Using Topic Modeling to Improve Prediction of Neuroticism and Depression in College Students

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1348–1353, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics Using Topic Modeling to Improve Prediction of Neuroticism and Depression in College Students Philip Resnik University of Maryland College Park, MD 20742 resnik@umd.edu Anderson Garron University of Maryland College Park, MD 20742 agarron@cs.umd.edu Rebecca Resnik Mindwell Psychology Bethesda 5602 Shields Drive Bethesda, MD 20817 drrebeccaresnik@gmail.com Abstract We investigate the value-add of topic model- ing in text analysis for depression, and for neu- roticism as a strongly associated personality measure. Using Pennebaker’s Linguistic In- quiry and Word Count (LIWC) lexicon to pro- vide baseline features, we show that straight- forward topic modeling using Latent Dirich- let Allocation (LDA) yields interpretable, psy- chologically relevant “themes” that add value in prediction of clinical assessments. 1 Introduction In the United States, where 25 million adults per year suffer a major depressive episode (NAMI, 2013), identifying people with mental health prob- lems is a key challenge. For clinical psychologists, language plays a central role in diagnosis: many clinical instruments fundamentally rely on what is, in effect, manual coding of patient language. Au- tomating language assessment in this domain poten- tially has enormous impact, for two reasons. First, conventional clinical assessments for affective dis- orders (e.g. The Minnesota Multiphasic Personal- ity Inventory, MMPI) are based on norm-referenced self-report, and therefore depend on patients’ will- ingness and ability to report symptoms. However, some individuals are motivated to underreport symp- toms to avoid negative consequences (e.g. active duty soldiers, parents undergoing child custody eval- uations), and others lack the self awareness to report accurately.1 Second, many people – e.g. those with- 1As evidenced by the fact that assessments such as the MMPI-2- RF include validity scales to detect, e.g., defensiveness, atyp- out adequate insurance or in rural areas – cannot ac- cess a clinician qualified to perform a psychological evaluation (Sibelius, 2013; APA, 2013). There is enormous value in inexpensive screening measures that could be administered in primary care and by social workers and other providers. We take as a starting point the well known lexicon-driven methods of Pennebaker and col- leagues (LIWC, Pennebaker and King (1999)), which relate language use to psychological vari- ables, and improve on them straightforwardly using topic modeling (LDA, Blei et al. (2003)). First, we show that taking automatically derived topics into account improves prediction of neuroticism (emo- tional instability, John and Srivastava (1999)), as measured by correlation with widely used clinical instruments, when compared with lexically-based prediction alone. Neuroticism is of particular inter- est as a personality measure because higher scores on neuroticism scales are consistent with increased distress and more difficulty coping; individuals with high levels of neuroticism may also be at higher risk of psychiatric problems categorized as Axis I in the DSM-IV (Association, 2000), including the inter- nalizing disorders (depression, anxiety).2 Second, we show a similar correlation improvement result for prediction of depression, adding improvement ical responses, and overly positive self-portrayals (Tellegen et al., …

Content truncated to 500 words