Empath: Neural-Extended Lexical Categories for Text Analysis
Empath is a 2016 HCI/NLP system for building and applying LIWC-style lexical categories with far less manual dictionary engineering than traditional psycholinguistic lexicons. Its central contribution is a semi-supervised construction pipeline: start with a few seed words, expand them through word-embedding similarity, then use lightweight human filtering to produce interpretable category dictionaries for large-scale text analysis.
Coverage note: verified through May 19, 2026.
Why Empath mattered
Empath sits at the boundary between classic dictionary-based text analysis and modern representation learning. Like LIWC, it treats text as a bag of category-bearing words and reports how much a document uses language associated with themes such as anger, work, money, death, religion, technology, or social media. Unlike LIWC, Empath was designed not only as a fixed dictionary but also as a lexicon generator: users can define a category from a small set of examples, then ask the system to find related terms through an embedding space. arXiv
The useful gloss for Empath is “neural-extended lexical categories”: a category begins as a human-readable lexical concept, but its candidate vocabulary is extended by a neural word embedding trained over a large corpus. In the CHI 2016 paper, Fast, Chen, and Bernstein describe Empath as mining a “living lexicon” from modern web text, using embeddings trained over more than 1.8 billion words of contemporary amateur fiction and then filtering candidates with crowd judgments. arXiv
Empath’s appeal is practical. Traditional lexicons are transparent and cheap to run, but slow to update; supervised classifiers can be more flexible, but require labeled data and are often less interpretable. Empath offered a middle path for Semi-Supervised Lexicon Induction: use a small amount of human input to steer an automatic expansion process, then keep the final artifact as an inspectable dictionary. arXiv+2arXiv+2
The problem: LIWC-style interpretability without LIWC-scale curation
LIWC became influential because it made psychologically meaningful language features easy to compute. Its basic operation is intentionally simple: read a text, match words or stems against dictionary entries, increment category counters, and report category percentages. LIWC-22’s official documentation describes more than 100 built-in dictionaries, a hierarchical category structure, and a dictionary development process involving word collection, human judge ratings, base-rate analyses, and psychometric evaluation. LIWC
That interpretability has a cost. LIWC-style dictionaries require humans to decide which words belong in which categories, how categories should be named, and whether the dictionary tracks contemporary language. The Empath paper explicitly frames LIWC as fast, transparent, and validated, but limited by category coverage: the authors note that LIWC’s topical and emotional categories were relatively small and did not naturally cover many modern or domain-specific topics such as social media, technology slang, or emergent cultural vocabulary. arXiv
Empath’s target was not to replace the theoretical apparatus of LIWC. It targeted the bottleneck of constructing and updating lexical categories. The paper’s benchmark claim is that Empath can generate roughly 200 built-in, human-validated categories and user-defined categories with far less manual effort than a fully hand-authored dictionary. The paper quantifies that reduction in terms of crowdsourcing cost rather than researcher-hours: naively asking crowd workers to validate 200 categories against a 5,000-word vocabulary would cost about $21,000, while Empath’s expansion-first pipeline validated 200 categories with a vocabulary of more than 10,000 words for about $840. arXiv
What Empath produces
Empath has two related outputs:
| Output | What it is | Why it matters |
|---|---|---|
| Built-in lexical categories | About 200 prebuilt thematic and affective dictionaries | Lets analysts run LIWC-like category counts without constructing a lexicon from scratch |
| User-generated categories | New dictionaries expanded from a few seed words | Lets analysts define domain-specific categories such as “social media,” “startup culture,” “bureaucracy,” or “alienation” |
| Document-level scores | Raw or normalized counts of category terms in a text | Makes Empath usable as a feature extractor in pipelines such as Psyche |
| Transparent word lists | Inspectable category vocabularies rather than opaque latent features | Allows auditing, pruning, and comparison across corpora |
The Python documentation presents Empath as a tool for analyzing text across lexical categories “similar to LIWC,” with an API that can return counts for all categories, selected categories, or custom categories. By default, it returns raw counts, but it can normalize by document length, which is usually necessary when comparing texts of different sizes. GitHub
For generated categories, the documentation exposes a create_category operation: for example, a user can seed a category with terms such as “red,” “blue,” and “green,” then let Empath build a broader “colors” category. The documentation also notes multiple category-generation models, including fiction, nytimes, and reddit, which matters because the embedding corpus changes what the model considers semantically or associatively nearby. GitHub
Construction method
Empath’s method can be summarized as a semi-supervised lexicon-construction pipeline. It is “semi-supervised” because humans provide category names, seed terms, and validation judgments, while the embedding model performs the large-scale candidate search.
