A Computational Cognitive Model for Semantic Sub-Network Extraction from Natural Language Queries

Proceedings of COLING 2012: Technical Papers, pages 727–744, COLING 2012, Mumbai, December 2012. A Computational Cognitive Model for Semantic Sub- network Extraction from Natural Language Queries Suman DEB ROY Wenjun ZENG Department of Computer Science University of Missouri – Columbia, USA. sdr5x8@mail.missouri.edu, zengw@missouri.edu ABSTRACT Semantic query sub-network is the representation of a natural language query as a graph of semantically connected words. Such sub-networks can be identified as sub-graphs in larger ontologies like DBpedia or Google knowledge graph, which allows for domain and concepts identification, especially in noisy queries. In this paper, we present a novel standalone NLP technique that leverages the cognitive psychology notion of semantic forms for semantic sub- network extraction from natural language queries. Semantic forms, borrowed from cognitive psychology models, are one of the fundamental structures employed by human cognition to construct semantic information in the brain. We propose a computational cognitive model by means of conditional random fields and explore the interaction patterns among such forms. Our results suggest that the cognitive abstraction provided by semantic forms during labelling can significantly improve parsing and sub-network extraction compared to pure lexical approaches like parts of speech tagging. We conduct experiments on approximately 5000 queries from three diverse datasets to demonstrate the robustness and efficiency of the proposed approach. KEYWORDS: Cognitive Linguistics, Forms, Query, Search, Subnets, Semantic 727 1 Introduction The efficiency of natural language (NL) search often depends on detection of keywords in a query, followed by construction of some meaningful connected network comprising of such keywords (Herdagdelen, 2010). This connected network of keywords is called a semantic subnet (Booth, 2009). These keywords together comprise what is called a semantic field (Croft, 2003). The goal of our research is to efficiently recover the semantic sub-network from NL queries. Prior research suggests three main motivations for extracting semantic subnets from NL queries. Firstly, extracted query subnets can be used to generate a candidate set of concepts within a larger ontology (like of DBpedia RDF network/ Google knowledge graph), which may align to the words in the query subnet and assist domain identification and query expansion (Booth, 2009). Secondly, a query subnet can act as a NL interface to concept graph databases (Popescu, 2003), facilitating semantic information retrieval (Kauffman, 2007) and improved query understanding and semantic search (Hu, 2009). Finally, semantic subnets enable identification of event structures within sentences (McClosky, 2011) and assist higher-level NLP tasks, like Question Answering (QA) (Huang, 2009). It is possible to detect semantic keywords in NL queries using methods like Named Entity Recognition (NER) or Semantic Role Labelling (SRL) (Collobert, 2011). Both techniques provide a higher level of abstraction than the basic syntax tree. However, our task goes a step further: we aim to find out how these keywords are semantically connected in terms of a network. This is very difficult to achieve using NER alone, since detecting the named entities provides limited information about their relations. SRL does a better job at concept level parsing using predicate logic, but is bound by …

Content truncated to 500 words