Table Filling Multi-Task Recurrent Neural Network for Joint Entity and Relation Extraction

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 2537–2547, Osaka, Japan, December 11-17 2016. Table Filling Multi-Task Recurrent Neural Network for Joint Entity and Relation Extraction Pankaj Gupta Corporate Technology, Siemens AG CIS, University of Munich (LMU) Munich, Germany gupta.pankaj.ext@siemens.com pankaj.gupta@campus.lmu.de Hinrich Sch¨utze CIS, University of Munich (LMU) inquiries@cis.lmu.org Bernt Andrassy Corporate Technology, Siemens AG Munich, Germany bernt.andrassy@siemens.com Abstract This paper proposes a novel context-aware joint entity and word-level relation extraction ap- proach through semantic composition of words, introducing a Table Filling Multi-Task Recurrent Neural Network (TF-MTRNN) model that reduces the entity recognition and relation classifica- tion tasks to a table-filling problem and models their interdependencies. The proposed neural network architecture is capable of modeling multiple relation instances without knowing the corresponding relation arguments in a sentence. The experimental results show that a simple approach of piggybacking candidate entities to model the label dependencies from relations to entities improves performance. We present state-of-the-art results with improvements of 2.0% and 2.7% for entity recognition and relation classification, respectively on CoNLL04 dataset. 1 Introduction Relation classification is defined as the task of predicting the semantic relation between the annotated pairs of nominals (also known as relation arguments). These annotations, for example named entity pairs participating in a relation are often difficult to obtain. Traditional methods are often based on a pipeline of two separate subtasks: Entity Recognition (ER1) and Relation Classification (RC), to first detect the named entities and then performing relation classification on the detected entity mentions, therefore ignoring the underlying interdependencies and propagating errors from the entity recognition to relation classification. The two subtasks together are known as End-to-End relation extraction. Relation classification is treated as a sentence-level multi-class classification problem, which often assume a single relation instance in the sentence. It is often assumed that entity recognition affects the relation classification, but it is not affected by relation classification. Here, we reason with experimental evidences that the latter is not true. For example, in Figure 1, relation Work For exists between PER and ORG entities, ORGBased in between ORG and LOC, while Located In between LOC and LOC entities. Inversely, for a given word with associated relation(s), the candidate entity types can be detected. For example, in Figure 2, for a given relation, say Located in, the candidate entity pair is (LOC, LOC). Therefore, the two tasks are interdependent and optimising a single network for ER and RC to model the interdependencies in the candidate entity pairs and corresponding relations is achieved via the proposed joint modeling of subtasks and a simple piggybacking approach. Joint learning approaches (Roth and Yih, 2004; Kate and Mooney, 2010) built joint models upon com- plex multiple individual models for the subtasks. (Miwa and Sasaki, 2014) proposed a joint entity and relation extraction approach using a history-based structured learning with a table representation; how- ever, they explicitly incorporate entity-relation label interdependencies, use complex features and search heuristics to fill table. In addition, their state-of-the-art method is structured prediction …

Content truncated to 500 words