Proceedings of the First Workshop on Computational Models of Reference, Anaphora and Coreference

Massimo Poesio, Vincent Ng, Maciej OgrodniczukNAACL 2018

TL;DR

This document presents the proceedings of the NAACL HLT 2018 workshop on Computational Models of Reference, Anaphora and Coreference. It showcases recent advancements in understanding and computationally modeling how language refers to entities, events, and concepts, addressing challenges in coreference resolution, anaphora resolution, and related linguistic phenomena across various languages and genres.

What Problem Does This Paper Address?

The field of computational linguistics has long grappled with the challenge of understanding how language refers to things in the world. This includes resolving pronouns (like "he" or "it") to their antecedents, identifying when different phrases refer to the same entity (coreference), and understanding more complex forms of reference like bridging anaphora (e.g., "I went to a restaurant. The waiter was friendly.") and discourse deixis. While significant progress has been made, dedicated workshops focusing on the broad spectrum of reference, anaphora, and coreference have been infrequent since the end of the DAARC series. This workshop aims to fill that gap by providing a forum for researchers to present work on all aspects of computational modeling of these phenomena, fostering collaboration and advancing the state-of-the-art.

Methods

The papers presented in this workshop cover a diverse range of methodologies, reflecting the multifaceted nature of reference, anaphora, and coreference resolution. These include:

  • Machine Learning Approaches: Various machine learning techniques are employed for anaphora and coreference resolution, including rule-based systems, learning-based methods, and neural network models.
  • Neural Network Relation Classifiers: Integration of predictions from neural network relation classifiers is explored to improve coreference and bridging resolution.
  • Data Analysis and Rule-based Experiments: For languages like German, researchers conduct data analysis to inform the development of rule-based systems for bridging resolution.
  • Corpus Development: The creation and utilization of annotated corpora are central to many of the presented works, enabling the training and evaluation of models.
  • Cognitive Modeling: Some research delves into modeling brain activity associated with pronoun resolution, bridging linguistic computation with cognitive processes.

Datasets

The workshop features research utilizing and contributing to several datasets:

  • ARRAU Corpus: This is a key dataset used in the workshop's shared task. It is an anaphorically annotated corpus of English that includes both written text and spoken dialogue, annotated for identity anaphora, bridging reference, and discourse deixis.
  • Twitter Conversations: One paper focuses on anaphora resolution in Twitter conversations, suggesting the use of this social media platform as a data source.
  • German Corpora: Research on German anaphora and coreference utilizes specific German corpora for data analysis and rule-based experiments.
  • PAWS (Parallel Treebank with Anaphoric Relations): This is a newly introduced multi-lingual parallel treebank with anaphoric relations, aiming to facilitate cross-lingual research.

Key Insights and Ideas

  • Bridging Reference is a Significant Challenge: Several papers highlight the difficulty of resolving bridging references, which often require world knowledge and deeper contextual understanding beyond simple antecedent identification.
  • Cross-lingual and Multi-lingual Approaches are Crucial: The development of resources like the PAWS treebank underscores the growing importance of multi-lingual approaches to anaphora and coreference resolution, acknowledging that linguistic phenomena manifest differently across languages.
  • Integrating Diverse Information Sources Improves Performance: Combining predictions from different types of models, such as neural network relation classifiers with traditional coreference systems, shows promise for enhanced accuracy.
  • Social Media Presents Unique Anaphora Challenges: The informal and often fragmented nature of social media, like Twitter, introduces specific challenges for anaphora resolution that require tailored approaches.
  • Cognitive Evidence Informs Computational Models: Investigating the neural correlates of pronoun resolution can provide valuable insights for developing more human-like computational models.

Key Results

  • The workshop accepted 11 out of 16 submissions, indicating a competitive acceptance rate of 68%.
  • The accepted papers cover anaphora annotation and resolution in 6 different languages.
  • Specific quantitative results are detailed within individual papers, but a general summary across all papers is not provided in the introductory sections. However, the diversity of topics suggests progress in areas like bridging resolution, coreference projection, and the modeling of anaphoric phenomena in various genres and languages.

Limitations and Future Work

  • Limited Dedicated Forums: The introduction notes the infrequency of dedicated workshops on this broad topic, highlighting a need for continued community engagement.
  • Under-investigated Phenomena: While this workshop aims for a broader focus, previous workshops have concentrated on under-investigated coreference phenomena, suggesting that some areas still require more attention.
  • Future Directions: The workshop aims to be the first in a series, indicating a commitment to continued research and community building in this area. Future work will likely involve developing more robust models for challenging phenomena like bridging reference, expanding multi-lingual resources, and integrating insights from cognitive science.

Summary updated

BibTeX
@inproceedings{poesio2018proceedings,
  author = {Massimo Poesio and Vincent Ng and Maciej Ogrodniczuk},
  title = {{Proceedings of the First Workshop on Computational Models of Reference, Anaphora and Coreference}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2018},
}
Figure