Improved Coreference Resolution Using Cognitive Insights

Coreference resolution is the task of extracting referential expressions, or mentions, in text and clustering these by the entity or concept they refer to. The sustained research interest in the task reflects the richness of reference expression usage in natural language and the difficulty in encoding insights from linguistic and cognitive theories effectively. In this thesis, we design and implement LIMERIC, a state-of-the-art coreference resolution engine. LIMERIC naturally incorporates both non-local decoding and entity-level modelling to achieve the highly competitive benchmark performance of 64.22% and 59.99% on the CoNLL-2012 benchmark with a simple model and a baseline feature set. As well as strong performance, a key contribution of this work is a reconceptualisation of the coreference task. We draw an analogy between shift-reduce parsing and coreference resolution to develop an algorithm which naturally mimics cognitive models of human discourse processing. In our feature development work, we leverage insights from cognitive theories to improve our modelling. Each contribution achieves statistically significant improvements and sum to gains of 1.65% and 1.66% on the CoNLL-2012 benchmark, yielding performance values of 65.76% and 61.27%. For each novel feature we propose, we contribute an accompanying analysis so as to better understand how cognitive theories apply to real language data. LIMERIC is at once a platform for exploring cognitive insights into coreference and a viable alternative to current systems. We are excited by the promise of incorporating our and further cognitive insights into more complex frameworks since this has the potential to both improve the performance of computational models, as well as our understanding of the mechanisms underpinning human reference resolution