Elamur Research⌕
☰
home / papers / 2020 / Givenness Hierarchy Theoretic Cognitive Status Filtering

Givenness Hierarchy Theoretic Cognitive Status Filtering

Paper · 2020 · Poulomi Pal, Lixiao Zhu, Andrea Golden-Lasher et al.

Summary

Robots that talk need to pick referring forms such as "it", "this" or "that NP" the way people do. The Givenness Hierarchy (GH) ties each form to the cognitive status a referent has for the listener: in focus, activated, familiar, and so on. Earlier full-GH systems by Williams and colleagues served understanding. Given a form, they searched status-indexed memory for candidate referents. Generation reverses the question, because the speaker already knows the entity and must estimate its status. The authors treat this as Bayesian filtering. Each object mentioned in the dialogue gets its own Cognitive Status Filter (CSF), which tracks a distribution over In Focus, Activated and Familiar. After every utterance, the filter updates based on whether the object was not mentioned, mentioned, or mentioned as topic. The transition probabilities were learned from Mechanical Turk judgments on a small portion of the OFAI Multimodal Task Description Corpus that had been rendered into silver-standard English. The CSF matched or slightly beat a rule-based FSM built from the GH coding protocol, and both far outperformed a random baseline. This suggests the protocol's heuristics can be learned from data and leaves room to add uncertain cues later.

Key points

  • Model: for each object, the status distribution after utterance t is derived from the previous status and the current linguistic status (N = not mentioned, M = non-topic mention, T = topic mention). Statuses are limited to I/A/F.
  • Familiarity is handled implicitly: an object counts as Familiar or higher simply if a CSF exists for it, meaning it has appeared in the conversation.
  • Data: Task 3 of the OFAI MTD corpus (a tube-mounting demonstration with 8 objects, including teacher and learner). Only 4 monologues of 4 utterances each were used, translated from German into silver-standard English. Ground truth came from a crowdsourced experiment.
  • Accuracy: U-model (uniform prior) 82.03, I-model (weakly informed prior) 81.25, FSM 81.25, random baseline (RB) 32.81.
  • Pairwise McNemar's tests found no significant difference among the CSF variants and the FSM (e.g., U vs FSM χ² ≈ 0, p = 1). Every model versus RB was significant (χ² ≈ 46–49, p < 0.0001). Data, scripts and outputs are on OSF (osf.io/qse7y).

Interesting ideas

  • I1. Each discourse entity carries its own small recursive filter, updated after every sentence from how (or whether) the entity was mentioned, which gives a cheap, inspectable salience state.
  • I2. The paper stresses an asymmetry between directions: understanding searches memory starting from the referring form, while generation starts from a known entity and asks for its status.
  • I3. Encoding "familiar" as the mere existence of a tracker avoids modeling a status that is trivially decidable.
  • I4. The hand-written rules from the coding protocol are encoded as an FSM baseline, and the learned model is compared against it and a random baseline using paired McNemar tests rather than raw accuracy alone.
  • I5. The authors admit their crowd question forced annotators to choose some referent for "it" even when none fit, and they propose adding a "none" option.

For Mova

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

  • I1 (frames): coreference could keep a per-entity I/A/F distribution and rank pronoun antecedents by it, logging each update so every resolution can be traced back to its mentions.
  • I2 (back): the generator could look up an entity's status in the meaning graph and choose "it", "that X" or a full NP accordingly, making pronoun choice in the text → graph → text compressor principled rather than ad hoc.
  • I3 (world): creating an entity record on first mention can double as the familiarity flag, with graded salience stored only for entities already in the world.
  • I4 (learn): when a learned component replaces a rule set, Mova could run McNemar on per-item outcomes against both the rules and a random negative control to show whether the change is real.
  • I5 (learn): prompts for the LLM teacher tiers and silver-data witnesses should always permit "no referent / no answer," so that forced guesses do not leak into the training data.

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

 Institutions

🎓Colorado School of MinesGolden · United States2 works · 2 citations

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 9,086 worksArtificial Intelligence · 8,889 worksArts and Humanities · 9,296 worksLanguage and Linguistics · 9,206 works

Cite

Cite as Poulomi Pal, Lixiao Zhu, Andrea Golden-Lasher, Akshay Swaminathan and Tom Williams. 2020. Givenness Hierarchy Theoretic Cognitive Status Filtering. arXiv:2005.11267.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2005.11267 · arXiv 2005.11267

@misc{pal2020givenness,
  title = {{Givenness Hierarchy Theoretic Cognitive Status Filtering}},
  author = {Poulomi Pal, Lixiao Zhu, Andrea Golden-Lasher, Akshay Swaminathan and Tom Williams},
  year = {2020},
  doi = {10.48550/arxiv.2005.11267},
  eprint = {2005.11267},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2005.11267}
}

