An I-vector Based Approach to Compact Multi-Granularity Topic Spaces Representation of Textual Documents

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 443–454, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics An I-vector Based Approach to Compact Multi-Granularity Topic Spaces Representation of Textual Documents Mohamed Morchid†, Mohamed Bouallegue†, Richard Dufour†, Georges Linar`es†, Driss Matrouf† and Renato de Mori†‡ †LIA, University of Avignon, France ‡McGill University, School of Computer Science, Montreal, Quebec, Canada {firstname.lastname}@univ-avignon.fr rdemori@cs.mcgill.ca Abstract Various studies highlighted that topic- based approaches give a powerful spo- ken content representation of documents. Nonetheless, these documents may con- tain more than one main theme, and their automatic transcription inevitably contains errors. In this study, we propose an orig- inal and promising framework based on a compact representation of a textual docu- ment, to solve issues related to topic space granularity. Firstly, various topic spaces are estimated with different numbers of classes from a Latent Dirichlet Allocation. Then, this multiple topic space representa- tion is compacted into an elementary seg- ment, called c-vector, originally developed in the context of speaker recognition. Ex- periments are conducted on the DECODA corpus of conversations. Results show the effectiveness of the proposed multi-view compact representation paradigm. Our identification system reaches an accuracy of 85%, with a significant gain of 9 points compared to the baseline (best single topic space configuration). 1 Introduction Automatic Speech Recognition (ASR) systems frequently fail on noisy conditions and high Word Error Rates (WER) make the analysis of the au- tomatic transcriptions difficult. Speech analyt- ics suffer from these transcription issues that may be overcome by improving the ASR robustness and/or the tolerance of speech analytic systems to ASR errors. This paper proposes a new method to improve the robustness of speech analytics by combining a semantic multi-model approach and a noise reduction technique based on the i-vector paradigm. This method is evaluated in the application framework of the RATP call centre (Paris Public Transportation Authority), focusing on the theme identification task (Bechet et al., 2012). Telephone conversations are a particular case of human-human interaction whose automatic processing raises problems, especially due to the speech recognition step required to obtain the transcription of the speech contents. First, the speaker’s behavior may be unexpected and the training/test mismatch may be very large. Second, the speech signal may be strongly impacted by various sources of variability: environment and channel noises, acquisition devices, etc. Telephone conversation issues Topics are related to the reason why the customer called. Various classes corresponding to the main customer’s requests are considered (lost and founds, traffic state, timelines, etc). In addition to classical issues in such adverse conditions, the topic-identification system should deal with problems related to class proximity. For example, a lost & found request is related to itinerary (where was the object lost?) or timeline (when?), that could appear in most of the classes. In fact, these conversations involve a relatively small set of basic concepts related to transportation issues. Figure 1 shows an example of a dialogue which is manually labeled by the agent …

Content truncated to 500 words