An Entity-Level Approach to Information Extraction

Aria Haghighi, Dan KleinACL 2010 · OpenReview

Abstract

We present a generative model of template-filling in which coreference resolution and role assignment are jointly determined. Underlying template roles first generate abstract entities, which in turn generate concrete textual mentions. On the standard corporate acquisitions dataset, joint resolution in our entity-level model reduces error over a mention-level discriminative approach by up to 20%.

Introduction

Traditional Information Extraction (IE) systems typically operate at the mention level, attempting to classify individual text spans (e.g., "Esso Inc." or "the company") into template roles in isolation. However, information about a single entity is often distributed across a document through various linguistic forms, including proper names, nominal descriptions, and pronouns. By treating coreference resolution as a separate post-processing step or ignoring it entirely, standard systems fail to leverage the collective context provided by all mentions of an entity.

We propose a shift toward entity-level extraction. By jointly modeling coreference and role assignment, our system can aggregate evidence across an entire document. For instance, a pronoun "it" appearing as the subject of "acquired" provides a strong signal for the "Purchaser" role, which can then be propagated back to the entity's proper name, even if the name itself appears in a less informative context.

Generative Entity-Level Model

Our approach utilizes a generative, hierarchical model that describes how a document is produced from a set of underlying template roles. Instead of classifying mentions, the model asks how the observed text was generated from latent entities.

The generative model architecture showing the hierarchy from roles to entities to mentions.

The generation process follows three primary stages:

  1. Roles to Entities: Template roles (e.g., PURCHASER, SELLER) generate abstract entities. Each role is characterized by a distribution over lexical items and frequencies.
  2. Entities to Mentions: These abstract entities generate concrete textual mentions. The model distinguishes between proper names, nominals, and pronouns, applying different generation rules for each type.
  3. Discourse Component: This component determines which entity is "evoked" at any given point in the text, facilitating the resolution of pronominal antecedents based on previous mentions.

To train the model, we employ Variational Expectation-Maximization (EM) to iteratively estimate the latent entity and role assignments that best fit the observed document text.

Experimental Results

The model was evaluated on the standard Corporate Acquisitions dataset and augmented with unannotated data from the Reuters-21857 corpus to demonstrate its ability to learn from unlabeled text.

Comparison of mention and entity accuracy across independent and joint models.

As shown in the results, the joint modeling of coreference and roles significantly outperforms independent baselines. The inclusion of pronominal mentions (JOINT+PRO) provides a further boost, improving mention labeling accuracy from 64.6% to 68.2%. Overall, the entity-level approach achieved a 20% error reduction over discriminative mention-level baselines. On the full extraction task, the model reached an F1 score of 52.9, establishing a new state-of-the-art on this benchmark compared to the previous best of 45.9.

Conclusion

By treating entities rather than mentions as the primary units of extraction, we demonstrate that coreference resolution and role assignment are mutually informative tasks. The generative framework allows for the effective use of pronouns and unlabeled data, leading to substantial improvements in extraction performance. Future work will address the current limitation of one-to-one role-to-entity mappings to handle more complex event structures involving multiple actors in the same role.

Summary updated

Abstract

We present a generative model of template-filling in which coreference resolution and role assignment are jointly determined. Underlying template roles first generate abstract entities, which in turn generate concrete textual mentions. On the standard corporate acquisitions dataset, joint resolution in our entity-level model reduces error over a mention-level discriminative approach by up to 20%.
BibTeX
@inproceedings{haghighi2010entity,
  author = {Aria Haghighi and Dan Klein},
  title = {{An Entity-Level Approach to Information Extraction}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2010},
  url = {https://openreview.net/forum?id=ByNRj2guWB},
}
Figure