Probabilistic Reasoning for Entity & Relation Recognition

Probabilistic Reasoning for Entity & Relation Recognition∗ Dan Roth Wen-tau Yih Department of Computer Science University of Illinois at Urbana-Champaign {danr, yih}@uiuc.edu Abstract This paper develops a method for recognizing rela- tions and entities in sentences, while taking mutual dependencies among them into account. E.g., the kill (Johns, Oswald) relation in: “J. V. Oswald was murdered at JFK after his assassin, K. F. Johns...” depends on identifying Oswald and Johns as people, JFK being identified as a location, and the kill relation between Oswald and Johns; this, in turn, enforces that Oswald and Johns are people. In our framework, classifiers that identify entities and relations among them are first learned from local infor- mation in the sentence; this information, along with con- straints induced among entity types and relations, is used to perform global inference that accounts for the mutual dependencies among the entities. Our preliminary experimental results are promising and show that our global inference approach improves over learning relations and entities separately. 1 Introduction Recognizing and classifying entities and relations in text data is a key task in many NLP problems such as in- formation extraction (IE) (Califf and Mooney, 1999; Freitag, 2000; Roth and Yih, 2001), question an- swering (QA) (Voorhees, 2000) and story comprehen- sion (Hirschman et al., 1999). In a typical IE application of constructing a jobs database from unstructured text, the system has to extract meaningful entities like title and salary and, ideally, to determine whether the entities are associated with the same position. In a QA system, many questions ask for specific entities involved in some rela- tions. For example, the question “Where was Poe born?” in TREC-9 asks for the location entity in which Poe was born. The question “Who killed Lee Harvey Oswald?” seeks a person entity that has the relation kill with the person Lee Harvey Oswald. In all earlier works we know of, the tasks of identify- ing entities and relations were treated as separate prob- lems. The common procedure is to first identify and clas- sify entities using a named entity recognizer and only ∗ Research supported by NSF grants CAREER IIS-9984168 and ITR IIS-0085836 and an ONR MURI Award. then determine the relations between the entities. How- ever, this approach has several problems. First, errors made by the named entity recognizer propagate to the relation classifier and may degrade its performance sig- nificantly. For example, if “Boston” is mislabeled as a person, it will never be classified as the location of Poe’s birthplace. Second, relation information is sometimes crucial to resolving ambiguous named entity recognition. For instance, if the entity “JFK” is identified as the vic- tim of the assassination, the named entity recognizer is unlikely to misclassify it as a location (e.g. JFK airport). This paper develops a novel approach for this prob- lem – a probabilistic framework for recognizing entities and relations together. In this framework, separate clas- sifiers are first trained for entities and relations. Their output is used to represent a conditional distribution …

Content truncated to 500 words