy-of-sciences/), [Harbin Institute of Technology](https://md.research.elamur.ai/free/en/universities/harbin-institute-of-technology/), [Hefei University of Technology](https://md.research.elamur.ai/free/en/universities/hefei-university-of-technology/), [Peking University](https://md.research.elamur.ai/free/en/universities/peking-university/) and [Beijing University of Posts and Telecommunications](https://md.research.elamur.ai/free/en/universities/beijing-university-of-posts-and-telecommunications/). In North America the list includes [Johns Hopkins University](https://md.research.elamur.ai/free/en/universities/johns-hopkins-university/), [Columbia University](https://md.research.elamur.ai/free/en/universities/columbia-university/), [Stanford University](https://md.research.elamur.ai/free/en/universities/stanford-university/), the [Massachusetts Institute of Technology](https://md.research.elamur.ai/free/en/universities/massachusetts-institute-of-technology/), the [University of Notre Dame](https://md.research.elamur.ai/free/en/universities/university-of-notre-dame/), [Google](https://md.research.elamur.ai/free/en/companies/google/) and the [National Research Council Canada](https://md.research.elamur.ai/free/en/government/national-research-council-canada/), home of the NRC lexicons. Elsewhere, the [Mohamed bin Zayed University of Artificial Intelligence](https://md.research.elamur.ai/free/en/universities/mohamed-bin-zayed-university-of-artificial-intelligence/), the [Indian Institute of Technology Bombay](https://md.research.elamur.ai/free/en/universities/indian-institute-of-technology-bombay/), [The University of Tokyo](https://md.research.elamur.ai/free/en/universities/the-university-of-tokyo/) and [RIKEN](https://md.research.elamur.ai/free/en/labs/riken/) appear among the leading institutions.

Where the authors on this topic work now University of Edinburgh is the current home of authors of 27 papers on this topic; counted are the authors with a page here, by their latest institution.

**The numbers**

| Institution | Papers |
| --- | --- |
| [University of Edinburgh](https://md.research.elamur.ai/free/en/universities/university-of-edinburgh/) | 27 papers |
| [Nanyang Technological University](https://md.research.elamur.ai/free/en/universities/nanyang-technological-university/) | 23 papers |
| [Chinese Academy of Sciences](https://md.research.elamur.ai/free/en/government/chinese-academy-of-sciences/) | 22 papers |
| [National University of Singapore](https://md.research.elamur.ai/free/en/universities/national-university-of-singapore/) | 22 papers |
| [Indian Institute of Technology Bombay](https://md.research.elamur.ai/free/en/universities/indian-institute-of-technology-bombay/) | 21 papers |
| [Mohamed bin Zayed University of Artificial Intelligence](https://md.research.elamur.ai/free/en/universities/mohamed-bin-zayed-university-of-artificial-intelligence/) | 18 papers |
| [Peking University](https://md.research.elamur.ai/free/en/universities/peking-university/) | 18 papers |
| [Columbia University](https://md.research.elamur.ai/free/en/universities/columbia-university/) | 16 papers |
| [Harbin Institute of Technology](https://md.research.elamur.ai/free/en/universities/harbin-institute-of-technology/) | 16 papers |
| [National Research Council Canada](https://md.research.elamur.ai/free/en/government/national-research-council-canada/) | 16 papers |

By country, the United States leads, followed by China, the United Kingdom, Singapore and Germany; Italy and Spain stand out for European work on hate speech and irony, and the United Arab Emirates for Arabic and offensive-language research.

Countries of the authors on this topic United States comes first: authors working there now wrote 206 papers on this topic.

**The numbers**

| Country | Papers |
| --- | --- |
| [United States](https://md.research.elamur.ai/free/en/countries/united-states/) | 206 papers |
| [China](https://md.research.elamur.ai/free/en/countries/china/) | 119 papers |
| [United Kingdom](https://md.research.elamur.ai/free/en/countries/united-kingdom/) | 91 papers |
| [Germany](https://md.research.elamur.ai/free/en/countries/germany/) | 42 papers |
| [Singapore](https://md.research.elamur.ai/free/en/countries/singapore/) | 42 papers |
| [India](https://md.research.elamur.ai/free/en/countries/india/) | 39 papers |
| [Italy](https://md.research.elamur.ai/free/en/countries/italy/) | 38 papers |
| [Canada](https://md.research.elamur.ai/free/en/countries/canada/) | 34 papers |
| [Australia](https://md.research.elamur.ai/free/en/countries/australia/) | 29 papers |
| [Spain](https://md.research.elamur.ai/free/en/countries/spain/) | 28 papers |

The venues show the role of shared tasks and resources. The LREC conferences, the SemEval workshops, the Arabic NLP workshops (WANLP and ArabicNLP) and the Workshop on Figurative Language Processing sit next to the main ACL, EMNLP and COLING conferences and the journal Computational Linguistics.

