Discourse Structure and Dialogue Acts in Multiparty Dialogue: the STAC Corpus

Authors: Nicholas Asher, Julie Hunter, Mathieu Morey, Farah Benamara, Stergos Afantenos · Year: 2016 · Venue: LREC 2016
Link: aclanthology.org/L16-1432 · Licence: CC BY-NC 4.0 (ELRA, LREC proceedings) — summary only; STAC data has its own terms — test use only

92 citations openalex-snapshot · 2026-10-14110 citations in our library of the 140673 papers whose bibliographies we readProceedings of the Tenth International Conference on Language Resources and Evaluation (LREC'16)licence

Publisher's page PDF DOI 10.63317/3dr9b3ndpn84 OpenAlex

Authors (5)

Nicholas Asher ORCID · Semantic Scholar · h 50 · 10780 citations · 474 works
Julie Hunter ORCID · Semantic Scholar · h 15 · 675 citations · 67 works
Mathieu Morey ORCID · Semantic Scholar · h 8 · 400 citations · 24 works
Farah Benamara ORCID · Semantic Scholar · h 23 · 2036 citations · 203 works
Stergos Afantenos Semantic Scholar · h 15 · 844 citations · 87 works

Citations and author metrics from openalex-snapshot (2026-10-14); authors with the same name are not always told apart.

What it is about

STAC is a corpus of chats from online games of The Settlers of Catan (36 games fully annotated in this release), in which players bargain for resources while competing, so they sometimes withhold information or mislead («anyone got some ore?» – «nope sorry» from a player holding two). The chats are messy (creative nouns, verb ellipsis such as «I can wheat for clay») and full of interleaved threads, and many moves only make sense against game events that nobody describes in words: a player steals a card, the victim answers «touché», «it» refers to the stolen resource. Annotation follows SDRT in two layers: dialogue acts (offers, counteroffers, acceptances, refusals), then discourse units linked by typed relations into a graph, not a tree, with complex units built from smaller ones. Graphs are needed because one move can attach to several earlier ones (a single «kk» acknowledges three refusals). The commonest relations are question–answer pair, question elaboration and acknowledgement; correction of an earlier move, rare in text schemes, is frequent.

Ideas

I1. Link a social act to the event it reacts to: a thank Say frame to the preceding giving or helping Event, an apologize to the harm, a «touché»-like express to the action just done to the speaker. In tales this lets the reader answer «Why did the girl thank the old woman?» from the event, which a Say frame alone cannot. A rule in prag/src/frames.rs (the nearest earlier Event whose agent is the addressee and whose patient or recipient is the speaker); measure with world knowledge diagnose on why-questions about speech.
I2. Add a correction link between frames: «No, it was the hare who won» corrects an earlier claim; the SLM student should then remember the corrected version and mark the old one as replaced. Detect it from a reply cue (no, not … but) plus the same predicate with a different role filler; test with minimal pairs in course-style lessons.
I3. Allow one reply to depend on several earlier lines (an acknowledgement or a summary answer to many questions) instead of exactly one, when the dependence link proposed under bunt-2020-iso-standard-dialogue-act-annotation I1 is built.
I4. Keep strategic speech in mind for the student: characters in tales also lie or withhold (the wolf at the door). A claim's truth is its principal's commitment until the story confirms it (see faller-2019-discourse-commitments-illocutionary-reportatives I1).

For Mova

languages/en/36-discourse-relations.md §5 (SDRT) and §11, 38-dialogue-and-discourse-markers.md §8–11. Code: prag/src/frames.rs (reaction and correction links), world/src/knowledge.rs (student memory). Measures: diagnosis of why-questions about speech; minimal-pair tests for corrections.

Library summary

The Research library's own summary of the paper, in our words — not the paper's text.

Written for the Research library (claude-opus-5-5) from the paper text. Original — aclanthology.org/L16-1432.pdf

Summary

Most discourse-annotated resources cover single-authored written text and assume tree-shaped structures. This leaves multi-party conversation poorly documented. The paper presents STAC, a collection of online chats recorded while people played a modified, open-source version of The Settlers of Catan and bargained over resources. The chats are annotated with full discourse structures following SDRT. Text is split into EDUs, units are attached to one another (sometimes grouped into CDUs), and each link is labelled with a relation type. Game events such as dice rolls, card plays and thefts are logged and aligned with the chat. These non-verbal moves can therefore act as arguments of discourse relations and provide antecedents for pronouns and ellipsis. The authors argue that some patterns cannot be represented by tree-based frameworks (RST, LDM, DLTAG). One example is a single acknowledgment aimed at several replies, which forms a "lozenge" graph. The authors describe STAC as, to their knowledge, the first resource with complete discourse structures for multi-party dialogue. It supports research on strategic, partly non-cooperative talk, interleaved threads, noisy creative language, and discourse parsing.

Key points

Data source: chat negotiations from an online Catan game. Game moves and chat are automatically aligned in .soclog logs, so whole games can be replayed. The raw, segmented and annotated files are publicly released.
The annotation model is graph-based SDRT. Its structure contains discourse units, relation edges, CDU-membership edges and relation labels, with no cycles. SDRT's Right Frontier Constraint is adapted for multi-party dialogue.
The most frequent relations are QAP, Question-Elaboration and Acknowledgment. Correction is also prominent. Temporal relations such as Narration and Background are rare.
The language is messy and inventive, with misspellings, novel nouns and verbs ("dolly" for sheep), and verb ellipsis lacking an overt antecedent. The data also contains crossing, non-projective threads.
Dialogue is mostly reactive: attachments 2–5 turns back are common, and links beyond five turns are rare. Players are competitors who may withhold or misrepresent information. The project was funded by ERC grant 269427 (Strategic Conversation).

Topics: Discourse analysis · Speech act theory

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.63317/3dr9b3ndpn84

@inproceedings{asher2016discourse,
  title = {{Discourse Structure and Dialogue Acts in Multiparty Dialogue: the STAC Corpus}},
  author = {Asher, Nicholas and Hunter, Julie and Morey, Mathieu and Benamara, Farah and Afantenos, Stergos},
  booktitle = {Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC'16)},
  year = {2016},
  doi = {10.63317/3dr9b3ndpn84},
  url = {https://doi.org/10.63317/3dr9b3ndpn84}
}