Pushing on Personality Detection from Verbal Behavior: A Transformer Meets Text Contours of Psycholinguistic Features

Evaluation and psychometric validity2022ACL AnthologyApproved editorial review

Authors: Elma Kerz, Yu Qiao, Sourabh Zanwar, Daniel Wiechmann

Keywords: personality detection, psycholinguistic features, BERT transformers, text contours, Big Five, MBTI, verbal behavior analysis

Source: Open primary source (opens in a new tab)

4
Authors
16
Findings
33
Limitations
18
Evidence

Editorial summary

English

This WASSA 2022 paper studies human-personality prediction from text; it does not assess an LLM's personality or show that BERT has traits of its own. It uses two benchmarks. Essays contains 2,468 student essays averaging 672 words with five binary labels derived from Big Five self-report. Kaggle MBTI aggregates PersonalityCafe posts from 8,675 users, averaging 1,288 words per user, with four self-reported MBTI dichotomies. The system extracts what it calls 437 sentence-level psycholinguistic features and retains their sequence or 'text contour': morphosyntactic complexity, lexical richness, readability, and sentiment/emotion/affect signals. The four reported groups actually sum to 436 (19+77+14+326), and the equations likewise use 436-dimensional vectors. The paper compares BLSTMs with and without attention, late-fusion combinations with `bert-base-uncased`, frozen versus fine-tuned BERT, and a stacked ensemble of ten instances. BERT sees only the first 512 tokens, while the contours represent document sentences. The paper reports grid search and ten repetitions of 10-fold cross-validation, but it does not document folds, seeds, the search grid, checkpoint selection, or an outer/nested evaluation for the stacking stage. The main table reports accuracy only. The ensemble averages 63.50% on Essays and 86.51% on MBTI. On Essays, this is 2.90 percentage points above the best displayed prior average of 60.6%. On MBTI, 86.51 minus 77.1 equals 9.41 points, not the 8.28% repeated in the abstract and results; 8.28 points corresponds to the best single fine-tuned model, 85.38%, relative to 77.1. Without BERT, ATTN-PSYLING reaches 60.04% and 75.29%. SP-LIME ranks sentiment/emotion/affect first for all nine targets and lexical features second except for P/J. This is a post-hoc explanation created by zeroing feature groups, not a retraining ablation or causal evidence that the cues express personality. Accuracy is particularly hard to interpret for MBTI because the paper's own figure shows class imbalance and it reports no balanced accuracy, macro-F1, intervals, or variability. It also does not control whether forum texts mention MBTI labels or associated vocabulary. There is no cross-domain validation, demographic analysis, bias audit, or consequence evaluation. Neither ACL nor the paper links code, and targeted searches found no implementation. Without extracted features, predictions, folds, and complete configurations, the results cannot be reproduced end to end. The defensible conclusion is narrow: on these two datasets and under a protocol that is not fully auditable, combining BERT representations with sequences of psycholinguistic measurements yields higher accuracy than the cited baselines; it does not establish a general, fair, or causal measurement of personality.

Español

Este artículo de WASSA 2022 estudia la predicción de personalidad humana a partir de textos; no evalúa la personalidad de un LLM ni demuestra que BERT tenga rasgos propios. Usa dos benchmarks. Essays contiene 2.468 ensayos de estudiantes, de unas 672 palabras, con cinco etiquetas binarias derivadas de autoinforme Big Five. Kaggle MBTI agrega publicaciones de PersonalityCafe de 8.675 usuarios, unas 1.288 palabras por usuario, con cuatro dicotomías MBTI autodeclaradas. El sistema extrae por frase lo que denomina 437 características psicolingüísticas y conserva su secuencia o «contorno textual»: complejidad morfosintáctica, riqueza léxica, legibilidad y señales de sentimiento, emoción o afecto. Sin embargo, los cuatro grupos declarados suman 436 (19+77+14+326) y las ecuaciones también usan vectores de 436 dimensiones. Compara BLSTM con y sin atención, combinaciones de esas redes con `bert-base-uncased` mediante fusión tardía, BERT congelado frente a ajuste fino y un ensemble apilado de diez instancias. BERT solo recibe los primeros 512 tokens, mientras que los contornos representan las frases del documento. El artículo declara grid search y 10-fold cross-validation repetida diez veces, pero no documenta particiones, seeds, rejilla, selección de checkpoint ni una evaluación externa o anidada de la segunda etapa del stacking. La tabla principal informa solo exactitud. El ensemble obtiene 63,50% de promedio en Essays y 86,51% en MBTI. En Essays, la mejora frente al mejor promedio previo mostrado, 60,6%, es de 2,90 puntos porcentuales. En MBTI, 86,51 menos 77,1 equivale a 9,41 puntos, no al 8,28% repetido en el abstract y en resultados; 8,28 puntos corresponde a la mejor red individual ajustada, 85,38%, frente a 77,1. En los modelos sin BERT, ATTN-PSYLING alcanza 60,04% y 75,29%. El análisis SP-LIME pone en primer lugar el grupo de sentimiento/emoción/afecto en los nueve objetivos y el grupo léxico en segundo lugar salvo P/J. Esto es una explicación post hoc obtenida poniendo grupos a cero, no una ablación por reentrenamiento ni evidencia causal de que esos rasgos expresen personalidad. La exactitud en MBTI es especialmente difícil de interpretar porque la propia figura muestra clases desbalanceadas y no se publican balanced accuracy, macro-F1, intervalos ni variabilidad. Tampoco se controla que los textos del foro mencionen etiquetas o vocabulario MBTI. No hay validación en otro dominio, análisis demográfico, auditoría de sesgo o evaluación de consecuencias. ACL y el PDF no enlazan código; las búsquedas dirigidas no localizaron una implementación. Sin features, predicciones, folds y configuración completa, los resultados no se pueden reproducir de extremo a extremo. La conclusión defendible es limitada: en estos dos datasets y bajo este protocolo no plenamente auditable, combinar representaciones BERT con secuencias de medidas psicolingüísticas produce exactitudes mayores que las baselines citadas; no establece una medición general, justa o causal de la personalidad.

