Toward Givenness Hierarchy Theoretic Natural Language Generation

What it is about

A robot that talks with people must choose between it, this N, that N and the N. The Givenness Hierarchy (GH) says that the form reflects the cognitive status a referent has in the hearer's mind: in focus, activated, familiar, uniquely identifiable, referential, type identifiable. Earlier work used GH data structures to understand anaphora. This paper addresses generation. It proposes a per-entity Bayesian filter that keeps a distribution over three statuses (in focus, activated, familiar), updated by how each mention is used (not mentioned, mentioned in a non-topic role, mentioned as topic). It also gives a Describe algorithm that picks a form from the status, uses only same-or-higher-status entities as distractors, and adjusts the form for distance. No experiment is reported; the authors plan a human study of preferred referring forms.

Ideas

• I1. A count-based status tracker: per entity, a state among the three statuses, changed by a fixed transition table keyed on the grammatical role of the mention (subject or topic moves to in focus; object to activated; others to familiar). Store the table in a BTreeMap so updates are deterministic.
• I2. Form rule: choose it if the entity is in focus with no competing distractor, this N if activated, that N if familiar, and the N otherwise. Check the rule against the pronoun and demonstrative choices in Mova's own coreference data.
• I3. Narrow the distractor set by status: only entities of the same or higher status need ruling out. This shrinks the candidate set checked in coref/src/sieve.rs.
• I4. Negative control: a random status assignment must give worse form accuracy than the tracker.

For Mova

Chapter: 35. Reference and coreference, and 12. Pronouns, determiners and numerals for the forms. Code: coref/src/sieve.rs and coref/src/mention.rs. Measure: form accuracy (pronoun, demonstrative, definite) on a held-out set of Mova texts, and the mean number of distractors checked per mention; two runs must match.

Cite

The paper itself, not this card. BibTeX from the library's record of the work. arXiv 2007.16009

@misc{pal2020toward,
  title = {{Toward Givenness Hierarchy Theoretic Natural Language Generation}},
  author = {Pal, Poulomi and Williams, Tom},
  year = {2020},
  eprint = {2007.16009},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2007.16009}
}

References

2 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Arnold, J. E. (2016). Explicit and emergent mechanisms of information status. Topics in Cog. Sci.. find
2. Callaway, C. B., & Lester, J. C. (2002). Pronominalization in generated discourse and dialogue. Proc. ACL (pp. 88–95). find

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/pal-2020-toward-givenness-hierarchy-theoretic-natural/