Regina Barzilay: Learning to Model Text Structure

Regina Barzilay: Learning to Model Text Structure

by Regina Barzilay

Formal & Physical Sciences AI — Models & Research UYQArtificial intelligence (AI)UYQLNatural language and machine translation
🎙 Regina Barzilay 👥 4K 📅 December 14, 2025 ⏱ 63 min 👁 51 📄 lecture 🧭 2026-08-16
Available in: English (current) Français

Keywords

entity gridcoherencecenteringdiscourseranking

Summary

Regina Barzilay presents her work on statistical models for text coherence. She begins by illustrating the problem of incoherent outputs from automatic summarization and generation systems, showing examples that are grammatically correct but semantically disjointed. She argues that current systems lack a mechanism to evaluate global text structure. To address this, she introduces the entity grid, a representation that abstracts a text into a matrix of discourse entities and their syntactic roles across sentences. This grid is then used to derive features based on transition patterns of entities, which are fed into a ranking model to distinguish coherent from incoherent texts. The approach is inspired by centering theory, which posits regularities in how discourse entities are introduced and maintained. Barzilay demonstrates the effectiveness of the model on two tasks: information ordering (distinguishing original from permuted texts) and summary coherence evaluation, outperforming a baseline based on latent semantic analysis. She also discusses the importance of feature selection and the use of n-grams of entity transitions. The talk concludes with a discussion of the model’s limitations and potential extensions.

177 words

Critical Evaluation

Value of the Information & Strength of the Argument

The talk provides a valuable contribution by proposing a novel, data-driven approach to modeling text coherence, moving beyond manual rule-based systems. The argumentation is solid: the problem is clearly motivated with concrete examples, the proposed method is well-defined and grounded in linguistic theory (centering), and the experimental results on multiple tasks demonstrate its practical utility. The speaker also addresses potential objections, such as the role of focus and the need for feature selection, showing a nuanced understanding of the challenges. The comparison with an LSA baseline effectively highlights the advantages of the entity grid approach, particularly in handling redundancy in summaries.

Scientific Rigor, Source Quality, Title Accuracy

The talk is scientifically rigorous, presenting a method that is both theoretically motivated and empirically validated. The speaker cites relevant linguistic theories (centering theory) and prior work (e.g., LSA by Foltz), and the experimental design is appropriate for the claims made. The title accurately reflects the content, which focuses on learning to model text structure. The talk is a lecture, not a peer-reviewed publication, but the speaker is a leading researcher in NLP, and the work has been published in reputable venues (though not explicitly mentioned in the talk). The description provides no additional sources, but the talk itself references the joint work with M. Lapata and L. Lee, which is a credible source.

230 words

Title / Content Match

The title accurately reflects the content, which focuses on statistical models for text structure, specifically coherence modeling.

Quality & Reliability

8/10

The talk presents a well-established research method (entity grid) with clear theoretical grounding in centering theory, and demonstrates results on multiple tasks. The speaker is a renowned researcher, and the content is technically rigorous, though it is a lecture rather than a peer-reviewed publication.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

  • Barzilay, R., & Lapata, M. (2008). Modeling Local Coherence: An Entity-Based Approach. — The talk is based on joint work with Mirella Lapata and Lillian Lee, and this paper is the primary reference for the entity grid method.
  • Foltz, P. W. (1996). Latent Semantic Analysis for Text-Based Research. — Mentioned as the baseline method for coherence evaluation.

Concurring Sources

Contribution & Novelties

The talk presents a novel, data-driven approach to modeling text coherence using entity grids, which captures distributional patterns of discourse entities. This is a significant departure from rule-based discourse theories, offering a scalable and linguistically motivated alternative. The method is shown to be effective on multiple tasks, including information ordering and summary evaluation, demonstrating its practical applicability.

Pour aller plus loin :

  • Centering Theory — Foundational linguistic theory that motivates the entity grid approach.
  • Entity Grid — The original paper by Barzilay and Lapata introducing the entity grid model.
  • Latent Semantic Analysis — The baseline method used for comparison in the talk.

102 words

Radar Profile

The radar profile shows high scores across all dimensions, indicating a well-balanced and reliable presentation. The talk is technically deep, provides substantial information, and is based on credible research, though it is not a peer-reviewed publication itself.

Reliability 8/10