Joint Inference for Knowledge Base Population

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1912–1923, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Joint Inference for Knowledge Base Population Liwei Chen1, Yansong Feng 1∗, Jinghui Mo1, Songfang Huang2, and Dongyan Zhao1 1ICST, Peking University, Beijing, China 2IBM China Research Lab, Beijing, China {chenliwei,fengyansong,mojinghui,zhaodongyan}@pku.edu.cn huangsf@cn.ibm.com Abstract Populating Knowledge Base (KB) with new knowledge facts from reliable text re- sources usually consists of linking name mentions to KB entities and identifying relationship between entity pairs. How- ever, the task often suffers from errors propagating from upstream entity linkers to downstream relation extractors. In this paper, we propose a novel joint infer- ence framework to allow interactions be- tween the two subtasks and find an opti- mal assignment by addressing the coher- ence among preliminary local predictions: whether the types of entities meet the ex- pectations of relations explicitly or implic- itly, and whether the local predictions are globally compatible. We further measure the confidence of the extracted triples by looking at the details of the complete ex- traction process. Experiments show that the proposed framework can significantly reduce the error propagations thus obtain more reliable facts, and outperforms com- petitive baselines with state-of-the-art re- lation extraction models. 1 Introduction Recent advances in natural language processing have made it possible to construct structured KBs from online encyclopedia resources, at an un- precedented scale and much more efficiently than traditional manual edit. However, in those KBs, entities which are popular to the community usu- ally contain more knowledge facts, e.g., the bas- ketball player LeBron James, the actor Nicholas Cage, etc., while most other entities often have fewer facts. On the other hand, knowledge facts should be updated as the development of entities, such as changes in the cabinet, a marriage event, or an acquisition between two companies, etc. In order to address the above issues, we could consult populating existing KBs from reliable text resources, e.g., newswire, which usually involves enriching KBs with new entities and populating KBs with new knowledge facts, in the form of <Entity, Relation, Entity> triple. In this paper, we will focus on the latter, identifying relationship be- tween two existing KB entities. This task can be intuitively considered in a pipeline paradigm, that is, name mentions in the texts are first linked to entities in the KB (entity linking, EL), and then the relationship between them are identified (re- lation extraction, RE). It is worth mentioning that the first task EL is different from the task of named entity recognition (NER) in traditional informa- tion extraction (IE) tasks, where NER recognizes and classifies the entity mentions (to several pre- defined types) in the texts, but EL focuses on link- ing the mentions to their corresponding entities in the KB. Such pipeline systems often suffer from errors propagating from upstream to downstream, since only the local best results are selected to the next step. One idea to solve the problem is to allow interactions among the local …

Content truncated to 500 words