Transfer learning for fine-grained entity typing

摘要

Fine-grained entity typing (FGET) is to classify the mentions of entities into hierarchical fine-grained semantic types. There are two main issues with existing FGET approaches. Firstly, the process of training corpora for FGET is normally to label the data automatically, which inevitably induces noises. Existing approaches either directly tweak noisy labels in corpora by heuristics or algorithmically retreat to parental types, both leading to coarse-grained type labels instead of fine-grained ones. Secondly, existing approaches usually use recurrent neural networks to generate feature representations of mention phrases and their contexts, which, however, perform relatively poor on long contexts and out-of-vocabulary (OOV) words. In this paper, we propose a transfer learning-based approach to extract more efficient feature representations and offset label noises. More precisely, we adopt three transfer learning schemes: (i) transferring sub-word embeddings to generate more efficient OOV embeddings; (ii) using a pre-trained language model to generate more efficient context features; (iii) using a pre-trained topic model to transfer the topic-type relatedness through topic anchors and select confusing fine-grained types at inference time. The pre-trained topic model can offset the label noises without retreating to coarse-grained types. The experimental results demonstrate the effectiveness of our transfer learning approach for FGET.