Towards a Structured Representation of Generic Concepts and Relations in Large Text Corpora

Proceedings of Recent Advances in Natural Language Processing, pages 65–73, Hissar, Bulgaria, 7-13 September 2013. Towards a Structured Representation of Generic Concepts and Relations in Large Text Corpora Archana Bhattarai Department of Computer Science The University of Memphis abhattar@memphis.edu Vasile Rus Department of Computer Science The University of Memphis vrus@memphis.edu Abstract Extraction of structured information from text corpora involves identifying entities and the relationship between entities ex- pressed in unstructured text. We propose a novel iterative pattern induction method to extract relation tuples exploiting lexi- cal and shallow syntactic pattern of a sen- tence. We start with a single pattern to illustrate how the method explores addi- tional paterns and tuples by itself with in- creasing amount of data. We apply fre- quency and correlation based filtering and ranking of relation tuples to ensure the cor- rectness of the system. Experimental eval- uation compared to other state of the art open extraction systems such as Reverb, textRunner and WOE shows the effective- ness of the proposed system. 1 Introduction Traditional information extraction methodologies tend to extract a predefined relation between named entities annotated in a different process. While this method might be useful and accurate for smaller data with limited entity types and re- lations, it cannot scale to extract entities and their relationships in web due to the sheer volume and heterogeneity of data. Thus open domain infor- mation extraction systems such as Reverb (Fader et al., 2011), TEXTRUNNER (Yates et al., 2007) and NELL (Carlson et al., 2010) have received added attention in recent times. Extracting ma- chine readable structured information from free text is the basis of most of the semantic analytical systems. With these units of semantic informa- tion, a lot of applications requiring semantic in- formation processing such as finding the semantic similarity between two unit of texts, semantic in- ference, automated question-answering etc can be visualized with better performance. Existing work on pre-defined relation extraction have implemented methods of supervised, semi- supervised, bootstrapped and unsupervised classi- fication(Zhao and Grishman, 2005), (Kambhatla, 2004) (Bunescu and Mooney, 2006) (Zelenko et al., 2003). For open information extraction meth- ods, since they do not have predefined relations, it is very hard if impossible to generate labeled data for all potential relations in large text cor- pora. In this paper, we propose an iterative pat- tern induction based extraction system CREATE (Concept Representation and Extraction through Heterogenous Evidence), to extract relation tuples from large text corpora. We will start with a single selective pattern and iteratively add tuples and pat- terns in the corresponding collection. This method is easily usable in any domain since it does not re- quire any labeled data. We ensure the selectivity of the pattern by filtering the patterns with statis- tics such as frequency and average pointwise mu- tual information (PMI) and specificity of the pat- tern. CREATE works under the assumption that sentences have a pattern of expressing information and this pattern is followed by multiple sentences. If we can explore these patterns …

Content truncated to 500 words