📄 Simultaneous Dialogue Act Segmentation and Labelling using Lexical and Syntactic Features

Summary

The paper tackles splitting spoken dialogue turns into dialogue act (DA) segments and labelling those segments at the same time, without assuming the turns are already segmented. The authors use the Switchboard corpus (SWBD) with SWBD-DAMSL labels. They parse the cleaned transcriptions with the Stanford Parser, which was trained on WSJ newswire text. For every word they record two features: the most general constituent that starts at it (MGSS) and the most general constituent that ends at it (MGSE). Their corpus analysis shows that the categories found at segment boundaries come from a small set. The base model is an HMM over words combined with an n-gram model over DA labels, decoded with Viterbi, so the segmentation falls out of the label sequence. The syntactic features enter as hard constraints during decoding: a boundary is allowed only where the MGSE or MGSS (or both) belong to a set of frequent boundary categories. The gains in labelling and segmentation are small but consistent. This suggests that even a crude use of syntax, taken from an out-of-domain parser, can guide the joint task.

Key points

• SWBD data: 1155 telephone conversations, about 205,000 utterances, 1.4M words, a vocabulary of about 22k words, and 42 SWBD-DAMSL labels. A turn holds 1.78 segments on average (std. dev. 1.41).
• Boundary statistics for intra-turn segments:
  • The last word of a segment most often closes ',' (37.1%), ROOT (33.9%), ':' (8.6%) or S (7.9%).
  • The first word of a segment most often opens ROOT (33.5%), CC (22.1%), NP (14.7%) or S (11.3%).
• Experimental setup:
  • Ten cross-validation runs, each with 1136 training dialogues and 19 test dialogues.
  • N-grams built with the SLM toolkit using Good-Turing discounting, and HMMs trained with Baum-Welch.
• Evaluation:
  • Labelling is scored with DA Error Rate (DAER) and with the share of turns labelled wrongly.
  • Segmentation is scored with an F-score on boundary positions.
• DAER results:
  • The baseline without syntax scores 54.41.
  • The best reported setting is MGSS restricted to the top 10 categories, at 52.76.
  • Some set sizes are slightly worse than the baseline, and allowing all categories gives 54.76.

Interesting ideas

• I1. Syntax as a decoding filter. Constituent information is not added as a learned feature; it only vetoes segment boundaries that fall inside unlikely constituents, which is a cheap way to inject structure into an existing probabilistic decoder.
• I2. Per-word features that summarise the tree. Each word gets the label of the largest constituent it opens and the largest one it closes, turning a whole parse into two simple tags that a sequence model can use.
• I3. Joint segmentation and labelling. Segment boundaries are read directly from where the most likely label sequence switches acts, so one search does both jobs and avoids errors piling up across two pipeline stages.
• I4. Tuning the size of the allowed set. The authors vary how many frequent boundary categories are permitted (5, 10, 20, all), and the results show that loosening the constraint fully erases the benefit.
• I5. Checking an out-of-domain parser before trusting it. A WSJ-trained parser is run on speech transcripts after normalising punctuation and stripping non-verbal tokens, and the boundary statistics are checked first to confirm the parse is still informative.

For Mova

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

• I1 (frames): the speech-act segmenter in prag could allow act boundaries only at UD subtree edges, so every proposed split can be explained by a tree position.
• I2 (tree): gparse output could be summarised the same way, as "largest subtree starting or ending here," to feed downstream clause and sentence splitters.
• I3 (frames): speech-act tagging of multi-act turns in dialogues within tales could produce segments and act types in a single Viterbi pass.
• I4 (learn): self-induced repair rules could be checked by sweeping their coverage threshold against held-out data, treating the unconstrained setting as a control.
• I5 (tokens): the tokenizer could normalise transcript markup and punctuation before parsing conversational text, then reinsert the markers afterwards, as the paper does.