Cross-sentence contexts in Named Entity Recognition with BERT

Abstract

Named entity recognition (NER) is a task under the broader scope of Natural Language Processing (NLP). The computational task of NER is often cast as a sequence classification task where the goal is to label each word (or token) in the input sequence with a class from a predefined set of classes. The development of deep transfer learning methodologies in recent years has greatly influenced both NLP and NER. There have been improvements in the performance of NER models but at the same time the use of cross-sentence context, the sentences around the sentence of interest, has diminished in NER methods. Many of the current methods use inputs that consist of only one sentence of text at a time. It is nevertheless clear that useful information for NER is often found also elsewhere in text. Recent self-attention models like BERT can both capture long-distance relationships in input and represent inputs consisting of several sentences. This creates opportunities for making use of cross-sentence information in NLP tasks. This thesis presents a systematic study exploring the use of cross-sentence information for NER using BERT models in five languages. The study shows that adding context as additional sentences to BERT input systematically increases NER performance. Adding multiple sentences in input samples also allows the study of predictions for the sentences in different contexts. A straightforward method of Contextual Majority Voting (CMV) is proposed to combine these different predictions. The study demonstrates that using CMV increases NER performance even further. Evaluation of the proposed methods on established datasets, including the Conference on Computational Natural Language Learning CoNLL'02 and CoNLL'03 NER benchmarks, demonstrates that the proposed approach can improve on the state-of-the-art NER results for English, Dutch, and Finnish, achieves the best reported BERT-based results for German, and is on par with other BERT-based approaches for Spanish. The methods implemented for this work are published under open licenses.