| Stage | Mechanism | Human role | Machine role |
|---|---|---|---|
| Category proposal | Choose a category name and 2–5 seed terms | Define the intended concept | None |
| Embedding training | Train a word-vector model over a large corpus | Choose or curate corpus | Learn distributional representations |
| Candidate expansion | Search nearest neighbors to the category vector | Optionally inspect candidates | Rank related words by cosine similarity |
| Crowdsourced filtering | Ask workers whether candidate words fit the category | Judge semantic relatedness | Aggregate votes |
| Text analysis | Count category words in documents | Interpret scores | Produce raw or normalized category frequencies |
1. Category names and seed terms
For the 200 built-in categories, Fast, Chen, and Bernstein generated many category candidates from ConceptNet relations conditioned on common corpus words, then manually refined them to final seed sets. The paper reports that each final category used two to five seed terms. For emotional categories, they bootstrapped from Parrott’s hierarchy of emotions rather than relying only on ConceptNet-derived topic names. arXiv
This design decision is important. A seed set does not merely “name” a category; it defines the direction from which embedding expansion begins. In distributional semantics, “violence” seeded by attack, fight, and kill will not produce exactly the same candidate list as “violence” seeded by abuse, threat, and harm. Empath’s generated categories are therefore operational definitions, not pure psychological constructs.
2. Word embedding over modern text
Empath used a neural vector-space model trained on more than 1.8 billion words from Wattpad fiction. The authors chose amateur fiction because it is rich in emotion, social description, and everyday concepts, which are useful for broad affective and topical categories. The model used a skip-gram architecture, 150 hidden dimensions, a context window of five words, minimum word count of 30, and negative sampling. arXiv
This was a methodological bet. Fictional narrative gives strong coverage for emotion, relationship, violence, social roles, and embodied experience. But it also biases the embedding space toward narrative co-occurrence patterns. The paper’s own limitations section gives examples: a fiction-trained embedding can associate “guns” with “gloveboxes” or other story-world artifacts in ways that may not match the analyst’s intended category boundary. arXiv
3. Expansion by embedding similarity
Empath represents each word in a 150-dimensional vector space and gathers candidate category words by cosine similarity. The query vector is built from the category name and seed terms, then used to retrieve nearby terms from a 40,000-word vocabulary. In the paper’s example, “depressed” is near words such as “lonely,” “confused,” “angry,” and “miserable,” illustrating how the embedding captures affective neighborhoods rather than strict dictionary synonymy. arXiv
This step is the “neural extension” in the system. The model does not know what a category should mean in a philosophical or psychometric sense. It only knows which terms appear in similar contexts. That makes the expansion powerful for discovery and dangerous for construct validity: distributional relatedness can mix synonymy, topical association, narrative setting, stereotype, and corpus artifact.