Where the papers on this topic were published Annual Meeting of the Association for Computational Linguistics (ACL) carries the most, 77 of 609.

**The numbers**

| Venue | Papers |
| --- | --- |
| [Annual Meeting of the Association for Computational Linguistics (ACL)](https://md.research.elamur.ai/free/en/venues/acl/) | 77 papers |
| [International Workshop on Semantic Evaluation (SemEval)](https://md.research.elamur.ai/free/en/venues/semeval/) | 70 papers |
| [Conference on Empirical Methods in Natural Language Processing (EMNLP)](https://md.research.elamur.ai/free/en/venues/emnlp/) | 65 papers |
| [Language Resources and Evaluation Conference (LREC)](https://md.research.elamur.ai/free/en/venues/lrec/) | 56 papers |
| [International Conference on Computational Linguistics (COLING)](https://md.research.elamur.ai/free/en/venues/coling/) | 29 papers |
| [Findings of the Association for Computational Linguistics](https://md.research.elamur.ai/free/en/venues/findings/) | 21 papers |
| [Arabic Natural Language Processing Workshop (WANLP)](https://md.research.elamur.ai/free/en/venues/wanlp/) | 20 papers |
| [Workshop on Computational Approaches to Subjectivity, Sentiment and Social Media Analysis](https://md.research.elamur.ai/free/en/venues/wassa/) | 16 papers |
| [Workshop on Figurative Language Processing](https://md.research.elamur.ai/free/en/venues/figlang/) | 15 papers |
| [Conference of the North American Chapter of the Association for Computational Linguistics (NAACL)](https://md.research.elamur.ai/free/en/venues/naacl/) | 14 papers |

## Open questions

- **What is the label?** Sentiment, emotion, offensiveness and sarcasm are judgements that people disagree on for good reasons. Whether to aggregate annotations into one gold label or to model the spread of opinions, as perspectivist corpora do, is unresolved.
- **Implicit sentiment.** Many evaluations carry no sentiment words: "the battery lasted until lunch" or a positive word about a bad situation. Models still struggle with sentiment that depends on world knowledge, implicature and context.
- **Neutral and mixed cases.** Error analyses again and again find that the neutral class and mixed opinions cause most mistakes, from [phrase-level polarity](https://md.research.elamur.ai/free/en/papers/wilson-2009-recognizing-contextual-polarity-exploration-features/) to [target-dependent LSTMs](https://md.research.elamur.ai/free/en/papers/tang-2016-effective-lstms-target-dependent-sentiment/).
- **Bias and fairness.** Systems score text differently depending on the names, dialects and identity terms it contains, and annotated data carries the biases of its annotators. How to measure and reduce this without suppressing minority voices is an open problem ([Social Biases in NLP Models](https://md.research.elamur.ai/free/en/papers/hutchinson-2020-social-biases-nlp-models-as/), [racial bias in abusive language datasets](https://md.research.elamur.ai/free/en/papers/davidson-2019-racial-bias-hate-speech-abusive/)).
- **Implicit and evolving abuse.** Coded language, irony and new slurs evade classifiers trained on explicit examples ([Latent Hatred](https://md.research.elamur.ai/free/en/papers/elsherief-2021-latent-hatred-benchmark-understanding-implicit/), [ToxiGen](https://md.research.elamur.ai/free/en/papers/hartvigsen-2022-toxigen-large-scale-machine-generated/)), and behavioural tests such as [HateCheck](https://md.research.elamur.ai/free/en/papers/rottger-2021-hatecheck-functional-tests-hate-speech/) show gaps that aggregate scores hide.
- **Languages and dialects.** Most resources are English; Arabic dialects, code-mixed text and African languages are catching up but remain thinly covered.
- **Large language models.** Prompted general models are strong on simple polarity but less reliable on structured, fine-grained and low-resource tasks. How to evaluate them fairly, and whether task-specific models still have a place, is debated.
- **Emotion theory.** Categorical and dimensional models of emotion still compete, and the choice affects both annotation and evaluation.

