Statistical natural language generation for dialogue systems based on hierarchical models

Abstract

Due to the increasing presence of natural-language interfaces in our life, natural language processing (NLP) is currently gaining more popularity every year. However, until recently, the main part of the research activity in this area was aimed to Natural Language Understanding (NLU), which is responsible for extracting meanings from natural language input. This is explained by a wider number of practical applications of NLU such as machine translation, etc., whereas Natural Language Generation is mainly used for providing output interfaces, which was considered more as a user interface problem rather than a functionality issue. Generally speaking, natural language generation (NLG) is the process of generating text from a semantic representation, which can be expressed in many different forms. The common application of NLG takes part in so called Spoken Dialogue System (SDS), where user interacts directly by voice with a computer- based system to receive information or perform a certain type of actions as, for example, buying a plane ticket or booking a table in a restaurant. Dialogue systems represent one of the most interesting applications within the field of speech technologies. Usually the NLG part in this kind of systems was provided by templates, only filling canned gaps with requested information. But nowadays, since SDS are increasing its complexity, more advanced and user-friendly interfaces should be provided, thereby creating a need for a more refined and adaptive approach. One of the solutions to be considered are the NLG models based on statistical frameworks, where the system’s response to user is generated in real-time, adjusting their response to the user performance, instead of just choosing a pertinent template. Due to the corpus-based approach, these systems are easy to adapt to the different tasks in a range of informational domain. The aim of this work is to present a statistical approach to the problem of utterance generation, which uses cooperation between two different language models (LM) in order to enhance the efficiency of NLG module. In the higher level, a class- based language model is used to build the syntactic structure of the sentence. Inthe second layer, a specific language model acts inside each class, dealing with the words. In the dialogue system described in this work, a user asks for an information regarding to a bus schedule, route schemes, fares and special information. Therefore in each dialogue the user has a specific dialogue goal, which needs to be met by the system. This could be used as one of the methods to measure the system performance, as well as the appropriate utterance generation and average dialogue length, which is important when speaking about an interactive information system. The work is organized as follows. In Section 2 the basic approaches to the NLG task are described, and their advantages and disadvantages are considered. Section 3 presents the objective of this work. In Section 4 the basic model and its novelty is explained. In Section 5 the details of the task features and the corpora employed are presented. Section 6 contains the experiments results and its explanation, as well as the evaluation of the obtained results. The Section 7 resumes the conclusions and the future investigation proposals.