A Comparative Study of Text Preprocessing Approaches for Topic Detection of User Utterances

A Comparative Study of Text Preprocessing Approaches for Topic Detection of User Utterances Roman Sergienko, Muhammad Shan and Wolfgang Minker Institute of Telecommunication Engeneering, Ulm University Albert-Einstein-Allee 43, 89081, Ulm, Germany {roman.sergienko, muhammad.shan, wolfgang.minker}@uni-ulm.de Abstract The paper describes a comparative study of existing and novel text preprocessing and classification techniques for domain detection of user utterances. Two corpora are considered. The first one contains customer calls to a call centre for further call routing; the second one contains answers of call centre employees with different kinds of customer orientation behaviour. Seven different unsupervised and supervised term weighting methods were applied. The collective use of term weighting methods is proposed for classification effectiveness improvement. Four different dimensionality reduction methods were applied: stop-words filtering with stemming, feature selection based on term weights, feature transformation based on term clustering, and a novel feature transformation method based on terms belonging to classes. As classification algorithms we used k-NN and a SVM-based algorithm. The numerical experiments have shown that the simultaneous use of the novel proposed approaches (collectives of term weighting methods and the novel feature transformation method) allows reaching the high classification results with very small number of features. Keywords: text classification, spoken language, term weighting 1. Introduction An example of a spoken language processing application is spoken dialogue systems (SDS). A standard SDS consists of subsequent processing steps including (Minker and Ben- nacef, 2004): speech recognition, semantic text analysis, dialogue management, text generation, and speech synthe- sis. The specific semantic model is designed for the problem domain of the SDS. Nowadays, multi-domain spoken di- alogue systems constitute a breakthrough in the area of human-machine interfaces (Lee et al., 2009; Komatani et al., 2009). Multi-domain SDS include different specific se- mantic models for each problem domain. Prior to a de- tailed semantic analysis the general topic domain of user utterance needs to be determined. Domain detection can be formulated as a text classification problem based on a ”bag-of-words” model. Designing domain-related seman- tic models after a domain detection step based on text clas- sification may be more effective than designing a complex semantic model that applies for all domains. Before text classification, it is necessary to perform text preprocessing which comprises three stages. The first one is the textual feature extraction based on raw document pre- processing. This includes procedures such as stop-words filtering (Fox, 1989) and stemming (Porter, 2001). The second stage is the numerical feature extraction based on term weighting. The most well-known unsupervised term weighting method is TF-IDF (Salton and Buckley, 1988). The following supervised term weighting methods are also considered in our paper: Gain Ratio (GR) (Debole and Sebastiani, 2004), Confident Weights (CW) (Soucy and Mineau, 2005), Term Second Moment (TM2) (Xu and Li, 2007), Relevance Frequency (RF) (Lan et al., 2009), Term Relevance Ratio (TRR) (Ko, 2012), and Novel Term Weighting (NTW) (Gasanova et al., 2014a); these methods involve information about the classes of the documents. In this paper we propose collectives of term weighting meth- ods that could …

Content truncated to 500 words