Paper

Global divergence and local convergence of utterance semantic representations in dialogue

---

Abstract

We use deep contextualized embedding models (BERT & ELMo) and shallow word embedding models (Fasttext & GloVe) to study the alignment between dialogue interlocutors at the semantic representation level, with the goal to examine the interactive alignment model (IAM) theory. We have observed both divergence and convergence patterns in dialogue: First, the semantic distance between two adjacent utterances increases with their relative positions within the dialogue, i.e., utterances at the later stage are more semantically apart than the earlier ones. Second, semantic distance also increases with the physical distance between utterances, i.e., utterances that are physically closer have more similar semantic meanings. We conclude that dialogue interlocutors demonstrate global divergence and local convergence patterns in semantic representation space. Our findings resolve the conflicts in previous studies, and challenge the claim from IAM that people gradually build alignment at higher representation levels in dialogue. The feasibility of using semantic representation techniques as psycholinguistic models of dialogue is discussed.

Keywords: linguistic alignment, interactive alignment model, dialogue, semantic models, BERT

How to Cite:

Xu, Y., (2021) “Global divergence and local convergence of utterance semantic representations in dialogue”, Society for Computation in Linguistics 4(1), 116-124. doi: https://doi.org/10.7275/3fgk-0y64