Decoding Affect in Dyadic Conversations: Leveraging Semantic Similarity through Sentence Embedding

Chen-Wei Yu, Yun-Shiuan Chuang, Alexandros N. Lotsos, Claudia M. HaasearXiv 2023 · arXiv · OpenReview

Abstract

Recent advancements in Natural Language Processing (NLP) have highlighted the potential of sentence embeddings in measuring semantic similarity. Yet, whether this approach can be used to analyze real-world dyadic interactions and predict people's emotional experiences remains largely uncharted. This study analyzes verbal conversations of 50 married couples engaging in naturalistic 10-minute conflict and 10-minute positive conversations. Using the Transformer-based model General Text Embeddings-Large (GTE-Large), we quantified conversation similarity via the average cosine similarity between adjacent utterances. Results indicate that lower similarity is associated with greater positive emotional experiences in pleasant conversations, but not in conflict. These findings remain stable when controlling for marital satisfaction and interaction volume, underscoring the potency of sentence embeddings in understanding the links between interpersonal dynamics and individual affect.

Introduction

In relationship science, the concept of "linguistic synchrony"—the tendency of partners to mirror each other’s language—has long been viewed as a hallmark of a healthy connection. Conventional wisdom suggests that "being on the same page" is always beneficial. However, this study, recently published in the Journal of Language and Social Psychology, challenges that "more is better" assumption. By applying modern transformer-based NLP to transcripts of married couples, the research demonstrates that the emotional value of semantic similarity depends entirely on the context of the interaction: whether a couple is resolving a conflict or sharing a pleasant memory.

Theoretical Framework

The researchers contrasted two psychological frameworks to understand how conversational flow relates to emotion:

  1. The Shared-Mind Perspective: This view predicts that high semantic similarity is universally positive, as it indicates that partners are mutually attentive and focused on a shared topic.
  2. The Broaden-and-Build Theory: This theory suggests that positive emotions expand our cognitive horizons. In a happy, low-stress conversation, this "broadening" should lead to a diverse flow of ideas, memories, and associations. Consequently, a joyful interaction might actually exhibit lower semantic similarity as the conversation wanders through a widened cognitive space.

Methodology

Moving beyond traditional word-counting methods like LIWC (Linguistic Inquiry and Word Count), the study utilized Sentence Embeddings to capture the nuanced meaning of dialogue.

The Model and Metric

The team employed gte-large (General Text Embeddings), a transformer model that converts sentences into 1024-dimensional vectors. Unlike simple keyword matching, these embeddings capture contextual meaning (e.g., the difference between a "river bank" and a "financial bank"). The primary metric, Latent Semantic Similarity (LSS), was calculated by measuring the cosine similarity between the embeddings of adjacent speaking turns.

Validation

To ensure LSS captured genuine interpersonal dynamics rather than random noise, the researchers performed three rigorous checks:

  • Temporal Decay: Confirming that utterances closer in time are more similar than those far apart.
  • Permutation Tests: Shuffling turn orders to demonstrate that the specific sequence of a conversation is meaningful.
  • Pseudo-Dyads: Pairing participants with strangers they did not actually interact with to prove that LSS captures a unique interdependence between real partners.

Experimental Design

The study analyzed a dataset of 57 long-term married couples (average marriage duration of 11 years). Each couple engaged in two 10-minute recorded conversations:

  • Conflict Context: Discussing a recurring topic of disagreement.
  • Pleasant Context: Discussing a topic of mutual enjoyment or a shared positive memory.

Following each session, participants provided a subjective rating of their emotional experience on a scale of 1 (very unpleasant) to 9 (very pleasant).

Key Results

The data provided strong support for the Broaden-and-Build theory in positive contexts while showing distinct patterns during conflict:

  • Pleasant Conversations: There was a significant negative association between LSS and positive emotions ($\beta = -0.30, p < .05$). In these interactions, lower similarity—or "topic-hopping"—was a predictor of higher happiness. This suggests that when couples are happy, they feel free to explore a wide range of topics.
  • Conflict Conversations: There was a positive but non-significant trend ($\beta = 0.19, p = .262$). While staying on-topic may be directionally helpful during a fight, it does not carry the same weight as the cognitive flexibility seen in pleasant exchanges.
  • Robustness: These findings held steady even when controlling for marital satisfaction, total word count, and "Linguistic Style Matching" (the mirroring of function words like "the" or "it"). The results were also algorithm-agnostic, remaining consistent when using SBERT (Sentence-BERT) instead of gte-large.

Discussion and Implications

This research suggests that LSS is a powerful "implicit" measure of relationship health. Unlike specific word choices, which individuals can consciously self-censor, the semantic flow of a conversation emerges spontaneously. The findings highlight that "synchrony" is not a monolithic good; rather, the ability to drift across diverse topics is a hallmark of shared joy and cognitive expansion.

Limitations and Future Directions

While groundbreaking, the study notes several areas for future exploration:

  • Sample Size: With 57 couples, the study may have been underpowered to detect a statistically significant effect in the conflict condition.
  • Directionality: As a correlational study, it remains unclear if a positive mood causes diverse conversation or if diverse conversation creates a positive mood.
  • Generalizability: The dynamics of long-term married couples may differ significantly from those of new daters or parent-child dyads.
  • Granularity: Future research could utilize "rating dials" to capture second-by-second emotional shifts, providing a more dynamic view of how semantic shifts influence affect in real-time.

Summary updated

Abstract

Recent advancements in Natural Language Processing (NLP) have highlighted the potential of sentence embeddings in measuring semantic similarity (hereafter similarity). Yet, whether this approach can be used to analyze real-world dyadic interactions and predict people's emotional experiences in response to these interactions remains largely uncharted. To bridge this gap, the present study analyzes verbal conversations of 50 married couples who engage in naturalistic 10-minute conflict and 10-minute positive conversations. Transformer-based model General Text Embeddings-Large is employed to obtain the embeddings of the utterances from each speaker. The overall similarity of the conversations is then quantified by the average cosine similarity between the embeddings of adjacent utterances. Results show that lower similarity is associated with greater positive emotional experiences in the positive (but not conflict) conversation. Follow-up analyses show that (a) findings remain stable when controlling for marital satisfaction and the number of utterance pairs and (b) the similarity measure is valid in capturing critical features of a dyadic conversation. The present study underscores the potency of sentence embeddings in understanding links between interpersonal dynamics and individual emotional experiences, paving the way for innovative applications of NLP tools in affective and relationship science.
BibTeX
@misc{yu2023decoding,
  author = {Chen-Wei Yu and Yun-Shiuan Chuang and Alexandros N. Lotsos and Claudia M. Haase},
  title = {{Decoding Affect in Dyadic Conversations: Leveraging Semantic Similarity through Sentence Embedding}},
  year = {2023},
  eprint = {2309.12646},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2309.12646},
}

Loading related content...

Figure