4. Crowdsourced filtering
After expansion, Empath filters candidate terms through Amazon Mechanical Turk. Workers rate whether candidate words are related to the proposed category; the paper describes tasks of 20 words, three workers per task, a four-point relatedness scale, and a majority-vote inclusion rule. A candidate term is retained if at least two of three workers judge it at least weakly related to the category. arXiv
The paper reports that workers judged 95% of unsupervised candidate words to be related, with an average 96% acceptance rate and unanimous worker agreement on related/unrelated status for 81% of words. Those numbers are best interpreted as evidence that the embedding expansion was usually plausible to crowd workers, not as proof that the categories are psychometrically valid. arXiv
5. Dictionary-style scoring
Once a category dictionary exists, Empath scores text by counting category terms. The documentation shows both raw counts and normalized counts; for example, a category count can be divided by total words to report the proportion of a document associated with that category. GitHub
That makes Empath a Closed-Vocabulary Text Analysis system in spirit. Even though the dictionary was generated through an embedding model, the final analysis still depends on whether a token appears in a fixed category vocabulary. It does not infer sarcasm, negation, metaphor, syntactic scope, discourse role, or the truth conditions of an utterance.
Empirical validation
The Empath paper validates the system in several ways: convergence with LIWC where overlapping categories exist, comparison with other lexical resources, human filtering statistics, and demonstrations on downstream corpora. The most widely cited result is the LIWC correlation benchmark: across overlapping categories, Empath with crowd filtering reached an average Pearson correlation of 0.906 with LIWC. arXiv
LIWC convergence
The validation corpus was a mixed text collection, evenly divided among tweets, StackExchange opinions, movie reviews, hotel reviews, and Project Gutenberg novels. The authors compared Empath and LIWC scores over shared categories using Pearson correlations. They capped Empath categories at 200 words and used a cosine-similarity threshold in the expansion process. arXiv
| Shared category | Empath correlation with LIWC | Empath + crowd correlation with LIWC |
|---|---|---|
| Positive emotion | 0.944 | 0.950 |
| Negative emotion | 0.941 | 0.936 |
| Sadness | 0.890 | 0.907 |
| Anger | 0.889 | 0.894 |
| Achievement | 0.915 | 0.903 |
| Religion | 0.893 | 0.908 |
| Work | 0.859 | 0.820 |
| Home | 0.919 | 0.941 |
| Money | 0.902 | 0.878 |
| Health | 0.866 | 0.898 |
| Sex | 0.928 | 0.935 |
| Death | 0.856 | 0.901 |
| Average | 0.900 | 0.906 |
The results support a narrow claim: for LIWC-overlapping categories, Empath’s scores are highly correlated with LIWC’s scores on that mixed corpus. They do not prove that Empath is equivalent to LIWC, nor that categories without LIWC analogues are valid. The paper itself says high Pearson correlations do not evaluate interpretability, false positives, false negatives, category balance, or category breadth, and that categories beyond LIWC were not tested in the same way. arXiv
Crowdsourcing and “internal consistency”
If “internal consistency” is read in the psychometric sense, Empath’s CHI paper does not report Cronbach-style coefficients for each category. That matters because LIWC documentation explicitly describes psychometric evaluation and internally consistent language dictionaries as part of its development process, while Empath’s paper reports crowd acceptance, worker agreement, and LIWC-convergence metrics instead. LIWC
A fairer description is that Empath reports lexical plausibility consistency rather than psychometric internal consistency. Crowd workers usually agreed that generated words belonged to their categories, and the resulting category counts usually tracked LIWC counts where a LIWC comparison was available. That is useful evidence for a dictionary generator, but it is weaker than construct validation for personality, affect, or clinical inference.
