Neural Fine-Grained Entity Type Classification with Hierarchy-Aware Loss

Proceedings of NAACL-HLT 2018, pages 16–25 New Orleans, Louisiana, June 1 - 6, 2018. c⃝2018 Association for Computational Linguistics Neural Fine-Grained Entity Type Classification with Hierarchy-Aware Loss Peng Xu Department of Computing Science University of Alberta Edmonton, Canada pxu4@ualberta.ca Denilson Barbosa Department of Computing Science University of Alberta Edmonton, Canada denilson@ualberta.ca Abstract The task of Fine-grained Entity Type Clas- sification (FETC) consists of assigning types from a hierarchy to entity mentions in text. Ex- isting methods rely on distant supervision and are thus susceptible to noisy labels that can be out-of-context or overly-specific for the train- ing sentence. Previous methods that attempt to address these issues do so with heuristics or with the help of hand-crafted features. Instead, we propose an end-to-end solution with a neu- ral network model that uses a variant of cross- entropy loss function to handle out-of-context labels, and hierarchical loss normalization to cope with overly-specific ones. Also, previous work solve FETC a multi-label classification followed by ad-hoc post-processing. In con- trast, our solution is more elegant: we use pub- lic word embeddings to train a single-label that jointly learns representations for entity men- tions and their context. We show experimen- tally that our approach is robust against noise and consistently outperforms the state-of-the- art on established benchmarks for the task. 1 Introduction Fine-grained Entity Type Classification (FETC) aims at labeling entity mentions in context with one or more specific types organized in a hier- archy (e.g., actor as a subtype of artist, which in turn is a subtype of person). Fine-grained types help in many applications, including rela- tion extraction (Mintz et al., 2009), question an- swering (Li and Roth, 2002), entity linking (Lin et al., 2012), knowledge base completion (Dong et al., 2014) and entity recommendation (Yu et al., 2014). Because of the high cost in labeling large training corpora with fine-grained types, current FETC systems resort to distant supervision (Mintz et al., 2009) and annotate mentions in the train- ing corpus with all types associated with the en- tity in a knowledge graph. This is illustrated in Figure 1, with three training sentences about en- tity Steve Kerr. Note that while the entity be- longs to three fine-grained types (person, athlete, and coach), some sentences provide evidence of only some of the types: person and coach from S1, person and athlete from S2, and just person for S3. Clearly, direct distant supervision leads to noisy training data which can hurt the accuracy of the FETC model. One kind of noise introduced by distant super- vision is assigning labels that are out-of-context (athlete in S1 and coach in S2) for the sentence. Current FETC systems sidestep the issue by ei- ther ignoring out-of-context labels or using simple pruning heuristics like discarding training exam- ples with entities assigned to multiple types in the knowledge graph. However, both strategies are in- elegant and hurt accuracy. Another source of noise introduced by distant supervision is when the type is overly-specific for the context. For instance, ex- ample …

Content truncated to 500 words