Joint Neural Entity Disambiguation with Output Space Search

Proceedings of the 27th International Conference on Computational Linguistics, pages 2170–2180 Santa Fe, New Mexico, USA, August 20-26, 2018. 2170 Joint Neural Entity Disambiguation with Output Space Search Hamed Shahbazi, Xiaoli Z. Fern, Reza Ghaeini, Chao Ma, Rasha Obeidat, Prasad Tadepalli Oregon State University, Corvallis, OR, USA {shahbazh, xfern, ghaeinim, machao, obeidatr, tadepall}@eecs.oregonstate.edu Abstract In this paper, we present a novel model for entity disambiguation that combines both local con- textual information and global evidences through Limited Discrepancy Search (LDS). Given an input document, we start from a complete solution constructed by a local model and conduct a search in the space of possible corrections to improve the local solution from a global view point. Our search utilizes a heuristic function to focus more on the least confident local decisions and a pruning function to score the global solutions based on their local fitness and the global coherences among the predicted entities. Experimental results on CoNLL 2003 and TAC 2010 benchmarks verify the effectiveness of our model. 1 Introduction The goal of entity disambiguation is to link a set of given query mentions in a document to their referent entities in a Knowledge Base (KB). As an essential and challenging task in Knowledge-Base Population (KBP) for text Analysis (Ji et al., 2014; Heng et al., 2015), entity disambiguation has attracted many research efforts from the NLP community. Recently, deep learning based approaches have demonstrated strong performances on this task (Ganea and Hofmann, 2017; Sil et al., 2018). A main challenge for entity disambiguation is to best identify and represent the appropriate context, which can be local or global. Different methods have been proposed to capture and represent different types of contexts. Textual context has been heavily investigated for local models that score each query’s candidates independently. Representations of the textual contexts range from weighted combination of the word embeddings based on attention (Ganea and Hofmann, 2017), to more fine-grained contextual representations using recurrent neural networks (Sil et al., 2018). The global context of other entities in the document has also been studied for a more global and joint prediction view on the problem. Ganea and Hofmann (2017) use a Conditional Random Field (CRF) based model to capture the interrelationship among entities in the same document, whereas Globerson et al. (2016) introduce a soft k-max attention model to weigh the importance of other entities in the document in making prediction for any given query. Working with the recently proposed models, we observe that local models that employ an appropriate attention mechanism often have a solid linking performance. In a single document, there are often a small number of hard queries for which the local model fails to make a correct decision. We conjecture that if some of these mistakes can be corrected, a global model that enforces coherence among entities will be able to propagate these corrections to improve the overall solution quite effectively. This inspires us to consider the Limited Discrepancy Search (LDS) framework (Doppa et al., 2014), which conducts …

Content truncated to 500 words