Knowledge-Enhanced Named Entity Disambiguation for Short Text

Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 10th International Joint Conference on Natural Language Processing, pages 735–744 December 4 - 7, 2020. c⃝2020 Association for Computational Linguistics 735 Knowledge-Enhanced Named Entity Disambiguation for Short Text Zhifan Feng, Qi Wang, Wenbin Jiang, Yajuan Lyu, Yong Zhu Baidu Inc., Beijing, China {fengzhifan,wangqi31,jiangwenbin,lvyajuan,zhuyong}@baidu.com Abstract Named entity disambiguation is an important task that plays the role of bridge between text and knowledge. However, the performance of existing methods drops dramatically for short text, which is widely used in actual applica- tion scenarios, such as information retrieval and question answering. In this work, we pro- pose a novel knowledge-enhanced method for named entity disambiguation. Considering the problem of information ambiguity and incom- pleteness for short text, two kinds of knowl- edge, factual knowledge graph and conceptual knowledge graph, are introduced to provide ad- ditional knowledge for the semantic matching between candidate entity and mention context. Our proposed method achieves significant im- provement over previous methods on a large manually annotated short-text dataset, and also achieves the state-of-the-art on three standard datasets. The short-text dataset and the pro- posed model will be publicly available for re- search use. 1 Introduction Name entity disambiguation (NED) aims to asso- ciate each entity mention in the text with its cor- responding entity in the knowledge graph (KG). It plays an important role in many text-related ar- tificial intelligent tasks such as recommendation and conversation, since it works as a bridge be- tween text and knowledge. In decades, researchers devoted their efforts to NED in many ways, in- cluding the rule-based methods (Shen et al., 2014), the conventional statistic methods (Shen et al., 2014) and the deep learning methods (Octavian- Eugen Ganea, 2017). On formal text, state-of-the- art methods achieve high performance thanks to the well-written utterance and rich context. However, experiments show that the performance of these methods degrades dramatically on informal text, for example, the short text widely used in many real application scenarios such as information re- trieval and human-machine interaction. It is diffi- cult for existing methods to make decisions on the non-standard utterance without adequate context. The discrimination procedure of NED depends on sufficient context in the input text, which is usually noisy and scarce in the short text used in in- formation retrieval and human-machine interaction. For example, an analysis based on search engine logs demonstrates that a search query contains 2.35 words on average (Yi Fang, 2011). Such short text could not provide adequate context which is neces- sary for NED models. In recent years, many efforts improve NED by exploiting more powerful models and richer context information (Shen et al., 2014). These methods mainly focus on the better utiliza- tion of existing context. Therefore, they can not improve NED effectively on short text since the problem of information shortage still exists. Intu- itively, it is hard for NED to achieve essential im- provement on short text if it can not exploit external …

Content truncated to 500 words