Low-Resource Named Entity Recognition Based on Multi-hop Dependency Trigger

Low-Resource Named Entity Recognition Based on Multi-hop Dependency Trigger Jiangxu Wu* wujx27@mail2sysu.edu.cn Peiqi Yan* yanpeiqiswu@163.com Abstract This paper introduces DepTrigger, a simple and effective model in low-resource named entity recognition (NER) based on multi-hop dependency triggers. Dependency triggers refer to salient nodes relative to an entity in the dependency graph of a context sentence. Our main observa- tion is that triggers generally play an important role in recognizing the location and the type of entity in a sentence. Instead of exploiting the manual labeling of triggers, we use the syntac- tic parser to annotate triggers automatically. We train DepTrigger using an independent model architectures which are Match Network encoder and Entity Recognition Network encoder. Com- pared to the previous model TriggerNER, DepTrigger outperforms for long sentences, while still maintain good performance for short sentences as usual. Our framework is significantly more cost-effective in real business. 1 Introduction Named Entity Recognition (NER) aims to detect the span from text belonging to the semantic category such as person, location, organization, etc. NER plays a core component in many NLP tasks and is widely employed in downstream applications, such as knowledge graph (Ji, 2021), question answering (Molla, 2004) and dialogue system (Peng, 2020). The deep-learning based approaches have shown remarkable success in NER, while it requires large corpora annotated with named entities. Moreover, in many practical settings, we wish to apply NER to domains with a very limited amount of labeled data since annotating data is a labor-intensive and time-consuming task. Therefore, it is an emergency to improve the performance of the deep-learning based NER model with limited labeled data. Previous work in low-resource NER mainly focused on meta-learning (Snell, 2017), distantly super- vision (Yang, 2018), transfer learning (Lin, 2017),et al. Recently, (CLin, 2020) proposed an approach based on entity trigger called TriggerNER. The key idea is that an entity trigger is a group of words that can help explain the recognition process of an entity in a sentence. Considering the sentence “Biden is the president of ”, we are able to infer that there is a country entity on the underline according to “the president of”. In this case, “the president of” is a group of triggers. Experiments reveal that the performance of utilizing 20% of the trigger-annotated sentences is comparable to that of exploiting 70% of conventional annotated sentences. However, crowd-sourced entity trigger annotations, which suffer from the same problem as traditional annotation, require labor costs and expert experience. Inspired by attribute triggers in Attribute Extraction (Huang2, 2017), this paper presents an alternative approach to automatically annotate the trigger in a sentence by utilizing the syntactic parse. Fig. 1 is the dependency parse result of the sentence “Alice was born in Beijing”, the relation “nsubj:pass” shows that the subject of “born” is “Alice”. According to the meaning of “born”, we are capable of inferring that “Alice” is a person entity. Inspired by this fact, we propose a novel model, namley DepTrigger, which explore the structures of dependency trees and utilize the …

Content truncated to 500 words