Exemplar coverage
Empath’s built-in categories were broader and more contemporary than the LIWC categories available at the time of the paper. The paper includes example category vocabularies for themes such as social media, war, violence, technology, fear, pain, hipster, and contempt; example terms include modern tokens such as “facebook,” “instagram,” “notification,” “selfie,” and “paypal.” arXiv
The practical coverage claim is not that every category is theoretically clean. It is that Empath can rapidly generate large enough category vocabularies to capture many topical signals that hand-built psycholinguistic dictionaries may omit. The paper reports 200 validated categories with a vocabulary of more than 10,000 words, produced at a much lower crowdsourcing cost than exhaustive category-by-word validation. arXiv
Crowd filtering was useful but not magical
The crowd filter improved the average LIWC correlation from 0.900 to 0.906, but the paper says that improvement was not statistically significant. In some categories, crowd filtering made agreement with LIWC worse: for example, the paper notes that workers removed terms such as “semester” from the work category, which may be reasonable in isolation but can reduce convergence with LIWC’s operational definition. arXiv
This is a key lesson for Crowdsourced Evaluation. Crowds can remove obvious false positives, but they do not automatically recover the “correct” construct. Workers judge everyday semantic relatedness; psychometric categories often depend on convention, instrument history, and downstream predictive behavior.
Empath versus LIWC
Empath and LIWC are often treated as substitutes because both produce category-frequency features. Methodologically, they are different artifacts.
| Dimension | LIWC | Empath |
|---|---|---|
| Primary design goal | Psychologically grounded dictionary analysis | Rapid generation and application of broad lexical categories |
| Construction style | Manual dictionary development, judge ratings, corpus base rates, psychometric evaluation | Seed terms, embedding expansion, crowd filtering |
| Category coverage | Stronger for validated psychological and linguistic constructs | Broader topical coverage, especially modern or domain-specific themes |
| Transparency | High: dictionary-style categories | High after generation: category word lists are inspectable |
| Update path | Requires dictionary revision process | Can generate new categories from seeds and retrain/use alternate embeddings |
| Noise profile | More conservative but still context-limited | Broader but somewhat noisier and corpus-dependent |
| Best use | Established psycholinguistic measurement | Exploratory lexical analysis, domain-specific topics, interpretable feature generation |
| Failure mode | Misses new vocabulary; word-sense ambiguity | Semantic drift, embedding artifacts, loose category boundaries |
LIWC’s official documentation emphasizes a long-running dictionary-development process, hierarchical categories, and psychometric evaluation. Empath’s paper emphasizes scale, customizability, and ease of generating new categories. LIWC+2LIWC+2
The “Empath is broader but slightly noisier” summary is directionally right, with one caveat: noise depends on the category, corpus, and downstream task. Empath’s crowd-filtered correlations with LIWC were high, but the paper explicitly warns that high correlation does not settle false positive, false negative, balance, interpretability, or non-LIWC-category validity questions. arXiv
Empath is also easier to update in a specific engineering sense: the construction process is partly automated and exposed through category generation. The documentation’s support for alternate category-generation models such as fiction, NYTimes, and Reddit makes the update path more modular than a purely hand-authored dictionary. GitHub
Relation to earlier semi-supervised lexicon work
Empath was not the first system to expand lexical categories from small supervision. Its contribution was to combine embedding-driven expansion, crowd validation, and a user-facing dictionary-analysis tool in a way that fit HCI and computational social science workflows.
Turney and Littman’s 2003 work on semantic orientation inferred word polarity from association with positive and negative paradigm words using PMI and latent semantic analysis, reporting 82.8% accuracy on a manually labeled set and higher accuracy when abstaining on ambiguous mild words. That work is an early example of seed-based lexical inference. arXiv
Hatzivassiloglou and McKeown’s 1997 work used linguistic constraints, especially conjunction patterns, to infer adjective polarity. Velikovich and colleagues later explored large-scale graph-propagation methods for deriving polarity lexicons from unlabeled web documents. These approaches share Empath’s core intuition: lexical categories can be induced from structure in unlabeled language, provided a small amount of supervision or a structural assumption anchors the interpretation. ACL Anthology
Hamilton, Clark, Leskovec, and Jurafsky’s 2016 work is especially close in spirit. They induced domain-specific sentiment lexicons by combining word embeddings with label propagation over small seed sets, showing that sentiment-bearing words can change polarity across domains and communities. Empath generalizes that kind of seed-and-embedding logic beyond positive/negative sentiment into many topical and affective categories. arXiv
The broader lesson for Distributional Semantics is that embeddings are not classifiers by themselves. They provide a geometry of usage. Turning that geometry into a lexicon requires naming categories, choosing seeds, selecting a corpus, defining inclusion thresholds, and validating the resulting word lists.
