Joint Bootstrapping of Corpus Annotations and Entity Types

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 436–446, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics Joint Bootstrapping of Corpus Annotations and Entity Types Hrushikesh Mohapatra Siddhanth Jain IIT Bombay Soumen Chakrabarti∗ Abstract Web search can be enhanced in powerful ways if to- ken spans in Web text are annotated with disambiguated entities from large catalogs like Freebase. Entity anno- tators need to be trained on sample mention snippets. Wikipedia entities and annotated pages offer high-quality labeled data for training and evaluation. Unfortunately, Wikipedia features only one-ninth the number of enti- ties as Freebase, and these are a highly biased sample of well-connected, frequently mentioned “head” entities. To bring hope to “tail” entities, we broaden our goal to a second task: assigning types to entities in Freebase but not Wikipedia. The two tasks are synergistic: know- ing the types of unfamiliar entities helps disambiguate mentions, and words in mention contexts help assign types to entities. We present TMI, a bipartite graphical model for joint type-mention inference. TMI attempts no schema integration or entity resolution, but exploits the above-mentioned synergy. In experiments involving 780,000 people in Wikipedia, 2.3 million people in Free- base, 700 million Web pages, and over 20 professional editors, TMI shows considerable annotation accuracy im- provement (e.g., 70%) compared to baselines (e.g., 46%), especially for “tail” and emerging entities. We also com- pare with Google’s recent annotations of the same corpus with Freebase entities, and report considerable improve- ments within the people domain. 1 Introduction Thanks to automatic information extraction and se- mantic Web efforts, keyword search over unstruc- tured Web text is rapidly evolving toward entity- and type-oriented queries (Guo et al., 2009; Pan- tel et al., 2012) over semi-structured databases such as Wikipedia, Freebase, and other forms of Linked Data. A key enabling component for such enhanced search capability is a type and entity catalog. This includes a directed acyclic graph of types under the subTypeOf relation between types, and entities at- tached to one or more types via instanceOf edges. ∗soumen@cse.iitb.ac.in YAGO (Suchanek et al., 2007) provides such a cat- alog by unifying Wikipedia and WordNet, followed by some cleanup. Another enabling component is an annotated cor- pus in which token spans (e.g., the word “Albert”) are identified as a mention of an entity (e.g., the Physicist Einstein). Equipped with suitable indices, a catalog and an annotated corpus let us find “sci- entists who played some musical instrument”, and answer many other powerful classes of queries (Li et al., 2010; Sawant and Chakrabarti, 2013). Consequently, accurate corpus annotation has been intensely investigated (Mihalcea and Csomai, 2007; Cucerzan, 2007; Milne and Witten, 2008; Kulkarni et al., 2009; Han et al., 2011; Ratinov et al., 2011; Hoffart et al., 2011). With two exceptions (Zheng et al., 2012; Gabrilovich et al., 2013) that we discuss later, public-domain corpus annotation work has almost exclusively used Wikipedia and deriva- tives, partly because Wikipedia provides not only a standardized space of entities, …

Content truncated to 500 words