References

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

  1. Bangalore, S., & Johnston, M. (2003). Balancing data-driven and rule-based approaches in the context of a multimodal conversational system. In Proceedings of the 2003 IEEE workshop on automatic speech recognition and understanding. find
  2. Bangalore, S., & Rambow, O. C. (2005). Probabilistic model for natural language generation. Google Patents. (US Patent 6,947,885) find
  3. Bostanci, B., & Bostanci, E. (2013). An evaluation of classification algorithms using McNemar’s test. In Bio-inspired computing: Theories and app. (BIC-TA). find
  4. Chai, J. Y., Hong, P., & Zhou, M. X. (2004). A probabilistic approach to reference resolution in multimodal user interfaces. In Intelligent user interfaces (IUI). find
  5. Chai, J. Y., Prasov, Z., & Qu, S. (2006). Cognitive principles in robust multimodal interpretation. Journal of Artificial Intelligence Research, 27, 55–83. ↗
  6. Clark, A. F., & Clark, C. (1999). Performance characterization in computer vision: a tutorial. find
  7. Cowan, N. (1998). Attention and memory: An integrated framework (Vol. 26). Oxford University Press. find
  8. Dale, R., & Reiter, E. (1995). Computational interpretations of the gricean maxims in the generation of referring expressions. Cognitive science, 19(2), 233–263. ↗
  9. Fay, M. P. (2011). Exact McNemar’s test and matching confidence intervals. find
  10. Grice, H. P. (1975). Logic and conversation. In Syntax and semantics 3: Speech acts (pp. 41–58). ↗
  11. Gundel, J. K., Bassene, M., Gordon, B., Humnick, L., & Khalfaoui, A. (2010). Testing predictions of the givenness hierarchy framework: A crosslinguistic investigation. Journal of Pragmatics, 42(7), 1770–1785. ↗
  12. Gundel, J. K., Hedberg, N., & Zacharski, R. (1993). Cognitive status and the form of referring expressions in discourse. Language, 274–307. ↗
  13. Gundel, J. K., Hedberg, N., Zacharski, R., Mulkern, A., Custis, T., Swierzbin, B., . . . Bassene, M. (2006). Coding protocol for statuses on the givenness hierarchy. Unpublished manuscript (1993/2006). http://www.sfu. ca/hedberg/Coding for Cognitive Status. pdf . ↗
  14. Hedberg, N. (2013). Applying the givenness hierarchy framework: Methodological issues. International workshop on information structure of Austronesian languages. find
  15. Kehler, A. (2000). Cognitive status and form of reference in multimodal human-computer interaction. In Proceedings of AAAI/IAAI. find
  16. Liddell, D. (1976). Practical tests of 2 × 2 contingency tables. Journal of the Royal Statistical Society: Series D (The Statistician), 25(4), 295–304. ↗
  17. McNemar, Q. (1947). Note on the sampling error of the difference between correlated proportions or percentages. Psychometrika, 12(2), 153–157. ↗
  18. Moore, C., Dunham, P. J., & Dunham, P. (2014). Joint attention: Its origins and role in development. Psychology Press. find
  19. Peeters, D., Azar, Z., & Özyürek, A. (2014). The interplay between joint attention, physical proximity, and pointing gesture in demonstrative choice. In Proceedings of the 36th annual meeting of the cognitive science society (pp. 1144–1149). find
  20. Scheutz, M., Williams, T., Krause, E., Oosterveld, B., Sarathy, V., & Frasca, T. (2019). An overview of the distributed integrated cognition affect and reflection DIARC architecture. In Cognitive architectures (pp. 165–193). Springer. ↗
  21. Schreitter, S., & Krenn, B. (2016). The OFAI multi-modal task description corpus. In Proceedings of the international conference on language resources and evaluation (LREC) (pp. 1408–1414). find
  22. Williams, T., Acharya, S., Schreitter, S., & Scheutz, M. (2016). Situated open world reference resolution for human-robot dialogue. In Proceedings of the international conference on human-robot interaction (HRI) (pp. 311–318). find
  23. Williams, T., Krause, E., Oosterveld, B., & Scheutz, M. (2018). Towards givenness and relevance-theoretic open world reference resolution. In RSS workshop on models and representations for natural human-robot communication. find
  24. Williams, T., & Scheutz, M. (2019). Reference in robotics: A givenness hierarchy theoretic approach. The Oxford Handbook of Reference. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/pal-2020-givenness-hierarchy-theoretic-cognitive-status/