Research question

Does textual classification of Big Five and MBTI labels improve by representing the phrase-by-phrase distribution of psycholinguistic features and combining those contours with BERT, and which feature groups contribute to the predictions?

Method

Psycholinguistic measures are extracted per phrase using Stanford CoreNLP, lexicons, lexical norms, and readability measures; the standardized sequence feeds a BLSTM with or without attention. In the hybrid models, that representation is concatenated with the average of embeddings from `bert-base-uncased` over the first 512 tokens, with BERT frozen or fine-tuned. An MLP classifier produces each binary label. The article declares grid search, 10-fold CV repeated ten times, and a stacking of ten predictions using logistic regression. SP-LIME zeroes out four feature groups to approximate their global importance.

Sample: 2,468 student essays, with a declared mean of 672 words and about 1.6 million words in total; and texts from 8,675 PersonalityCafe users, with a mean of 1,288 words and about 11.2 million words. Each document or user contributes five Big Five binary targets or four MBTI binary targets, respectively. The fold identifiers and the effective number of examples after preprocessing are not published.

Findings

  • The BERT+PSYLING ensemble obtains 63.50% mean accuracy on Essays and 86.51% on Kaggle MBTI.
  • On Essays, the ensemble achieves O=71.95, C=61.38, E=63.01, A=60.16, and N=60.98.
  • On MBTI, the ensemble achieves I/E=85.47, N/S=92.27, T/F=85.70, and P/J=82.58.
  • The improvement on Essays over the best previous mean shown, 60.6, is 2.90 percentage points.
  • The improvement of the MBTI ensemble over the best previous mean shown, 77.1, is 9.41 points; the declared 8.28 corresponds to 85.38 minus 77.1, that is, to the best individual model with attention and fine-tuning, not to the ensemble of 86.51.
  • ATTN-PSYLING, without BERT, averages 60.04 on Essays and 75.29 on MBTI; BLSTM-PSYLING obtains 58.68 and 74.38.
  • Hybrid models with fine-tuned BERT outperform their frozen BERT variants on all nine targets in the table.
  • The attention mechanism improves all corresponding BLSTM comparisons except MBTI N/S in the purely psycholinguistic variant.
  • The sentiment/emotion/affect group receives the highest SP-LIME importance across all five Big Five targets and all four MBTI targets.
  • Lexical features rank second across all targets except P/J, where the readability group ranks second.
  • The four groups declared as 19, 77, 14, and 326 features sum to 436, not 437; the encoder equations use 436 dimensions.
  • BERT processes only the first 512 tokens of documents averaging 672 and 1,288 words, so the Transformer representation truncates an unquantified portion of the text.
  • Figure 3 shows a strong imbalance in several MBTI dichotomies; aggregate accuracy may be dominated by the majority class.
  • Descriptive feature differences associate higher extraversion with lower morphosyntactic complexity and more positive vocabulary; higher neuroticism with self-reference and sadness or anxiety vocabulary; and higher conscientiousness with more prevalent words, affiliation, and academic n-grams.
  • These associations are computed as standardized mean differences between labels and do not control for covariates, domain, or causality.
  • The article does not publish code, features, predictions, folds, or per-repetition results; no repository was located through targeted searches by title and internal model names.

