C-HTS: A Concept-based Hierarchical Text Segmentation approach

C-HTS: A Concept-based Hierarchical Text Segmentation approach Mostafa Bayomi, Séamus Lawless ADAPT Centre, Knowledge and Data Engineering Group, School of Computer Science and Statistics, Trinity College Dublin, Dublin, Ireland bayomim@tcd.ie, seamus.lawless@scss.tcd.ie Abstract Hierarchical Text Segmentation is the task of building a hierarchical structure out of text to reflect its sub-topic hierarchy. Current text segmentation approaches are based upon using lexical and/or syntactic similarity to identify the coherent segments of text. However, the relationship between segments may be semantic, rather than lexical or syntactic. In this paper we propose C-HTS, a Concept-based Hierarchical Text Segmentation approach that uses the semantic relatedness between text constituents. In this approach, we use the explicit semantic representation of text, a method that replaces keyword-based text representation with concept-based features, automatically extracted from massive human knowledge repositories such as Wikipedia. C-HTS represents the meaning of a piece of text as a weighted vector of knowledge concepts, in order to reason about text. We evaluate the performance of C-HTS on two publicly available datasets. The results show that C-HTS compares favourably with previous state-of-the-art approaches. As Wikipedia is continuously growing, we measured the impact of its growth on segmentation performance. We used three different snapshots of Wikipedia from different years in order to achieve this. The experimental results show that an increase in the size of the knowledge base leads, on average, to greater improvements in hierarchical text segmentation. Keywords: Hierarchical Text Segmentation, Explicit Semantic Analysis, Semantic Relatedness, Wikipedia 1. Introduction Text segmentation aims to divide text into coherent segments which reflect the sub-topic structure of the text. It is widely used as a pre-processing task for Information Retrieval (Prince and Labadié, 2007) and several Natural Language Processing (NLP) tasks such as Text Summarization (Boguraev and Neff, 2000) and Question Answering (Tellex et al., 2003). Text segmentation is essentially based on measuring the coherence between atomic units of text (sentences or paragraphs). Some approaches use lexical similarity, which compares units of text that are represented as vectors of the same, or similar, words (Choi, 2000). These approaches are limited, in that they rely on endogenous knowledge extracted from the documents themselves. Relying on such knowledge does not reveal much about the meaning beyond the text. Some approaches started to enrich the text representation by exploiting its semantic meaning by using the Latent Semantic Analysis (LSA) (Choi et al., 2001). However, these approaches require a very large corpus, and consequently the pre-processing effort required is significant. On the other hand, some other approaches started to use external resources such as WordNet to enrich text (Stokes et al., 2004). However, such resources cover only a small fragment of the language lexicon. Furthermore, the use of such lexical resources offers little knowledge about the different word representations. Buchanan and Feigenbaum (1982) stated that: “The power of an intelligent program to perform its task well depends primarily on the quantity and quality of knowledge it has about that task.” Hence, in this research, we are trying to enrich the text representation …

Content truncated to 500 words