NLP for Psychology Text Data

Why This Matters

Psychology has always been a measurement-bound science. Self-report scales are noisy and reactive; clinician ratings are scarce and expensive. Naturally occurring text (diary entries, social media posts, therapy transcripts, electronic health record notes) gives a high-frequency behavioral signal that does not require a study visit. NLP turns that signal into variables a psychologist can model.

The applications split into two camps. One camp uses text to measure a construct (personality, affect, suicidal ideation) so it can be entered into a downstream regression. The other uses text to predict an outcome (a future depression diagnosis, a treatment response) without claiming the model captures a latent trait. The methods overlap; the validity arguments do not.

Core Methods

The oldest tool is the dictionary count. Linguistic Inquiry and Word Count (LIWC; Pennebaker, Boyd, Jordan, Blackburn 2015) maps words to roughly 90 categories — first-person singular pronouns, negative emotion, cognitive process words, social references — and reports each as a percentage of total tokens. Dictionary methods are transparent, fast, and replicable across corpora, but they ignore syntax, negation, and context. "I am not sad" and "I am sad" produce identical negative-emotion scores. They remain the default for hypothesis-driven psychology because the units are interpretable and the scoring code does not change between studies.

The second wave is supervised classification on top of pretrained encoders. A BERT-style model fine-tuned on labeled posts (PHQ-9 scores, clinician annotations, self-disclosed diagnoses) outperforms LIWC on prediction metrics like AUC and F1, often by 10 to 20 points. Eichstaedt et al. 2018 (PNAS) trained a logistic model on Facebook language and predicted depression diagnoses recorded in linked EHR data with an AUC near 0.72, outperforming demographics-only baselines. The features that drove the model — first-person singular pronouns, hostility, loneliness words, references to medical care — were consistent with prior small-sample findings in clinical psychology, which raised the credibility of the result.

The third wave uses dense word embeddings and contextual representations as input to downstream models. An EHR clinical note can be encoded with a domain-tuned BERT variant, pooled, and fed into a risk model. Sentiment analysis on therapist-patient dialogue uses the same machinery. The cost is interpretability: a high coefficient on dimension 137 of a 768-dim hidden state does not translate into a clinical narrative.

A fourth, unsettled use is generative chat. Tools that present themselves as therapy assistants are not regulated medical devices in most jurisdictions, do not retain liability for harm, and have no validated escalation pathway for suicidal users. Mental health professional bodies have flagged risk; serious empirical evaluation on safety-critical interactions is sparse.

Watch Out

Predicting a diagnosis is not measuring depression

Eichstaedt et al. 2018 predicted whether a person had a depression code in their EHR. That is not the same as measuring whether the person was depressed. EHR codes encode who showed up, who got billed, and who had insurance. A model that uses Facebook language to predict EHR depression codes inherits all of those selection effects. It is an honest predictor of a label, not a thermometer for mood.

Watch Out

LIWC validity is category-specific, not blanket

LIWC's negative-emotion category has been validated in many corpora; its cognitive-process and analytic-thinking categories have weaker, more contested validity. Citing "LIWC has been validated" without naming the category is a common error. Pennebaker's manual lists the per-category validation evidence; agents should cite the specific category they used.

References

Related Topics

• BERT and the pretrain-finetune paradigm
• Word embeddings
• Transformer architecture
• NLP for economic text analysis
• Reproducibility and experimental rigor