## Neighbouring topics

The largest overlap in the library is with [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/), reflecting a large body of Arabic sentiment, sarcasm and offensive-language work presented at the Arabic NLP workshops; related work on [Hebrew & Aramaic](https://md.research.elamur.ai/free/en/topics/hebrew-aramaic/) is a smaller part of it. [Bias, fairness & safety](https://md.research.elamur.ai/free/en/topics/bias-fairness-safety/) shares the hate speech and fairness-audit papers. [Vision & language](https://md.research.elamur.ai/free/en/topics/multimodal/) shares multimodal sentiment, sarcasm and hateful memes, and [Speech & spoken language](https://md.research.elamur.ai/free/en/topics/speech/) shares the acoustic side of emotion recognition. [Dialogue systems](https://md.research.elamur.ai/free/en/topics/dialogue-systems/) connects through emotion recognition in conversation and empathetic response generation.

[Multilingual & cross-lingual NLP](https://md.research.elamur.ai/free/en/topics/multilingual-nlp/) and [Low-resource languages](https://md.research.elamur.ai/free/en/topics/low-resource-languages/) cover cross-lingual transfer and new-language benchmarks. [Evaluation & benchmarks](https://md.research.elamur.ai/free/en/topics/evaluation-benchmarks/) shares the shared-task papers. [Language models & pretraining](https://md.research.elamur.ai/free/en/topics/language-models/) supplies the encoders, and [Reasoning](https://md.research.elamur.ai/free/en/topics/reasoning/) touches stance, sarcasm and implicit sentiment. On the linguistic side, [Discourse analysis](https://md.research.elamur.ai/free/en/topics/discourse-analysis/) (for instance [RST-based document sentiment](https://md.research.elamur.ai/free/en/papers/bhatia-2015-better-document-level-sentiment-analysis/)), [Speech act theory](https://md.research.elamur.ai/free/en/topics/speech-act-theory/), [Implicature](https://md.research.elamur.ai/free/en/topics/implicature/) and [Idioms & multiword expressions](https://md.research.elamur.ai/free/en/topics/multiword-expressions/) explain how evaluative meaning arises beyond single words. [Summarization](https://md.research.elamur.ai/free/en/topics/summarization/) meets the topic in opinion summarisation, [Interpretability](https://md.research.elamur.ai/free/en/topics/interpretability/) in explaining classifiers, and [Knowledge graphs & ontologies](https://md.research.elamur.ai/free/en/topics/knowledge-graphs/) in commonsense-enhanced stance and emotion models. Most papers here stand on the [Linguistics & languages](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) and [Natural language processing](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) shelves, with a few on [Machine translation](https://md.research.elamur.ai/free/en/topics/machine-translation/), [AI & machine learning](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) and [Society & ethics](https://md.research.elamur.ai/free/en/shelves/social-sciences/sociology-and-political-science/).

Neighbouring topics: papers shared Semitic languages shares the most: 145 of the 609 papers here (24%) stand on both.

**The numbers**

| Topic | Papers shared |
| --- | --- |
| [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/) | 145 shared · 24% |
| [English](https://md.research.elamur.ai/free/en/topics/english/) | 92 shared · 15% |
| [Bias, fairness & safety](https://md.research.elamur.ai/free/en/topics/bias-fairness-safety/) | 50 shared · 8% |
| [Vision & language](https://md.research.elamur.ai/free/en/topics/multimodal/) | 35 shared · 6% |
| [Multilingual & cross-lingual NLP](https://md.research.elamur.ai/free/en/topics/multilingual-nlp/) | 30 shared · 5% |
| [Evaluation & benchmarks](https://md.research.elamur.ai/free/en/topics/evaluation-benchmarks/) | 20 shared · 3% |
| [Dialogue systems](https://md.research.elamur.ai/free/en/topics/dialogue-systems/) | 19 shared · 3% |
| [Language models & pretraining](https://md.research.elamur.ai/free/en/topics/language-models/) | 13 shared · 2% |
| [Idioms & multiword expressions](https://md.research.elamur.ai/free/en/topics/multiword-expressions/) | 9 shared · 1% |
| [Reasoning](https://md.research.elamur.ai/free/en/topics/reasoning/) | 9 shared · 1% |

The shelves this topic stands on Artificial Intelligence: 590 of the 609 papers here.

