From text to talk: Harnessing conversational corpora for humane and diversity-aware language technology

Mark Dingemanse, Andreas LiesenfeldACL 2022 · DOI

Introduction

Modern Natural Language Processing (NLP) is increasingly "divorced" from the natural habitat of human language: informal social interaction. While Large Language Models (LLMs) achieve remarkable fluency by training on massive text scrapes like Wikipedia and web crawls, they suffer from two critical gaps. First, the Interactional Gap—text lacks the timing, turn-taking, and pragmatic interjections (mhm, uh-huh) humans use to coordinate understanding. Second, the Diversity Gap—NLP remains biased toward a handful of high-resource languages, neglecting the conversational richness of the world's 7,000+ languages.

Methodology

To address these gaps, this study analyzed a curated collection of conversational corpora representing 63 languages across 26 language families. Moving beyond standard token-based analysis, the researchers treated language as talk-in-interaction. Their methods included temporal analysis of "Floor Transfer Offsets" (the timing between speakers), activity mapping to distinguish between equitable "chat" and storytelling "chunks," and distributional analysis of entire conversational turns.

Key Insights

Talk is Not Text

A fundamental finding is that the linguistic tools humans use to calibrate understanding in real-time are often absent or rare in written corpora. A comparison of word frequencies reveals that the most essential items in talk are frequently the least common in text.

Scattertext visualization comparing word frequencies in talk versus written text.

The Universal Infrastructure of Turn-Taking

Despite vast cultural and linguistic differences, the study identifies a universal interactional infrastructure. Humans across all 10 sampled language families aim for a "no-gap-no-overlap" target, typically transitioning between speakers in approximately 200ms. This speed implies that humans plan their responses while the other person is still speaking, a level of coordination currently missing from AI systems.

Distribution of turn-taking offsets across diverse languages, showing a universal tendency toward minimal gaps.

Cultural Tuning and Formulaic Language

While the temporal structure of turn-taking is universal, the frequency of feedback is culturally tuned. For example, Korean speakers utilize "continuers" (like eung) twice as often as English speakers use mhm. Furthermore, the researchers found that recurring conversational turns follow a power-law distribution similar to Zipf’s Law for words, suggesting that the "turn" may be a more fundamental unit of social language than the individual word.

Conclusion and Future Work

The study confirms a mean transition time of 59ms across the full dataset, with 46% of turns involving slight overlaps. However, challenges remain regarding data granularity and the exclusion of sign languages. The authors conclude with a call to action for the NLP community: to move from "text to talk" to ensure future conversational AI is not just fluent, but interactionally competent and culturally flexible.

Summary updated

BibTeX
@inproceedings{dingemanse2022from,
  author = {Mark Dingemanse and Andreas Liesenfeld},
  title = {{From text to talk: Harnessing conversational corpora for humane and diversity-aware language technology}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2022},
  doi = {10.18653/v1/2022.acl-long.385},
}
Figure