Modeling Empathetic Alignment in Conversation

TL;DR

The paper targets the gap in NLP empathy work by explicitly modeling cognitive alignment between a Target and an Observer using Appraisal Theory. It introduces AloE, a span-annotated dataset of 9,284 appraisals and 3,262 target–observer alignments from Reddit, and develops models for appraisal prediction and for alignment via a Siamese architecture on $T\times O$ pairs with labels $Y\in \{0,1\}^{k\times l}$. Large-scale analysis of 2.3M posts and 8.9M comments reveals that observers partially align with targets, often through advice rather than mirroring appraisals, with mental health professionals showing higher alignment than laypeople. The work provides datasets, models, and insights to improve empathetic communication in online support and to inform downstream NLP tasks that assist in generating or highlighting empathetic, aligned responses.

Abstract

Empathy requires perspective-taking: empathetic responses require a person to reason about what another has experienced and communicate that understanding in language. However, most NLP approaches to empathy do not explicitly model this alignment process. Here, we introduce a new approach to recognizing alignment in empathetic speech, grounded in Appraisal Theory. We introduce a new dataset of over 9.2K span-level annotations of different types of appraisals of a person's experience and over 3K empathetic alignments between a speaker's and observer's speech. Through computational experiments, we show that these appraisals and alignments can be accurately recognized. In experiments in over 9.2M Reddit conversations, we find that appraisals capture meaningful groupings of behavior but that most responses have minimal alignment. However, we find that mental health professionals engage with substantially more empathetic alignment.

Figures (14)

• Figure 1: Subreddits arranged according to their distribution of Target appraisals.
• Figure 2: Subreddits arranged according to their distribution of Observer appraisals.
• Figure 3: The probabilities of a Target's appraisal type (row) having the specific appraisal (col) in the aligned span of the Observer's. Empty cells indicate the aligned pairs are rare or non-existent in the data.
• Figure 4: Mean alignment by profession; the error bar for laypeople is too small to be seen.
• Figure 5: Difference of mean alignment between professionals and laypeople by appraisal; bars pointing right are appraisals more commonly aligned by professionals, left for laypeople.