**The numbers**

| Shelf | Papers |
| --- | --- |
| [Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) | 590 papers |
| [Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) | 438 papers |
| [Experimental and Cognitive Psychology](https://md.research.elamur.ai/free/en/shelves/psychology/experimental-and-cognitive-psychology/) | 126 papers |
| [Social Psychology](https://md.research.elamur.ai/free/en/shelves/psychology/social-psychology/) | 49 papers |
| [Sociology and Political Science](https://md.research.elamur.ai/free/en/shelves/social-sciences/sociology-and-political-science/) | 41 papers |
| [Computer Vision and Pattern Recognition](https://md.research.elamur.ai/free/en/shelves/computer-science/computer-vision-and-pattern-recognition/) | 23 papers |
| [Communication](https://md.research.elamur.ai/free/en/shelves/social-sciences/communication/) | 21 papers |
| [General Social Sciences](https://md.research.elamur.ai/free/en/shelves/social-sciences/general-social-sciences/) | 12 papers |
| [Safety Research](https://md.research.elamur.ai/free/en/shelves/social-sciences/safety-research/) | 12 papers |
| [Human-Computer Interaction](https://md.research.elamur.ai/free/en/shelves/computer-science/human-computer-interaction/) | 8 papers |
| [Literature and Literary Theory](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/literature-and-literary-theory/) | 8 papers |
| [Signal Processing](https://md.research.elamur.ai/free/en/shelves/computer-science/signal-processing/) | 5 papers |

## Key works in this library

- [An Argument for Basic Emotions](https://md.research.elamur.ai/free/en/papers/ekman-1992-argument-basic-emotions/) — the psychological case for a few discrete emotions.
- [Thumbs up?](https://md.research.elamur.ai/free/en/papers/pang-2002-thumbs-up-sentiment-classification-machine/) — supervised review polarity, the field's starting point.
- [A Sentimental Education](https://md.research.elamur.ai/free/en/papers/pang-2004-sentimental-education-sentiment-analysis-subjectivity/) — subjectivity filtering with minimum cuts.
- [Seeing Stars](https://md.research.elamur.ai/free/en/papers/pang-2005-seeing-stars-exploiting-class-relationships/) — rating inference on ordinal scales.
- [Determining the Sentiment of Opinions](https://md.research.elamur.ai/free/en/papers/kim-2004-determining-sentiment-opinions/) — opinion holders and WordNet-grown lexicons.
- [Recognizing Contextual Polarity](https://md.research.elamur.ai/free/en/papers/wilson-2005-recognizing-contextual-polarity-phrase-level/) — prior versus contextual polarity.
- [SENTIWORDNET](https://md.research.elamur.ai/free/en/papers/esuli-2006-sentiwordnet-publicly-available-lexical-resource/) — graded sentiment for every WordNet sense.
- [WordNet Affect](https://md.research.elamur.ai/free/en/papers/strapparava-2004-wordnet-affect-affective-extension-wordnet/) — an affective label layer over WordNet.
- [Biographies, Bollywood, Boom-boxes](https://md.research.elamur.ai/free/en/papers/blitzer-2007-biographies-bollywood-boom-boxes-blenders/) — domain adaptation and the Amazon review corpus.
- [Cross-lingual co-training](https://md.research.elamur.ai/free/en/papers/wan-2009-co-training-cross-lingual-sentiment/) — machine translation as a second view.
- [Twitter as a sentiment corpus](https://md.research.elamur.ai/free/en/papers/pak-2010-twitter-as-corpus-sentiment-analysis/) — emoticon-labelled tweets as training data.
- [Lexicon-based methods (SO-CAL)](https://md.research.elamur.ai/free/en/papers/taboada-2011-lexicon-based-methods-sentiment-analysis/) — the SO-CAL dictionary approach.
- [Word vectors for sentiment](https://md.research.elamur.ai/free/en/papers/maas-2011-learning-word-vectors-sentiment-analysis/) — polarity-aware vectors and the IMDB benchmark.
- [Baselines and Bigrams](https://md.research.elamur.ai/free/en/papers/wang-2012-baselines-bigrams-simple-good-sentiment/) — NBSVM, a strong simple baseline.
- [Sarcasm as contrast](https://md.research.elamur.ai/free/en/papers/riloff-2013-sarcasm-as-contrast-between-positive/) — bootstrapped sarcasm patterns.
- [SemEval-2014 ABSA](https://md.research.elamur.ai/free/en/papers/pontiki-2014-semeval-2014-task-4-aspect/) — the founding ABSA benchmark.
- [Attention-based LSTM for aspects](https://md.research.elamur.ai/free/en/papers/wang-2016-attention-based-lstm-aspect-level/) — aspect-conditioned attention.
- [Hateful symbols or people?](https://md.research.elamur.ai/free/en/papers/waseem-2016-hateful-symbols-or-hateful-people/) — an early public hate speech corpus.
- [Tensor Fusion Network](https://md.research.elamur.ai/free/en/papers/zadeh-2017-tensor-fusion-network-multimodal-sentiment/) — explicit cross-modal interactions.
- [DeepMoji](https://md.research.elamur.ai/free/en/papers/felbo-2017-millions-emoji-occurrences-learn-any/) — pretraining on emoji as noisy labels.
- [MELD](https://md.research.elamur.ai/free/en/papers/poria-2018-meld-multimodal-multi-party-dataset/) — emotion in multi-party conversation.
- [Affect in Tweets](https://md.research.elamur.ai/free/en/papers/mohammad-2018-semeval-2018-task-1-affect/) — emotion intensity in three languages.
- [Racial bias in hate speech detection](https://md.research.elamur.ai/free/en/papers/sap-2019-risk-racial-bias-hate-speech/) — dialect bias in labels and models.
- [ASGCN](https://md.research.elamur.ai/free/en/papers/zhang-2019-aspect-based-sentiment-classification-aspect/) — dependency graphs for aspects.
- [A reality check for sentiment analysis with LLMs](https://md.research.elamur.ai/free/en/papers/zhang-2023-sentiment-analysis-era-large-language/) — how prompted LLMs fare on sentiment tasks.

