Information extraction provides the basic technical support for knowledge graph construction and Web applications. Named entity recognition(NER) is one of the fundamental tasks of information extraction. Recognizing unseen entities from numerous contents with the support of only a few labeled samples, also termed as few-shot learning, is a crucial issue to be studied. Few-shot NER aims at identifying emerging named entities from the context with the support of a few labeled samples. Existing methods mainly use the same strategy to construct a single prototype for each entity or non-entity class, which has limited expressiveness power and even biased representation. In this work, we propose a novel hybrid multi-prototype class representation approach. Specifically, for entity classes, we first insert labels after entities in support sentences to enrich the learned token and label embeddings with more contextual information. Then, for each entity span, the contextual token embeddings are averaged to form its entity-level prototype, while the contextual label embedding is considered as its label-level prototype. The set of prototypes for all entities in a class constitute the multi-prototype of this entity class. For non-entity class, we directly use the set of token embeddings to represent it, where multi-prototype refers to the multiple token embeddings. By treating the entity and non-entity classes differently, our hybrid strategy can extract more precise class representations from the support examples. Furthermore, we establish a harder and more reasonable experimental setting of few-shot NER by offering a rigorous sampling strategy. Extensive empirical results show that our proposal improves F1 scores by 3%∼10% absolute points over prior models on popular benchmark Few-NERD under both loose and our proposed rigorous sampling constraints, achieving state-of-the-art performance.

Information extraction provides the basic technical support for knowledge graph construction and Web applications. Named entity recognition(NER) is one of the fundamental tasks of information extraction. Recognizing unseen entities from numerous contents with the support of only a few labeled samples, also termed as few-shot learning, is a crucial issue to be studied. Few-shot NER aims at identifying emerging named entities from the context with the support of a few labeled samples. Existing methods mainly use the same strategy to construct a single prototype for each entity or non-entity class, which has limited expressiveness power and even biased representation. In this work, we propose a novel hybrid multi-prototype class representation approach. Specifically, for entity classes, we first insert labels after entities in support sentences to enrich the learned token and label embeddings with more contextual information. Then, for each entity span, the contextual token embeddings are averaged to form its entity-level prototype, while the contextual label embedding is considered as its label-level prototype. The set of prototypes for all entities in a class constitute the multi-prototype of this entity class. For non-entity class, we directly use the set of token embeddings to represent it, where multi-prototype refers to the multiple token embeddings. By treating the entity and non-entity classes differently, our hybrid strategy can extract more precise class representations from the support examples. Furthermore, we establish a harder and more reasonable experimental setting of few-shot NER by offering a rigorous sampling strategy. Extensive empirical results show that our proposal improves F1 scores by 3%∼10% absolute points over prior models on popular benchmark Few-NERD under both loose and our proposed rigorous sampling constraints, achieving state-of-the-art performance.

In recent years, Coronavirus disease 2019 (COVID-19) has become a global epidemic, and some efforts have been devoted to tracking and controlling its spread. Extracting structured knowledge from involved epidemic case reports can inform the surveillance system, which is important for controlling the spread of outbreaks. Therefore, in this paper, we focus on the task of Chinese epidemic event extraction (EE), which is defined as the detection of epidemic-related events and corresponding arguments in the texts of epidemic case reports. To facilitate the research of this task, we first define the epidemic-related event types and argument roles. Then we manually annotate a Chinese COVID-19 epidemic dataset, named COVID-19 Case Report (CCR). We also propose a novel hierarchical EE architecture, named multi-model fusion-based hierarchical event extraction (MFHEE). In MFHEE, we introduce a multi-model fusion strategy to tackle the issue of recognition bias of previous EE models. The experimental results on CCR dataset show that our method can effectively extract epidemic events and outperforms other baselines on this dataset. The comparative experiments results on other generic datasets show that our method has good scalability and portability. The ablation studies also show that the proposed hierarchical structure and multi-model fusion strategy contribute to the precision of our model.

No abstract