The proposed project aims to develop a natural language generation system designed for easy and broad adaptability using statistical methods. First, the system will be able to adapt the structure and lexical choice of generated sentences to the communication goal and to the particular user (e.g., by aligning the vocabulary choice to the expressions uttered by the user). In comparison to current systems, which are typically based on hand-written rules or templates, this will render a significant improvement in the naturalness of the output. Second, the system aims to provide an easier domain and language adaptability in comparison to the current systems – the possibility to retrain most parts of the generator from data. The usage of statistical methods and trainable modules has recently been spreading also in
the area of natural language generation; however, mainly in spoken dialogue systems, it remains limited to the initial steps of utterance generation for the most part.
The resulting system will be used to generate utterances in the spoken dialogue system currently under development within the VYSTADIAL grant project at the Institute of Formal and Applied Linguistics.