Role in Psyche’s analysis pipeline
In Psyche, Empath functions as the lexical-analysis layer: a local, auditable feature extractor that can sit alongside or substitute for LIWC-style analysis when proprietary LIWC access is unavailable or when broader categories are useful. The Psyche GitHub README describes an analysis pipeline with corpus parsers, quality filters, LLM inference, Empath analysis over 200+ lexical categories mapped to Big Five traits, and weighted synthesis. It also lists Empath linguistic analysis as a LIWC-validated evidence source with weight 0.10. GitHub
That stated role should be interpreted carefully. A 2026 Psyche project writeup says Empath was originally used as a fourth inference method with weight 0.10, but subsequent analysis found that Empath measured topic and genre more than personality; the writeup says Empath is now used for comparative corpus characterization, with z-scored lexical frequencies across contexts. raw.ashitaorbis.com
That is the right caution. Empath can tell a pipeline that one corpus contains more work, death, family, technology, aggression, or social-media vocabulary than another. It cannot, by itself, infer stable personality traits. If Psyche uses Empath features for Personality Inference, those features should be treated as weak lexical evidence, not as direct trait measurements.
A useful division of labor is:
| Pipeline layer | Empath’s role | What it should not be asked to do |
|---|---|---|
| Corpus characterization | Compare topic and affect vocabulary across documents, contexts, or time windows | Infer personality from vocabulary alone |
| Feature extraction | Provide interpretable lexical features for models or reports | Replace construct validation |
| Sanity checking | Detect whether an LLM-based interpretation is tracking obvious lexical signals | Decide whether an interpretation is psychologically true |
| Privacy-sensitive local analysis | Run cheap, local dictionary counts without sending text to an API | Understand context like a language model |
In Psyche-like systems, Empath is most defensible as a transparent evidence channel. It is less defensible as a standalone psychometric instrument.
Limitations
1. Empath is still closed-vocabulary in spirit
Empath’s generation phase is neural, but its analysis phase is dictionary matching. Once a category is built, any text that expresses the concept without using category words is invisible. Conversely, any text that uses a category word in an irrelevant sense may be counted. This is the same broad failure mode that affects LIWC, though the specific dictionary contents differ. LIWC’s own documentation notes word-sense ambiguity problems, such as a word like “mad” appearing in different psychological contexts. LIWC
This limitation is not a minor implementation detail. It defines the epistemology of the method. Empath measures lexical surface evidence, not latent mental state.
2. Embedding quality determines category quality
The paper’s “garbage in, garbage out” warning is central. Empath’s categories inherit the biases, genres, time period, and word senses of the embedding corpus. The authors report that a Google News embedding version performed worse overall than their fiction-trained embedding, with average LIWC correlation dropping to 0.84 and especially poor results for categories such as religion and work. arXiv
This does not mean fiction is always the best corpus. It means the embedding corpus must match the analysis domain. A fashion forum, Hacker News archive, clinical interview corpus, legal corpus, or fanfiction corpus will encode different neighborhoods. The paper itself suggests that different corpora may be appropriate for different generated categories. arXiv
3. Seed choice is construct choice
Empath’s generated categories depend on seed words. The paper explicitly notes that seed choice is important. A category seeded by polite institutional language will differ from one seeded by slang, taboo terms, or clinical vocabulary. arXiv
This is especially dangerous for philosophical or psychological categories such as “agency,” “alienation,” “neuroticism,” “care,” or “moral injury.” These constructs have competing theories. A seed set can smuggle one theory into the lexicon while giving the output an appearance of empirical neutrality.
