Lexicons for Sentiment, Affect, and Connotation

Emotion, sentiment, personality, mood and attitudes. Closely related is subjectivity: the study of a speaker’s or writer’s evaluations, opinions, emotions and speculations.

Attitudes: what people like or dislike, e.g. from reviews, editorials or tweets.

Emotion/mood: a confused or engaged student in a tutoring system, a frustrated caller, signs of depression. Stance: friendliness or awkwardness, conversational hot spots for meeting summarization. Personality: conversational agents work better when they match users’ personality expectations.

Lists of words carrying strong cues to affect or sentiment, used instead of all words as features. They presuppose that words have connotations: aspects of meaning related to a writer’s or reader’s emotions, sentiment, opinions or evaluations.

(1) Basic emotions: a small fixed set of atomic units from which others derive (Ekman, Plutchik). (2) Emotion as a point in a 2- or 3-dimensional space (valence, arousal, dominance).

Surprise, happiness, anger, fear, disgust, sadness, proposed to be universal across cultures.

8 basic emotions in four opposing pairs, plus emotions derived from them:

joy	↔	sadness
anger	↔	fear
trust	↔	disgust
anticipation	↔	surprise

Valence: the pleasantness of the stimulus. Arousal: the level of alertness, activeness or energy it provokes. Dominance: the degree of control exerted by the stimulus or the emotion.

It is a special case: the valence dimension (how pleasant or unpleasant a word is) is often used directly as a measure of sentiment.

Lexicon models fix a word’s affect regardless of context, dialect or culture. Appraisal theory treats emotions as complex cognitive processes: a person considers how an event fits their goals, given its agency, certainty, urgency, novelty and control.

The General Inquirer (oldest, 1966, from content analysis), the MPQA Subjectivity lexicon (also labels strongly/weakly subjective), and the Hu & Liu (2004) polarity lexicon (from product reviews, bootstrapped from WordNet).

Scores for valence, arousal and dominance for about 20,000 words (Mohammad 2018), e.g. high valence vacation, low torture; high arousal enraged, low napping; high dominance powerful, low weak.

The NRC Word-Emotion Association Lexicon: binary associations of about 14,000 words with Plutchik’s 8 emotions plus positive/negative, e.g. reward: anticipation, joy, surprise, trust, positive.

Real-valued association scores (rather than binary) for anger, fear, joy and sadness, e.g. outraged is very high on anger, nurture very low.

Linguistic Inquiry and Word Count: 73 hand-built lexicons for social-psychological tasks, with categories like positive and negative emotion, anger, sadness, insight, inhibition, family, negation. Entries may be prefixes (appreciat*).

The General Inquirer’s dimensions (strong/weak, active/passive, overstated/understated; pleasure, pain, virtue, vice…) and concreteness ratings (Brysbaert et al.): crowdsourced 1–5 ratings, e.g. banana 5, belief about 1.

By crowdsourcing: break the task into small pieces and distribute them to many annotators.

Each item first asked a multiple-choice synonym question that primed the intended sense (e.g. which word is closest to startle: automobile, shake, honesty, entertain?), then asked for association with each of the 8 emotions.

Ratings on a not/weakly/moderately/strongly scale were filtered for outliers, the majority class taken (ties to the stronger), then mapped to binary: not/weak → 0, moderate/strong → 1.

Annotators get N items (usually 4) and choose the best (highest) and worst (lowest) on some property. A word’s score is the proportion of times chosen best minus the proportion chosen worst. Used for the NRC VAD lexicon.

A measure of annotation agreement: split the annotations in half and correlate the scores computed from the two halves.

Start from seed words defining two poles of a semantic axis (good vs bad) and label each word by its similarity to the two seed sets; two families are axis-based and graph-based.

Either start with one large seed lexicon and let induction adapt it to the domain, or choose different seeds per genre (e.g. separate seeds for general text, Twitter and finance, where profit, gains vs litigation, losses matter).

Embed the seed words (off-the-shelf, trained on the domain, or fine-tuned); take each pole’s centroid \(V^+\), \(V^-\); the axis is \(V_{axis} = V^+ - V^-\); score each word by \(\cos(E(w), V_{axis})\).

Apply a threshold (or similar method) to split the scored words into positive and negative lists.

(1) Build a graph connecting each word to its k nearest embedding neighbors, edges weighted by cosine-based similarity. (2) Pick positive and negative seeds. (3) Run random walks from each seed set; a word’s raw score is the probability of landing on it. (4) Combine: \(score^+(w) = \frac{raw^+(w)}{raw^+(w) + raw^-(w)}\). (5) Bootstrap over seed subsets for confidence.

Syntactic cues: adjectives frequently conjoined by and tend to share polarity (fair and legitimate, corrupt and brutal); those conjoined by but tend to be opposite (fair but brutal).

Morphological negation: adjectives with the same root differing by un-, im-, -less (adequate/inadequate, thoughtful/thoughtless) tend to have opposite polarity.

Iteratively: Lex⁺ gets synonyms of positive words and antonyms of negative words; Lex⁻ gets synonyms of negative words and antonyms of positive words.

