Models Cascade for Tree-Structured Named Entity Detection

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 1269–1278, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Models Cascade for Tree-Structured Named Entity Detection Marco Dinarelli LIMSI-CNRS / Orsay Cedex, France marcod@limsi.fr Sophie Rosset LIMSI-CNRS / Orsay Cedex, France rosset@limsi.fr Abstract Named Entity Recognition (NER) is a well-known Natural Language Processing (NLP) task, used as a preliminary process- ing to provide a semantic level to more complex tasks. In this paper we describe a new set of named entities having a multi- level tree structure, where base entities are combined to define more complex ones. This definition makes the NER task more complex than previous tasks, even more due to the use of noisy data for the anno- tation: transcriptions of French broadcast data. We propose an original and effec- tive system to tackle this new task, putting together the strengths of solutions for se- quence labeling approaches and syntactic parsing via cascading of different mod- els. Our system was evaluated in the 2011 Quaero named entity detection evaluation campaign and ranked first, with results far better than those of the other participating systems. 1 Introduction Named Entity Detection is a well-known NLP task used to extract semantic information in other more complex tasks such as Relation Extrac- tion (Doddington et al., 2004) or Question An- swering (Voorhees, 2001). After its definition in MUC-6 (Grishman and Sundheim, 1996), the NER task evolved increasing its complexity. Cur- rent definitions provide a fine-grained semantic in- formation level with a broad coverage (Sekine, 2004). This has increased the interests in devel- oping named entity detection systems. In this paper we describe a new set of named en- tities defined recently(Grouin et al., 2011). These named entities have a multilevel tree structure where components are combined to define more complex and general entity structures. This defi- nition increases significantly the complexity of the NER task, even more due to the type of data used for the annotation: manual and automatic tran- scriptions of French broadcast data. Given such a definition, it is not possible to tackle the task with traditional sequence labeling approaches. At the same time, solutions able to re- construct tree structures from flat sequences, like syntactic parsing solutions, may have serious lim- itations, due to the noisy data used. In order to solve these problems, we studied and implemented a two-stage system that put together the strengths of the two approaches: i) a complex linear-chain Conditional Random Field (CRF) model, integrat- ing a huge number of features, takes the noisy data as input and generates the corresponding sequence of components; ii) a syntactic parsing model based on Probabilistic Context Free Grammar (PCFG) reconstructs the tree-structured named entities. We describe our solution showing that it is equivalent, in the processing steps, to a syntac- tic parsing analysis solution, and that it can better handle noisy data. Additionally, given the amount of data used and the relatively slow training time of CRF models when using large amount …

Content truncated to 500 words