4. Semantic drift and associative drift
Embedding expansion retrieves related words, not necessarily category members. Relatedness can mean synonymy, antonymy, scene co-occurrence, social stereotype, genre convention, or causal association. A “violence” category may pull in weapons, police, fear, and injury terms; depending on the downstream question, those may be valid indicators or contaminating context.
This is why crowdsourced filtering helps but does not solve the problem. Crowd workers can reject obvious mismatches, but they usually cannot adjudicate whether a candidate term belongs in a construct for a specific research design.
5. Category multiplicity invites false positives
Empath’s breadth can tempt analysts to run hundreds of categories and then interpret whichever correlations look interesting. The paper warns against this kind of uncontrolled search, noting the risk of false positives when many categories are tested and recommending corrections such as Bonferroni, Holm, false-discovery-rate methods, or regularization approaches like LASSO. arXiv
This is a serious concern for blog-scale and self-quantification projects. A 200-category lexicon can produce a compelling story from noise unless the analyst preregisters hypotheses, holds out data, or treats the output as exploratory.
6. Empath validates overlap better than novelty
Empath’s strongest quantitative validation is against LIWC on overlapping categories. That is useful, but it leaves a gap: the categories that make Empath most interesting are often the ones LIWC does not already cover. The paper is explicit that categories beyond LIWC were not tested in the same way. arXiv
That does not invalidate novel Empath categories. It means they require local validation. A generated “AI anxiety” category for 2026 engineering blogs, for example, should be inspected, tested against human labels, and checked for genre confounds before being used as evidence.
Has LLM classification superseded Empath?
For many inference tasks, especially personality and affect inference, LLM-Based Text Classification is now a stronger default candidate than dictionary counting. Large language models can use context, paraphrase, discourse, and pragmatic cues that lexicons miss. Recent work reports that GPT-4-based systems can infer Big Five personality from free-form user interactions, with the strongest performance when the model is prompted to elicit personality-relevant information; Peters, Cerf, and Matz report mean correlation around r = .443 in that elicitation setting, while naturalistic and generic-assistant settings perform lower. arXiv
Other work shows LLMs can infer psychological dispositions from social media text in zero-shot settings. Peters and Matz report average Big Five correlation around r = .29 from Facebook status updates using GPT-3.5 and GPT-4, comparable to supervised machine-learning baselines in their study. arXiv
But “LLMs supersede Empath” is too broad. The better claim is task-dependent:
| Task | Empath | LLM classifier |
|---|---|---|
| Cheap corpus-wide lexical trend analysis | Strong | Often unnecessary and expensive |
| Transparent feature extraction | Strong | Weaker unless explanations are separately audited |
| Local privacy-preserving analysis | Strong | Depends on local model availability |
| Context-sensitive affect inference | Weak | Stronger candidate |
| Personality inference | Weak as standalone evidence | Stronger, but still validation-sensitive |
| New category exploration | Strong for interpretable seed expansion | Strong for open-ended labeling, weaker for stable measurement |
| Production classification with labels | Useful as features | Fine-tuned models often stronger |
The current evidence does not show that zero-shot generative LLMs dominate all text classification settings. Bucher and Martini report that fine-tuned smaller BERT-style models can outperform larger zero-shot generative models across several classification tasks, while broader reviews note continuing tradeoffs among accuracy, labeled-data requirements, cost, inference latency, and language coverage. arXiv
For personality and affect inference, LLMs reduce some of Empath’s core weaknesses: they can interpret context, combine evidence across passages, and reason over statements that do not contain obvious category words. They introduce different weaknesses: opacity, prompt sensitivity, training-data uncertainty, potential demographic bias, privacy exposure, and difficulty reproducing results across model versions.
The open question is therefore not whether Empath has been “beaten” in a leaderboard sense. It is whether a given analysis needs measurement transparency or semantic competence. Empath is still useful when the analyst wants auditable lexical features. LLMs are more appropriate when the construct cannot be reduced to category-word frequencies.
