Measuring the Language of Self-Disclosure across Corpora

Findings of the Association for Computational Linguistics: ACL 2022, pages 1035 - 1047 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Measuring the Language of Self-Disclosure across Corpora Ann-Katrin Reuel1, Sebastian Peralta1, João Sedoc2, Garrick Sherman1, Lyle Ungar1 1University of Pennsylvania, SEAS, 220 S 33rd St, Philadelphia, PA 19104 2New York University Stern, 44 West 4th Street, New York, NY 10012 {akreuel, peraltas, garricks, ungar}@seas.upenn.edu, jsedoc@stern.nyu.edu Abstract Being able to reliably estimate self-disclosure – a key component of friendship and intimacy – from language is important for many psy- chology studies. We build single-task models on five self-disclosure corpora, but find that these models generalize poorly; the within- domain accuracy of predicted message-level self-disclosure of the best-performing single- task model (mean Pearson’s r=0.69) is much higher than the respective across data set accu- racy (mean Pearson’s r=0.32), due to both vari- ations in the corpora (e.g., medical vs. general topics) and labelling instructions (target vari- ables: self-disclosure, emotional disclosure, intimacy). However, some lexical features, such as expression of negative emotions and use of first person personal pronouns such as ’I’ reliably predict self-disclosure across cor- pora. We develop a multi-task model that im- proves results, with an average Pearson’s r of 0.37 for out-of-corpora prediction. 1 Introduction Interpersonal exchanges are a core component in human relationships. They are determined by inti- macy, which in turn is characterized by the willing- ness of the involved parties to self-disclose (Rubin and Shenker, 1978). In general, self-disclosure can be defined as “revealing intimate information about one’s self” (Derlega et al., 1993). Note that self- disclosure, which often involves revealing embar- rassing facts about oneself that are considered vio- lations of social norms ("I flunked my exam." or "I have a growth on my butt"), is different from reveal- ing personally identifiable information (PII). Self- disclosure encompasses the sharing of thoughts, aspirations, feelings, likes and dislikes, while PII, such as date of birth or social security number, is used to unambiguously identify a person. Unlike self-disclosing, sharing PII does not necessarily suggest an intimate relationship between two peo- ple. Figure 1: Two sentences from the Med and the Emp- Con data set. In our work, we predict the associated self-disclosure and assess the most important features in both sentences, which are highlighted in the exam- ples. NLP researchers have labeled a variety of data sets with self-disclosure or some approximation of self-disclosure such as "intimacy", which is more accurately viewed as being a property of the re- lationship between two people than of an utter- ance. In this paper, we build models to predict self-disclosure from text, and assess how well these models generalize across five different corpora. We find that they mostly generalize poorly, but that there are some reliable linguistic markers of self- disclosure. We draw on multiple corpora labeled for self- disclosure: conversations from an online breast cancer support community (Wang et al., 2015); annotated conversational turns (Omitaomu et al., 2022); medical posts from patient.info and Red- dit (Valizadeh et al., …

Content truncated to 500 words