Limitations

  • The work calls the feature set 437, but the four categories sum to 436 and the architecture defines 436-dimensional vectors.
  • The complete inventory of the 436/437 variables is not published, nor is an executable version of the ATA system.
  • Some sources, such as LIWC, may require licensed resources; no reproducible path is documented to obtain exactly the same features.
  • There is no linked repository, code, environment files, checkpoints, extracted features, predictions, or evaluation artifacts.
  • No seeds, PyTorch, Transformers, CoreNLP, or lexical resource versions, hardware, or determinism are reported.
  • Batch size, number of epochs, stopping criterion, scheduler, or checkpoint selection are not reported.
  • The hyperparameter grid and the criterion used to choose the optimal configuration are not published.
  • It is not clarified whether grid search is nested within each fold; selecting with the same data used to estimate performance can introduce optimism.
  • The second stacking stage is described with out-of-fold predictions, but no external evaluation or independent nested CV for the meta-model is specified.
  • Fold identifiers are not published, nor is it demonstrated that all normalization and selection steps are fit exclusively on the train fold.
  • The table only reports point accuracy: there are no standard deviations, confidence intervals, distributions across the ten repetitions, or paired tests.
  • Results are compared with numbers from other articles without guaranteeing the same partitions, preprocessing, or validation protocol.
  • The abstract uses percentages where the table allows calculating percentage points; it does not distinguish absolute from relative improvement.
  • The 8.28% for MBTI is incompatible with the ensemble row that the text calls the best model: 86.51 versus 77.1 is 9.41 points.
  • Kaggle MBTI is imbalanced, but balanced accuracy, macro-F1, sensitivity, specificity, confusion matrix, or majority baseline are not reported.
  • MBTI labels are self-reported on a thematic forum and are not validated with supervised administration of the instrument.
  • It is not analyzed whether posts contain type names, dichotomies, or MBTI jargon that could allow direct or indirect label leakage.
  • The article treats MBTI alongside Big Five as one of the two dominant models without discussing their psychometric differences or the information loss from binarizing traits.
  • The five Essays labels are used as binary; measurement error, reliability, or sensitivity to the binarization threshold are not studied.
  • The two datasets come from very specific domains and populations; there is no external validation on another corpus, language, textual genre, or application context.
  • Sufficient demographic composition is not reported to evaluate representativeness or performance by groups.
  • BERT receives only the first 512 tokens and the article does not quantify how many documents are truncated or compare segmentation strategies.
  • The text calls BPE tokens "word tokens"; the relationship between 510 subwords and the declared length in words is not quantified.
  • The fusion combines a truncated view of BERT with potentially complete contours, but does not isolate whether the improvement comes from greater coverage, more parameters, or complementary information.
  • SP-LIME zeroes out groups without retraining the model; describing it as ablation may suggest a stronger intervention than was performed.
  • Zero is the center of a standardized variable, but the paper does not verify that the perturbed combinations form plausible texts or feature profiles.
  • SP-LIME magnitudes are aggregated local explanations of the predictor and do not validate that the groups are causal psychological constructs.
  • The high/low difference graphs do not control for length, topic, age, gender, region, or other variables that may explain the associations.
  • There is no error analysis, examples of false positives/negatives, or qualitative evaluation of model decisions.
  • Calibration, robustness to paraphrases, adversaries, temporal shifts, or domain shift are not evaluated.
  • There is no limitations section, ethics statement, privacy, consent, impact, or potential misuse of inferring personality.
  • Biases or disparities by groups are not audited, even though the work itself mentions sensitive applications such as health, criminal justice, and news.
  • The conclusion proposes studying demographic metadata in PANDORA, but does not evaluate whether using it would be valid, fair, or privacy-respecting.

What the study does not establish

  • It does not demonstrate that BERT, BLSTM, or the ensemble possess personality of their own.
  • It does not demonstrate that human personality can be validly inferred in general from any text.
  • It does not validate MBTI or the binary Big Five labels as error-free ground truth.
  • It does not demonstrate that the improvement generalizes beyond Essays and PersonalityCafe.
  • It does not establish performance in languages other than English.
  • It does not demonstrate that comparisons with prior work are statistically significant under identical folds.
  • It does not prove that the ensemble achieves 8.28 points of improvement on MBTI; the table implies 9.41 over the best previous mean shown.
  • It does not allow knowing how much variability exists between initializations or folds.
  • It does not demonstrate that stacking was evaluated without leakage or selection optimism.
  • It does not establish that sentiment or vocabulary signals cause or directly reveal psychological traits.
  • It does not demonstrate that SP-LIME recovers causal or stable importances.
  • It does not rule out that topic, length, demographics, or forum jargon explain part of the prediction.
  • It does not rule out MBTI label leakage within the textual content.
  • It does not demonstrate fairness across demographic groups or safety in sensitive applications.
  • It does not provide a reproducible end-to-end implementation.
  • It does not justify deploying the system for clinical, employment, educational, judicial, or personalization decisions.

