Elamur Research⌕
☰
home / papers / 2009 / Annotating Spoken Dialogs: From Speech Segments to Dialog Acts and Frame Semantics

Annotating Spoken Dialogs: From Speech Segments to Dialog Acts and Frame Semantics

Summary

Most Spoken Language Understanding systems either use hand-crafted semantic grammars, which have weak coverage and robustness, or flat attribute-value labeling, which does not capture how concepts relate to each other. Within the EU LUNA project, the authors built a multi-layer annotated corpus of Italian spoken dialogs from a software/hardware help desk. The goal is to support statistical semantic parsers and richer dialog systems. The corpus combines real human-human (HH) calls and human-machine (HM) dialogs collected with a Wizard of Oz setup. Its layers are:

  • POS and morphosyntax (EAGLES)
  • domain attribute-value concepts
  • syntactic parses
  • dialog acts, based on a compact taxonomy derived from DAMSL, TRAINS and DIT++
  • predicate argument structures following FrameNet, extended with new frames where English ones were missing

The paper describes the annotation pipeline and tools and gives frequency statistics for both corpora. It also estimates mutual information between frames and dialog acts. Most frames are strongly linked to a single dialog act, so shallow semantics looks like a useful feature for dialog act classification. The authors present this as the first SDS corpus to combine HH and HM dialogs with lexical, semantic and dialog layers.

Key points

  • Corpus plan and taxonomy: The corpus is planned at 1000 dialogs, split evenly between HH and HM. The dialog act taxonomy has 16 tags in four groups: core, conventional, feedback/turn management, and Other. It is also used by the statistical Dialogue Manager of the ADAMACH project.
  • Tools: Utterance segmentation and dialog act labeling were done by hand in MMAX. For PAS annotation, utterances were parsed with Bikel's parser trained for Italian, and frames were annotated in Salto, which stores data in TIGER-XML.
  • Segmentation differences: HH dialogs required 167 turn mergings and 225 splittings. HM dialogs needed only 158 splittings, and dialog act variety was higher in HH.
  • HM dialog acts: The HM corpus contains 947 dialog acts, led by Info-request (26.3%), Answer (18.1%) and Inform (17.2%). Its most frequent frame bigram is Greeting–Being named (17.1%).
  • Mutual information analysis: Using 20 HH dialogs and 50 HM dialogs, the authors used weighted conditional entropy to approximate mutual information between frames and dialog acts. Examples of strong links:
    • In HM, info-req pairs with Being in operation, and act-req pairs with Replacing.
    • In HH, quit pairs with Greeting.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Linguistics & languages

Topics: Speech act theory

Cited by

18 papers of our library cite it (of the 140673 whose bibliographies we read); 2 of them have a page here, 16 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{dinarelli2009annotating,
  title = {{Annotating Spoken Dialogs: From Speech Segments to Dialog Acts and Frame Semantics}},
  author = {Dinarelli, Marco and Quarteroni, Silvia and Tonelli, Sara and Moschitti, Alessandro and Riccardi, Giuseppe},
  booktitle = {Proceedings of SRSL 2009, the 2nd Workshop on Semantic Representation of Spoken Language},
  year = {2009},
  doi = {10.3115/1626296.1626301},
  url = {https://doi.org/10.3115/1626296.1626301}
}

References

12 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. C. F. Baker, C. J. Fillmore, and J. B. Lowe. 1998. The Berkeley FrameNet Project. In Proceedings of ACL/Coling’98, pages 86–90. find our card
  2. F. Bechet, G. Riccardi, and D. Hakkani-Tur. 2004. Mining spoken dialogue corpora for system evaluation and modeling. In Proceedings of EMNLP’04, pages 134–141. find
  3. H. Bunt. 2005. A framework for dialogue act specica-tion. In Proceedings of SIGSEM WG on Representation of Multimodal Semantic Information. find
  4. A. Burchardt, K. Erk, A. Frank, A. Kowalski, S. Padó, and M. Pinkal. 2006. Salto - a versatile multi-level annotation tool. In Proceedings of LREC 2006, pages 517–520, Genoa, Italy. find
  5. A. Corazza, A. Lavelli, and G. Satta. 2007. Analisi sintattica-statistica basata su costituenti. Intelligenza Artificiale, 4(2):38–39. find
  6. M. G. Core and J. F. Allen. 1997. Coding dialogs with the DAMSL annotation scheme. In Proceedings of the AAAI Fall Symposium on Communicative Actions in Humans and Machines. find
  7. R. De Mori, F. Bechet, D. Hakkani-Tur, M. McTear, G. Riccardi, and G. Tur. 2008. Spoken language understanding: A survey. IEEE Signal Processing magazine, 25(3):50–58. find
  8. G. Leech and A. Wilson. 2006. EAGLES recommendations for the morphosyntactic annotation of corpora. Technical report, ILC-CNR. find
  9. C. Müller and M. Strube. 2003. Multi-level annotation in MMAX. In Proceedings of SIGDIAL’03. find
  10. J. M. Sinclair and R. M. Coulthard. 1975. Towards an Analysis of Discourse: The English Used by Teachers and Pupils. Oxford University Press, Oxford. find
  11. D. Traum. 1996. Conversational agency: The TRAINS-93 dialogue manager. In Proceedings of TWLT 11: Dialogue Management in Natural Language Systems, pages 1–11, June. find
  12. S. Varges, G. Riccardi, and S. Quarteroni. 2008. Persistent information state in a data-centric architecture. In Proceedings of SIGDIAL’08. find

↑ back to top

This page in Markdown →