📄 Analyzing the Effect of Entrainment on Dialogue Acts

Summary

Speakers in a conversation tend to converge on each other's behaviour. This convergence is called entrainment. It is well studied for word choice, but little is known about whether it also shapes the structure of a dialogue. The authors ask three questions. Do partners align in the dialogue acts they choose? Does lexical alignment depend on the dialogue act of an utterance? Does alignment grow as a conversation goes on? Working on the Switchboard Corpus, they build on the Nenkova et al. word-distribution score with two changes. First, the raw score is replaced by a rank-based Entrainment Score Ratio, which counts how often the real partner beats every non-partner speaker. Second, per-speaker unigram distributions are Dirichlet-smoothed against a corpus-wide background model, which reduces sparsity effects. The findings are:

• Dialogue-act entrainment exists.
• Lexical entrainment is strongest in low-content acts such as greetings, closings, backchannels and agreement.
• Entrainment rises slightly from the first half of a dialogue to the second, but not significantly.

These results suggest that dialogue systems should adapt to users at the act level and should do so progressively.

Key points

• Data: Switchboard Corpus with dialogue-act labels. The per-act analysis covers N = 2310 target speakers.
• Baseline check on the 25MFC word class gives a score of -0.211 for partners and -0.248 for non-partners. This reproduces the earlier finding that real partners align more.
• Word classes from prior work are 25MFC, 25MFD, ACW (affirmative cue words) and FP (filled pauses). 25MFC is the main class used.
• Conventional-closing shows the strongest lexical entrainment, with Cohen's d 1.50 and Ratio 0.703. Effects are counted as notable at d > 0.5 or Ratio > 0.55. Acts such as Apology, Action-directive and several question types show almost no gain.
• The smoothing weight α is fitted by maximizing likelihood under a Dirichlet process prior with Newton's method.
• The partner versus non-partner difference is already significant in the earlier half of dialogues. A "Cross" condition tests whether one speaker's earlier words reappear later in the partner's speech.

Interesting ideas

• I1. Compare each real pair against many shuffled non-partner pairs and report the win rate, not the raw similarity; this makes measures over sets of very different sizes comparable.
• I2. Smooth small per-slice distributions (per speaker, act and dialogue part) toward a global background model with a fitted Dirichlet weight, so that sparse slices do not look artificially dissimilar.
• I3. Condition lexical measurements on the dialogue act of an utterance, since formulaic acts behave very differently from content-bearing ones.
• I4. The authors admit that their measure cannot separate fixed adjacency responses (coordination) from true entrainment, which flags a confound to design against.
• I5. Split each dialogue into earlier and later halves and add a cross condition to detect delayed influence of one speaker on the other.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (learn): use partner-vs-random win rates as a negative control when testing whether two parsers' agreement or a frame match exceeds chance.
• I2 (tokens): smooth per-genre or per-author lemma and tag statistics toward corpus-wide counts before comparing texts.
• I3 (frames): tag greeting, closing and backchannel speech acts in prag as formulaic, so that they are not over-interpreted as carrying content or implicature.
• I4 (frames): model conventional responses such as hi→hello or bye→bye as explicit adjacency-pair rules in prag, kept apart from learned stylistic alignment.
• I5 (world): track when a character's wording or state first appears and then recurs in another character's speech, giving the event-order model evidence of influence.