## Papers (609)

**1** [2](https://md.research.elamur.ai/free/en/topics/sentiment-emotion/page/2/) … [7](https://md.research.elamur.ai/free/en/topics/sentiment-emotion/page/7/) [next →](https://md.research.elamur.ai/free/en/topics/sentiment-emotion/page/2/)

1. [**An Argument for Basic Emotions**Paul Ekman*1992 · 9,903 citations*](https://md.research.elamur.ai/free/en/papers/ekman-1992-argument-basic-emotions/)
2. **[Thumbs up? Sentiment Classification using Machine Learning Techniques](https://md.research.elamur.ai/free/en/papers/pang-2002-thumbs-up-sentiment-classification-machine/) [Conference on Empirical Methods in Natural Language Processing (EMNLP)](https://md.research.elamur.ai/free/en/venues/emnlp/)**Bo Pang, Lillian Lee, Shivakumar Vaithyanathan*2002 · 7,762 citations*
3. **[Learning Word Vectors for Sentiment Analysis](https://md.research.elamur.ai/free/en/papers/maas-2011-learning-word-vectors-sentiment-analysis/) [Annual Meeting of the Association for Computational Linguistics (ACL)](https://md.research.elamur.ai/free/en/venues/acl/)**Andrew L. Maas, Raymond E. Daly, Peter T. Pham, et al.*2011 · 4,290 citations*
4. **[Recognizing Contextual Polarity in Phrase-Level Sentiment Analysis](https://md.research.elamur.ai/free/en/papers/wilson-2005-recognizing-contextual-polarity-phrase-level/) [Human Language Technology Conference (HLT)](https://md.research.elamur.ai/free/en/venues/hlt/)**Theresa Wilson, Janyce Wiebe, Paul Hoffmann*2005 · 3,991 citations*
5. **[A Sentimental Education: Sentiment Analysis Using Subjectivity Summarization Based on Minimum Cuts](https://md.research.elamur.ai/free/en/papers/pang-2004-sentimental-education-sentiment-analysis-subjectivity/) [Annual Meeting of the Association for Computational Linguistics (ACL)](https://md.research.elamur.ai/free/en/venues/acl/)**Bo Pang, Lillian Lee*2004 · 3,849 citations*
6. **[Lexicon-Based Methods for Sentiment Analysis](https://md.research.elamur.ai/free/en/papers/taboada-2011-lexicon-based-methods-sentiment-analysis/) [Computational Linguistics](https://md.research.elamur.ai/free/en/venues/cl/)**Maite Taboada, Julian Brooke, Milan Tofiloski, et al.*2011 · 3,563 citations*
7. **[SentiWordNet 3.0: An Enhanced Lexical Resource for Sentiment Analysis and Opinion Mining](https://md.research.elamur.ai/free/en/papers/baccianella-2010-sentiwordnet-3-0-enhanced-lexical/) [Language Resources and Evaluation Conference (LREC)](https://md.research.elamur.ai/free/en/venues/lrec/)**Stefano Baccianella, Andrea Esuli, Fabrizio Sebastiani*2010 · 3,076 citations*
8. **[Twitter as a Corpus for Sentiment Analysis and Opinion Mining](https://md.research.elamur.ai/free/en/papers/pak-2010-twitter-as-corpus-sentiment-analysis/) [Language Resources and Evaluation Conference (LREC)](https://md.research.elamur.ai/free/en/venues/lrec/)**Alexander Pak, Patrick Paroubek*2010 · 2,915 citations*
9. **[Seeing Stars: Exploiting Class Relationships for Sentiment Categorization with Respect to Rating Scales](https://md.research.elamur.ai/free/en/papers/pang-2005-seeing-stars-exploiting-class-relationships/) [Annual Meeting of the Association for Computational Linguistics (ACL)](https://md.research.elamur.ai/free/en/venues/acl/)**Bo Pang, Lillian Lee*2005 · 2,733 citations*
