Hierarchical Topic Structuring: From Dense Segmentation to Topically Focused Fragments via Burst Analysis

Proceedings of Recent Advances in Natural Language Processing, pages 588–595, Hissar, Bulgaria, Sep 7–9 2015. Hierarchical Topic Structuring: From Dense Segmentation to Topically Focused Fragments via Burst Analysis Anca-Roxana Simon Universit´e de Rennes 1 IRISA & INRIA Rennes Pascale S´ebillot INSA de Rennes IRISA & INRIA Rennes anca-roxana.simon@irisa.fr pascale.sebillot@irisa.fr guig@irisa.fr Guillaume Gravier CNRS IRISA & INRIA Rennes Abstract Topic segmentation traditionally relies on lexical cohesion measured through word re-occurrences to output a dense segmen- tation, either linear or hierarchical. In this paper, a novel organization of the topical structure of textual content is proposed. Rather than searching for topic shifts to yield dense segmentation, we propose an algorithm to extract topically focused frag- ments organized in a hierarchical manner. This is achieved by leveraging the tem- poral distribution of word re-occurrences, searching for bursts, to skirt the limits im- posed by a global counting of lexical re- occurrences within segments. Comparison to a reference dense segmentation on var- ied datasets indicates that we can achieve a better topic focus while retrieving all of the important aspects of a text. 1 Introduction Being aware of the topical structure of texts or automatic transcripts is known to be helpful for multiple natural language processing tasks such as summarization, question answering, etc. Vari- ous solutions have emerged to obtain such a struc- ture, the most interesting ones being generic so- lutions that can be applied on any kind of textual data. These generic solutions are generally based on lexical cohesion, i.e., on identifying segments with a consistent use of vocabulary, in particular measured via word re-occurrences. Their output is a dense segmentation, i.e., contiguous segments, most of the time linear even if the structure of dis- course is known to have a hierarchical form (Grosz and Sidner, 1986; Marcu, 2000). Dense segmentation, linear or hierarchical, is however not necessarily appropriate to reflect the fact that some fragments of the data bear im- portant ideas while others are simple fillers, i.e., they do not bring additional important informa- tion. This notion of irrelevant ideas was also mentioned in (Choi, 2000) where the author notes that skipping irrelevant fragments improves nav- igation. In addition, lexical re-occurrence is not sufficient for this type of segmentation, as we will demonstrate. In particular in the hierarchical case, segments get smaller as we go towards fine grain segmentation: As a consequence, there is a re- duced number of words per segment and neigh- boring segments might refer to the same general topic and hence exhibit high lexical coherence. To skirt these limits, we investigate a different way of organizing the topical structure of textual content. We rely on the fact that some words ap- pear in bursts, i.e., with a frequency higher than normal at specific locations in the text. The key idea that we leverage is that the presence of lexi- cal bursts usually indicates a strong topical focus, as we will highlight. As an alternative to dense hierarchical topic segmentation, we propose to de- rive a hierarchy …

Content truncated to 500 words