Automatic analysis of real dialogues and generating of training corpora

The development of computerized information retrieval dialogue systems communicating with the user in natural language requires the implementation of an effective training procedure with the aid of which the main modules of the dialogue system can be partly automatically developed. The presented paper describes an attempt to create the generating sentence templates automatically, using a special program package implementing an especially developed method of a quantitative linguistic analysis of transcribed real dialogues. Firstly, the program package generates a set of templates, i. e. formulas consisting of elements (terminal and non-terminal symbols) of a special grammar and describing the syntactic structure of a required subset of Czech sentences. Secondly, it generates a large corpus of unique training sentences using the sentence templates and a stochastic context-free grammar. The experimentally created corpus was used for the training of modules of a city information dialogue system.