10. **[SENTIWORDNET: A Publicly Available Lexical Resource for Opinion Mining](https://md.research.elamur.ai/free/en/papers/esuli-2006-sentiwordnet-publicly-available-lexical-resource/) [Language Resources and Evaluation Conference (LREC)](https://md.research.elamur.ai/free/en/venues/lrec/)**Andrea Esuli, Fabrizio Sebastiani*2006 · 2,722 citations*
11. **[Attention-based LSTM for Aspect-level Sentiment Classification](https://md.research.elamur.ai/free/en/papers/wang-2016-attention-based-lstm-aspect-level/) [Conference on Empirical Methods in Natural Language Processing (EMNLP)](https://md.research.elamur.ai/free/en/venues/emnlp/)**Yequan Wang, Minlie Huang, Xiaoyan Zhu, et al.*2016 · 2,619 citations*
12. **[SemEval-2014 Task 4: Aspect Based Sentiment Analysis](https://md.research.elamur.ai/free/en/papers/pontiki-2014-semeval-2014-task-4-aspect/) [International Workshop on Semantic Evaluation (SemEval)](https://md.research.elamur.ai/free/en/venues/semeval/)**Maria Pontiki, Dimitris Galanis, John Pavlopoulos, et al.*2014 · 2,409 citations*
13. **[Biographies, Bollywood, Boom-boxes and Blenders: Domain Adaptation for Sentiment Classification](https://md.research.elamur.ai/free/en/papers/blitzer-2007-biographies-bollywood-boom-boxes-blenders/) [Annual Meeting of the Association for Computational Linguistics (ACL)](https://md.research.elamur.ai/free/en/venues/acl/)**John Blitzer, Mark Dredze, Fernando Pereira*2007 · 2,329 citations*
14. **[Hateful Symbols or Hateful People? Predictive Features for Hate Speech Detection on Twitter](https://md.research.elamur.ai/free/en/papers/waseem-2016-hateful-symbols-or-hateful-people/) [NAACL Student Research Workshop](https://md.research.elamur.ai/free/en/venues/naacl-student-research-workshop/)**Zeerak Waseem, Dirk Hovy*2016 · 2,167 citations*
15. **[Tensor Fusion Network for Multimodal Sentiment Analysis](https://md.research.elamur.ai/free/en/papers/zadeh-2017-tensor-fusion-network-multimodal-sentiment/) [Conference on Empirical Methods in Natural Language Processing (EMNLP)](https://md.research.elamur.ai/free/en/venues/emnlp/)**Amir Zadeh, Minghai Chen, Soujanya Poria, et al.*2017 · 2,113 citations*
16. **[A Survey on Hate Speech Detection using Natural Language Processing](https://md.research.elamur.ai/free/en/papers/schmidt-2017-survey-hate-speech-detection-natural/) [International Workshop on Natural Language Processing for Social Media](https://md.research.elamur.ai/free/en/venues/socialnlp/)**Anna Schmidt, Michael Wiegand*2017 · 1,805 citations*
17. **[Document Modeling with Gated Recurrent Neural Network for Sentiment Classification](https://md.research.elamur.ai/free/en/papers/tang-2015-document-modeling-gated-recurrent-neural/) [Conference on Empirical Methods in Natural Language Processing (EMNLP)](https://md.research.elamur.ai/free/en/venues/emnlp/)**Duyu Tang, Bing Qin, Ting Liu*2015 · 1,684 citations*
18. **[Determining the Sentiment of Opinions](https://md.research.elamur.ai/free/en/papers/kim-2004-determining-sentiment-opinions/) [International Conference on Computational Linguistics (COLING)](https://md.research.elamur.ai/free/en/venues/coling/)**Soo-Min Kim, Eduard Hovy*2004 · 1,679 citations*
