Text Classification into Abstract Classes Based on Discourse Structure

Proceedings of Recent Advances in Natural Language Processing, pages 200–207, Hissar, Bulgaria, Sep 7–9 2015. Text Classification into Abstract Classes Based on Discourse Structure Boris Galitsky Knowledge Trail Inc. San-Francisco, USA bgalitsky@hotmail.com Dmitry Ilvovsky National Research Universi- ty Higher School of Eco- nomics, Moscow, Russia dilvovsky@hse.ru Sergey O. Kuznetsov National Research Universi- ty Higher School of Eco- nomics, Moscow, Russia skuznetsov@hse.ru Abstract The problem of classifying text with respect to belonging to a document or a meta-document is formulated and its application areas are pro- posed. An algorithm is proposed for document classification tasks where counts of words is insufficient do differentiate between such ab- stract classes of text as metalanguage and ob- ject-level. We extend the parse tree kernel method from the level of individual sentences towards the level of paragraphs, based on anaphora, rhetoric structure relations and communicative actions linking phrases in dif- ferent sentences. Tree kernel learning tech- nique is applied to these extended trees to lev- erage of additional discourse-related infor- mation. We evaluate our approach in the do- main of action-plan documents. 1 Introduction Solving text classification problems, keywords and their topicality usually suffice. These fea- tures provide abundant information to determine a topic of a text or document, such as apple vs banana, or adventures vs relaxing travel. At the same time, there is a number of document classi- fication domains where distinct classes have sim- ilar words. In this case, style, phrasings and other kinds of text structure information need to be leveraged. To perform text classification in such domains, one needs to employ discourse infor- mation such as anaphora, rhetoric structure, enti- ty synonymy and ontology, if available (Wu et al., 2011). In this study, an issue of classifying a text with respect to being metalanguage or language object is addressed. We are concerned with differentiat- ing between object-level documents, which in- form us on how to do things, or how something has been done, and meta-documents, specifying how to write a document which explains how to do things, or how things have been done. Meta- language is a symbolic system intended to ex- press information, or analyze another language or symbolic system. In a natural language docu- ment, metalanguage is used as a special expres- sive means to ascend to the desired level of ab- straction. To automatically recognize metalan- guage patterns in text, one needs some implicit signals at the syntactic level. Naturally, just us- ing keyword statistics is insufficient to differen- tiate between texts in metalanguage and lan- guage-object. A presence of verbs for speech acts and men- tal states (such as knowing) may help to identify metalanguage patterns, but is an unreliable crite- rion: I know the location of the highest mountain vs I know what he thinks about the highest moun- tain in the world. The latter sentence contains a meta-predicate think (who, about-what) with the second variable ranging over a set of (object- level) expressions for thoughts about the highest mountain. Relying on syntactic parse trees …

Content truncated to 500 words