Leveraging multi-token entities in document-level named entity recognition
Proposed an attention-based document-level NER model that uses global context features (document-level) as well as local context features (sentence-level) for NER! Most NER model focuses on each sentence-level extraction independently and this could easily lead to confusion in entity types due to lack of a macro context. This paper focuses specifically on news articles and proposed MEID, Multi-token Entity Informed Document-level model for document-level NER. The whole idea being that we want to utilise the context information of multi-token entities as they are usually less ambiguous than single-token entities. In fact, 26.62% of single-token entities in CoNLL-2003 dataset are constituents of multi-token entities in the same document. The paper also designed an auxiliary task of Multi-token Entity Classification to detect multi-token entities and this is trained together with the document-level NER model without extra annotation!
Multi-grained named entity recognition
Proposed MGNER, Multi-Grained Named Entity Recognition, that is able to extract BOTH non-overlapping and nested entities. This is in contrast to previous work where models can either do one or the other! NER is common treated as a sequence labelling problem but this caused a major problem that it only trains models to recognise non-overlapping entities. There are now many approaches that can explicitly detect nested entities but they usually don’t perform well on non-overlapping entities. MGNER works by first detecting entity positions in different granularities using a Detector and then classify these entities into different categories using a Classifier. MGNER is highly modularised with 5 different modules:
This modularisation allows people to have high customisability in their NER models! MGNER achieved new SOTA results on both Nested and Non-overlapping NER tasks!