Coreference Resolution as Query-based Span Prediction (1911.01746v4)

Abstract: In this paper, we present an accurate and extensible approach for the coreference resolution task. We formulate the problem as a span prediction task, like in machine reading comprehension (MRC): A query is generated for each candidate mention using its surrounding context, and a span prediction module is employed to extract the text spans of the coreferences within the document using the generated query. This formulation comes with the following key advantages: (1) The span prediction strategy provides the flexibility of retrieving mentions left out at the mention proposal stage; (2) In the MRC framework, encoding the mention and its context explicitly in a query makes it possible to have a deep and thorough examination of cues embedded in the context of coreferent mentions; and (3) A plethora of existing MRC datasets can be used for data augmentation to improve the model's generalization capability. Experiments demonstrate significant performance boost over previous models, with 87.5 (+2.5) F1 score on the GAP benchmark and 83.1 (+3.5) F1 score on the CoNLL-2012 benchmark.

• The paper introduces CorefQA, reframing coreference resolution as query-based span prediction, offering flexibility, deeper context analysis, and data augmentation opportunities.
• CorefQA achieves significant performance gains, with F1 scores of 83.1 on CoNLL-2012 and 87.5 on GAP, improving upon previous models.
• This framework reduces the need for domain-specific datasets and suggests potential for transforming other NLP tasks into question-answering formats.

CorefQA: Coreference Resolution as Query-based Span Prediction

The paper "CorefQA: Coreference Resolution as Query-based Span Prediction" introduces an innovative approach to coreference resolution by reframing the task as a span prediction problem, akin to question answering. This methodological shift offers notable advantages over traditional coreference resolution systems.

Formulation and Advantages

CorefQA formulates coreference resolution as a query-based task, where each candidate mention generates a query using its surrounding context. It then employs a span prediction module to identify coreferences within the document. This formulation extends the capabilities of coreference resolution systems in three key ways:

1. Retrieval Flexibility: By framing the problem in a question answering context, CorefQA allows the system to retrieve mentions that may have been missed during the mention proposal stage.
2. Contextual Examination: Encoding a mention and its context in a query enables a deeper analysis of contextual cues embedded around coreferent mentions.
3. Data Augmentation: The framework can leverage existing question answering datasets for data augmentation, which significantly enhances the model's generalization capability.

Experimental Insights

The implementation of CorefQA demonstrates substantial performance improvements. Specifically:

• On the CoNLL-2012 benchmark, CorefQA achieves an F1 score of 83.1, which is a notable 3.5 point increase over previous models.
• On the GAP benchmark, it attains an F1 score of 87.5, outperforming existing systems by 2.5 points.

These results underscore the effectiveness of the query-based span prediction framework in improving coreference resolution performance.

Practical and Theoretical Implications

The proposed framework has significant implications both in practice and theory:

• Practical Benefits: By allowing the reuse of question answering datasets, CorefQA reduces the overhead associated with creating domain-specific coreference datasets. This practical advantage facilitates broader application and scalability across various NLP tasks.
• Theoretical Contributions: This work challenges traditional paradigms in NLP by integrating coreference resolution with question answering, paving the way for more nuanced cross-task methodologies.

Future Directions in AI

The success of CorefQA hints at future explorations where more NLP tasks may be efficiently transformed into question answering frameworks. Such approaches could be valuable in advancing models that are both cross-linguistic and domain-adaptive, leading to AI systems with enhanced reasoning capabilities.

Overall, the work presented in "CorefQA: Coreference Resolution as Query-based Span Prediction" marks a progressive step towards more versatile and cognitively robust natural language understanding systems. It opens avenues for further research in combining different NLP tasks to exploit synergies among them, potentially leading to advancements in how LLMs are structured and trained.