19. **[MELD: A Multimodal Multi-Party Dataset for Emotion Recognition in Conversations](https://md.research.elamur.ai/free/en/papers/poria-2018-meld-multimodal-multi-party-dataset/) [Annual Meeting of the Association for Computational Linguistics (ACL)](https://md.research.elamur.ai/free/en/venues/acl/)**Soujanya Poria, Devamanyu Hazarika, Navonil Majumder, et al.*2018 · 1,671 citations*
20. **[SemEval-2016 Task 5: Aspect Based Sentiment Analysis](https://md.research.elamur.ai/free/en/papers/pontiki-2016-semeval-2016-task-5-aspect/) [International Workshop on Semantic Evaluation (SemEval)](https://md.research.elamur.ai/free/en/venues/semeval/)**Maria Pontiki, Dimitris Galanis, Haris Papageorgiou, et al.*2016 · 1,659 citations*
21. **[WordNet Affect: an Affective Extension of WordNet](https://md.research.elamur.ai/free/en/papers/strapparava-2004-wordnet-affect-affective-extension-wordnet/) [Language Resources and Evaluation Conference (LREC)](https://md.research.elamur.ai/free/en/venues/lrec/)**Carlo Strapparava, Alessandro Valitutti*2004 · 1,577 citations*
22. **[Multimodal Language Analysis in the Wild: CMU-MOSEI Dataset and Interpretable Dynamic Fusion Graph](https://md.research.elamur.ai/free/en/papers/zadeh-2018-multimodal-language-analysis-wild-cmu/) [Annual Meeting of the Association for Computational Linguistics (ACL)](https://md.research.elamur.ai/free/en/venues/acl/)**AmirAli Bagher Zadeh, Paul Pu Liang, Soujanya Poria, et al.*2018 · 1,509 citations*
23. **[Sentiment Analysis of Twitter Data](https://md.research.elamur.ai/free/en/papers/agarwal-2011-sentiment-analysis-twitter-data/) [Workshop on Language in Social Media](https://md.research.elamur.ai/free/en/venues/workshop-on-language-in-social-media/)**Apoorv Agarwal, Boyi Xie, Ilia Vovsha, et al.*2011 · 1,461 citations*
24. **[SemEval-2015 Task 12: Aspect Based Sentiment Analysis](https://md.research.elamur.ai/free/en/papers/pontiki-2015-semeval-2015-task-12-aspect/) [International Workshop on Semantic Evaluation (SemEval)](https://md.research.elamur.ai/free/en/venues/semeval/)**Maria Pontiki, Dimitris Galanis, Haris Papageorgiou, et al.*2015 · 1,448 citations*
25. **[Semi-Supervised Recursive Autoencoders for Predicting Sentiment Distributions](https://md.research.elamur.ai/free/en/papers/socher-2011-semi-supervised-recursive-autoencoders-predicting/) [Conference on Empirical Methods in Natural Language Processing (EMNLP)](https://md.research.elamur.ai/free/en/venues/emnlp/)**Richard Socher, Jeffrey Pennington, Eric H. Huang, et al.*2011 · 1,420 citations*

**1** [2](https://md.research.elamur.ai/free/en/topics/sentiment-emotion/page/2/) … [7](https://md.research.elamur.ai/free/en/topics/sentiment-emotion/page/7/) [next →](https://md.research.elamur.ai/free/en/topics/sentiment-emotion/page/2/)

Updated 24 October 2026

- [← previous topic Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/)
- [next topic → Speech & spoken language](https://md.research.elamur.ai/free/en/topics/speech/)
- [↑ back to topics](https://md.research.elamur.ai/free/en/topics/)