Distant Supervision for Entity Linking

Distant Supervision for Entity Linking Miao Fan∗, Qiang Zhou and Thomas Fang Zheng CSLT, Division of Technical Innovation and Development Tsinghua National Laboratory for Information Science and Technology Department of Computer Science and Technology, Tsinghua University, Beijing, 100084, China ∗fanmiao.cslt.thu@gmail.com Abstract Entity linking is an indispensable oper- ation of populating knowledge reposito- ries for information extraction. It stud- ies on aligning a textual entity mention to its corresponding disambiguated entry in a knowledge repository. In this paper, we propose a new paradigm named dis- tantly supervised entity linking (DSEL), in the sense that the disambiguated entities that belong to a huge knowledge reposi- tory (Freebase) are automatically aligned to the corresponding descriptive webpages (Wiki pages). In this way, a large scale of weakly labeled data can be generat- ed without manual annotation and fed to a classifier for linking more newly dis- covered entities. Compared with tradi- tional paradigms based on solo knowl- edge base, DSEL benefits more via joint- ly leveraging the respective advantages of Freebase and Wikipedia. Specifically, the proposed paradigm facilitates bridging the disambiguated labels (Freebase) of entities and their textual descriptions (Wikipedi- a) for Web-scale entities. Experiments conducted on a dataset of 140,000 items and 60,000 features achieve a baseline F1- measure of 0.517. Furthermore, we ana- lyze the feature performance and improve the F1-measure to 0.545. 1 Introduction To build the “Digital Alexandria Library” for our human race, researchers in the NLP community have dedicated themselves to Information Extrac- tion (Sarawagi, 2008) over the past decades. In- formation extraction focuses on processing natu- ral language text to produce structured knowledge, which is usually represented as triples (two entities and their relation) for the convenience of storage in a database, retrieval, or even automatic reason- ing. For example, if we send a natural language sentence, Michael Jordan visited CMU yesterday, to the pipeline of information extraction machine, it will be processed by three operations in advance, i.e., • Named Entity Recognition (Nadeau and Sekine, 2007): Entities should firstly be i- dentified and classified into predefined cate- gories, such as person (PER), location (LOC) and organization (ORG). The sentence will be annotated as [Michael Jordan]/PER vis- ited [CMU]/ORG yesterday, after being pro- cessed by this operation. • Coreference Resolution (Ng, 2010): Some entities may have alias or abbreviations. It is well known that CMU is the abbreviation for Carnegie Mellon University. The knowl- edge repository may only store the regular- ized name, e.g., Carnegie Mellon University, for this named entity, so coreference resolu- tion is indeed necessary. • Relation Extraction (Bach and Badaskar, 2007): After both of the named entities ([Michael Jordan]/PER and [Carnegie Mel- lon University]/ORG) are recognized and regularized, we begin to study on the rela- tion between them. In this case, we extrac- t the verb visited and map it to the relation visit. Then the output will be a triple, i.e., (Michael Jordan [PER], visit, Carnegie Mel- lon University [ORG]). So far, we only abstract the triple as the struc- tured knowledge from …

Content truncated to 500 words