Polarity assigned to WordNet synsets rather than words: each sense gets positive, negative and objective scores (summing to 1), e.g. estimable “deserving of esteem” is positive, “may be computed” purely objective. A classifier on glosses plus a random walk assigns the scores.

Embedding cosine, coordination (and/but), morphology, and thesaurus structure, scoring words by similarity to positive seeds and dissimilarity to negative ones.

Online review scores (stars): positive words appear more in 5-star reviews, negative in 1-star ones. A word can be represented by its distribution over rating levels, not just a binary polarity.

\(P(w \mid c) = \frac{count(w, c)}{\sum_{w \in C} count(w, c)}\), \(PottsScore(w) = \frac{P(w \mid c)}{\sum_c P(w \mid c)}\): the word’s likelihood in each rating category, normalized over categories. It is a variant of PMI without the log.

Strongly positive words (excellent): J-shape. Strongly negative (terrible): reverse J. Weakly polarized words: a hump, peaking below the middle (disappointing) or above it (good).

Emphatics (totally, absolutely) have a J-shape or a U-shape (frequent in both extremes); attenuators (somewhat, fairly, pretty) have a hump, favoring the middle of the scale.

They overemphasize differences in very rare words (tiny counts fluctuate, or are zero in one corpus) and often in very frequent words.

\(llr(w) = \log\frac{P^i(w)}{P^j(w)} = \log\frac{f^i(w)}{n^i} - \log\frac{f^j(w)}{n^j}\), with \(f\) the word count and \(n\) the corpus size.

\(lor(w) = \log\frac{f^i(w)}{n^i - f^i(w)} - \log\frac{f^j(w)}{n^j - f^j(w)}\): compares the word’s odds rather than its probabilities.

It adds scaled counts from a large background corpus as a prior (\(\alpha_w\), \(\alpha_0\)): \(\delta_w^{(i-j)} = \log\frac{f^i_w + \alpha_w}{n^i + \alpha_0 - (f^i_w + \alpha_w)} - \log\frac{f^j_w + \alpha_w}{n^j + \alpha_0 - (f^j_w + \alpha_w)}\), then divides by its estimated standard deviation, \(\sigma^2 \approx \frac{1}{f^i_w + \alpha_w} + \frac{1}{f^j_w + \alpha_w}\), giving a z-score.

It uses background-corpus counts as a prior (shrinking toward expected frequencies) and z-scores that control for the variance in a word’s frequency.

When there is not enough labeled training data. Count (optionally weighted) positive and negative lexicon words, \(f^+\) and \(f^-\): classify positive if \(f^+/f^- > \lambda\), negative if \(f^-/f^+ > \lambda\), otherwise neutral.

Lexicon counts (weighted or normalized) become features in a classifier alongside other lexical and non-lexical features.

Supervised classification: with large, similar training and test sets, all words or bigrams as features to logistic regression or an SVM is hard to beat; affect detection is treated as document classification.

Keep only words and phrases used by at least 1% of subjects, and keep 2- and 3-grams only with high PMI, \(\log\frac{p(phrase)}{\prod_{w \in phrase} p(w)} > 2 \cdot length\); counts were normalized by each subject’s total word use.

An indicator (1 if the text has any word from lexicon \(L\)), the count of tokens in \(L\), \(\sum_{w \in L} count(w)\), or a weighted count \(\sum_{w \in L} \theta^L_w count(w)\); counts may be logged or normalized per writer.

Inspect the feature weights, e.g. in logistic regression the weight indicates how associated the lexicon feature is with each class.

Training: for each word, average its contextual embeddings (e.g. BERT, no fine-tuning) and fit three regression models predicting its NRC VAD scores. Inference: embed an entity mention in context, apply the regressors, and average over coreferent mentions.

Sentiment, Agency and Power.

Archetypes: protagonist and antagonist both have high power and agency but differ in sentiment; a love interest may have low power and agency but high sentiment.

A lexicon combining affect with frame semantics: a predicate expresses connotations about its arguments, e.g. sentiment of writer and reader toward each role, and of roles toward each other.

Effect (something bad happened to x), value (x is valuable), mental state (x is distressed), power differential between arguments, and agency of the agent.

He implored the tribunal: agent has less power than the theme (power(AG<TH)); demanded implies the reverse. The princess waited for her prince: low agency for the agent (agency(AG)=−).

By hand (crowdsourced, as in Sap et al.) or learned by supervised classifiers for each relation (e.g. whether S(writer→Role1) is + or −), with global constraints across relations.

Osgood et al. (1957): participants rated words on scales, and factor analysis found the most significant factor was the evaluative dimension (good/bad, valuable/worthless, pleasant/unpleasant). This influenced content-analysis dictionaries like the General Inquirer.

Das and Chen (2001), measuring market sentiment from stock message board words, with a hand-made lexicon weighted by how well words discriminate classes like buy vs sell.

The power of using all the words as features for sentiment classification, without a sentiment lexicon.

An algorithm (Rothe et al. 2016) that learns to transform an embedding space so that sentiment (or another property) is concentrated in a few dimensions.