FINET: Context-Aware Fine-Grained Named Entity Typing

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 868–878, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. FINET: Context-Aware Fine-Grained Named Entity Typing Luciano Del Corro∗ Abdalghani Abujabal∗ Rainer Gemulla† Gerhard Weikum∗ ∗ Max-Planck-Institut f¨ur Informatik, Saarbr¨ucken, Germany {delcorro, abujabal, weikum}@mpi-inf.mpg.de † Universit¨at Mannheim, Mannheim, Germany rgemulla@uni-mannheim.de Abstract We propose FINET, a system for detect- ing the types of named entities in short inputs—such as sentences or tweets—with respect to WordNet’s super fine-grained type system. FINET generates candidate types using a sequence of multiple extrac- tors, ranging from explicitly mentioned types to implicit types, and subsequently selects the most appropriate using ideas from word-sense disambiguation. FINET combats data scarcity and noise from ex- isting systems: It does not rely on supervi- sion in its extractors and generates training data for type selection from WordNet and other resources. FINET supports the most fine-grained type system so far, includ- ing types with no annotated training data. Our experiments indicate that FINET out- performs state-of-the-art methods in terms of recall, precision, and granularity of ex- tracted types. 1 Introduction Named entity typing (NET) is the task of detecting the type(s) of a named entity in context. For in- stance, given “John plays guitar on the stage”, our goal is to infer that “John” is a guitarist or a mu- sician and a person. We propose FINET, a system for detecting the types of named entities in short inputs—such as sentences or tweets—with respect to WordNet’s super fine-grained type system (16k types of organizations, persons and locations). Named entity typing is a fundamental building block for many natural-language processing tasks. NET is at the heart of information extraction meth- ods for finding types for entities in a knowledge base1 (KB) (Mitchell et al., 2015). Likewise, NET aids named entity disambiguation by reducing the candidate space for a given entity mention. Entity types are an important resource for entity-based retrieval or aggregation tasks, such as semantic search (Hoffart et al., 2014) or question answer- ing (Yahya et al., 2013). Finally, type information helps to increase the semantic content of syntactic patterns (Nakashole et al., 2012) or in open infor- mation extraction (Lin et al., 2012). The extraction of explicit types has been studied in the literature, most prominently in the context of taxonomy induction (Snow et al., 2006). Ex- plicit types occur, for example, in phrases such as “Steinmeier, the German Foreign Minister, [...]” or “Foreign Minister Steinmeier.” These explicit types are often extracted via patterns, such as the well-known Hearst patterns (Hearst, 1992), and subsequently integrated into a taxonomy. Pattern- based methods often have high precision but low recall: Types are usually mentioned when a named entity is introduced or expected to be unknown to readers, but often are not explicitly stated. The NET problem differs from taxonomy induction in that (1) the type system is prespecified, (2) types are disambiguated, and (3) types are associated with each occurrence of named entity in context. Our FINET system …

Content truncated to 500 words