Practical interpretation rules
Empath results should be read as lexical signals, not psychological facts. A high “anger” score means the text contains more words in Empath’s anger dictionary. It does not necessarily mean the author is angry. A high “technology” score means the document uses technology-associated vocabulary. It does not necessarily mean the document is about technology in the analyst’s intended sense.
The safest use pattern is comparative and corpus-aware:
| Safer use | Riskier use |
|---|---|
| “This corpus uses more family vocabulary than that corpus.” | “This author is more family-oriented.” |
| “These posts have more death-related lexical markers after the event.” | “The author is clinically depressed.” |
| “The LLM’s interpretation may be influenced by high work/money vocabulary.” | “Empath confirms the LLM’s personality diagnosis.” |
| “This generated category needs inspection and validation.” | “The generated category is valid because the embedding found related words.” |
Empath is best understood as an interpretable measurement scaffold. It helps analysts ask better questions about language, but it does not answer the psychological or philosophical question by itself.
References
[Fast, Chen & Bernstein 2016 CHI] Ethan Fast, Binbin Chen, and Michael S. Bernstein, “Empath: Understanding Topic Signals in Large-Scale Text,” CHI 2016. Primary source for Empath’s method, built-in categories, LIWC correlation benchmark, crowd filtering, and limitations. HCI Group
[Empath Python documentation] Official Empath package documentation and examples for category analysis, normalized counts, and generated categories. GitHub+2GitHub+2
[LIWC official documentation] Official LIWC documentation describing dictionary matching, LIWC-22 categories, dictionary construction, and psychometric evaluation. LIWC
[Turney & Littman 2003] Peter Turney and Michael Littman, “Measuring Praise and Criticism: Inference of Semantic Orientation from Association.” Early seed-word method for inferring semantic orientation. arXiv
[Hatzivassiloglou & McKeown 1997] Vasileios Hatzivassiloglou and Kathleen McKeown, “Predicting the Semantic Orientation of Adjectives.” Early work on inducing lexical polarity from linguistic constraints. ACL Anthology
[Velikovich et al. 2010] Leonid Velikovich and colleagues, “The Viability of Web-Derived Polarity Lexicons.” Large-scale graph-propagation approach to deriving polarity lexicons from unlabeled web text. research.google.com
[Hamilton et al. 2016] William Hamilton, Kevin Clark, Jure Leskovec, and Dan Jurafsky, “Inducing Domain-Specific Sentiment Lexicons from Unlabeled Corpora.” Embedding-plus-label-propagation method showing that lexical polarity varies across domains and communities. arXiv
[Psyche GitHub README] Project documentation for Psyche’s pipeline, including Empath analysis as a local lexical-analysis component. GitHub
[Psyche project writeup 2026] Project writeup describing Psyche’s shift from using Empath as a personality-inference channel toward using it for comparative corpus characterization. raw.ashitaorbis.com
[Peters, Cerf & Matz 2024] Research on LLM-based personality inference from free-form user interactions. arXiv
[Peters & Matz 2023/2024] Research on LLM inference of psychological dispositions from social media text. arXiv
[Bucher & Martini 2024] Evidence that fine-tuned smaller models can outperform zero-shot generative LLMs on multiple text-classification tasks. arXiv
Companion entries
Core theory: LIWC, Closed-Vocabulary Text Analysis, Distributional Semantics, Word Embeddings, Semi-Supervised Lexicon Induction, Construct Validity, Psychometrics for Text Analysis
Methods: Crowdsourced Evaluation, Dictionary-Based NLP, Embedding Similarity, Label Propagation, Feature Extraction for Personality Models, False Discovery Rate in Text Mining
Practice: Psyche, Personality Inference, Corpus Characterization, Lexical Analysis Pipeline, Interpretable NLP Features, Local-First Text Analysis
Counterarguments and limitations: LLM-Based Text Classification, Semantic Drift, Prompt-Based Psychological Inference, Measurement Error in Computational Social Science, Genre Confounds, Open-Vocabulary Models