Traceability

Scope: Full text

Version: WASSA 2022 final published version, ACL Anthology 2022.wassa-1.17; DOI 10.18653/v1/2022.wassa-1.17

Consulted source: https://aclanthology.org/2022.wassa-1.17.pdf

Review: Codex full-text, bilingual-fidelity, visual, publication-version, arithmetic-consistency, feature-count, psychometric-label, class-imbalance, truncation, cross-validation, stacking, interpretability, leakage, reproducibility, fairness, ethics and evidence-level audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • BLSTM-PSYLING, 3-layer bidirectional LSTM contour encoder
  • ATTN-PSYLING, 3-layer BLSTM contour encoder with feature-level attention
  • BERT+BLSTM-PSYLING late-fusion model, feature-based/frozen and fine-tuned variants
  • BERT+ATTN-PSYLING late-fusion model, feature-based/frozen and fine-tuned variants
  • BERT+PSYLING stacked ensemble of ten fine-tuned BERT+ATTN-PSYLING instances
  • bert-base-uncased from Hugging Face Transformers
  • Three-layer feed-forward PReLU classifier
  • Logistic-regression stacking model

Instruments and metrics

  • Sentence-level automated text analysis with sliding windows
  • Morphosyntactic complexity features
  • Lexical richness, diversity and sophistication features
  • Readability features
  • Ten sentiment, emotion and affect lexicons including LIWC, NRC, SenticNet and DepecheMood++
  • Stanford CoreNLP tokenization, sentence splitting, POS tagging, lemmatization and PCFG parsing
  • 10-fold cross-validation repeated ten times
  • Classification accuracy
  • SP-LIME group perturbation explanations
  • Stacked generalization with logistic regression

Data used

  • Big Five Essays dataset: 2,468 stream-of-consciousness student essays with five binary self-report labels
  • Kaggle MBTI dataset: PersonalityCafe text from 8,675 users with four self-reported binary MBTI dimensions
  • COCA register subcorpora used to derive n-gram frequency features
  • External lexical norms and sentiment/emotion resources used in the feature extractor

Evidence and location

  • Bibliographic identity, publication, DOI, and complete abstract: ACL Anthology 2022.wassa-1.17; publisher PDF p. 1 (proceedings p. 182)
  • Problem definition and scope of the two benchmarks: Publisher PDF pp. 1–3 (proceedings pp. 182–184), sections 1–3.1
  • Size, origin, and length of Essays and Kaggle MBTI: Publisher PDF p. 3 (proceedings p. 184), section 3.1
  • Four groups and discrepancy 437 versus sum/vector 436: Publisher PDF pp. 3–5 (proceedings pp. 184–186), sections 3.2 and 4.1
  • Linguistic resources and contour construction: Publisher PDF pp. 3–5 (proceedings pp. 184–186), section 3.2 and Figure 1
  • BERT architecture, truncation to 512, and pooling: Publisher PDF p. 5 (proceedings p. 186), section 4.1
  • BLSTM, attention, fusion, and fine-tuning: Publisher PDF pp. 5–6 (proceedings pp. 186–187), sections 4.1–4.2
  • Grid search, repeated 10-fold CV, and absence of selection detail: Publisher PDF pp. 4 and 6 (proceedings pp. 185 and 187), sections 4 and 4.2
  • Two-stage stacking procedure: Publisher PDF p. 6 (proceedings p. 187), section 4.2
  • SP-LIME and zeroing out of groups: Publisher PDF pp. 6–7 (proceedings pp. 187–188), section 4.3
  • Complete results by trait and model: Publisher PDF p. 7 (proceedings p. 188), Table 1
  • Inconsistency 8.28 versus ensemble result: Publisher PDF pp. 1 and 7 (proceedings pp. 182 and 188), abstract, section 5 and Table 1
  • Importance of the four groups: Publisher PDF p. 8 (proceedings p. 189), Table 2 and section 5
  • Visual label imbalance: Publisher PDF p. 11 (proceedings p. 192), Figures 2–3
  • Features descriptively associated with each label: Publisher PDF pp. 8, 12–13 (proceedings pp. 189, 193–194), section 5 and Figures 4–5
  • Conclusions and proposed future generalization: Publisher PDF pp. 8–9 (proceedings pp. 189–190), section 6
  • Visual inspection: All 13 pages of the final ACL PDF rendered and visually inspected, including five figures, two tables and equations; checked 15 Jul 2026
  • Absence of linked implementation: ACL record, final PDF and targeted GitHub searches for exact title, BERT+ATTN-PSYLING and ATTN-PSYLING; checked 15 Jul 2026