IJCSNS International Journal of Computer Science and Network Security, VOL.8 No.12, December 2008 253 Manuscript received December 5, 2008 Manuscript revised December 20, 2008 Classification-and-Ranking Architecture for Response Generation based on Intentions Aida Mustapha, Md. Nasir Sulaiman, Ramlan Mahmod and Hasan Selamat University Putra Malaysia, Serdang, Selangor Darul Ehsan, Malaysia Summary Grammar-based natural language generation is lacking robustness in implementation because it is virtually incapable for learning. Statistical generation through language models is expensive due to overgeneration and its bias to short strings. Because dialogue utterances render intentions, learning model for the response generation systems should consider all utterances as equally good regardless of length or grammar. An intention-based architecture has been developed to generate response utterances in dialogue systems. This architecture is called classification-and-raking. In this architecture, response is deliberately chosen from dialogue corpus rather than wholly generated, such that it allows short ungrammatical utterances as long as they satisfy the intended meaning of input utterance. The proposed architecture is tested on 64 mixed-initiative, transaction dialogue corpus in theater domain. The results from the comparative experiment show 91.2% recognition accuracy in classification-and-ranking as opposed to an average of 68.6% accuracy in overgeneration-and-ranking. Keywords: Intentions, Speech Acts, Dialogue System, Natural Language Generation, Classification-and-Ranking. 1. Introduction In human-human conversation, dialogue is mutually structured and timely negotiated between dialogue participants. Speakers take turns when they interact, interrupt each other, but their speeches seldom overlap. Similarly, human-machine conversation using dialogue systems exhibits comparable qualities. A response generation system is the natural language generation component in dialogue systems, which is responsible for providing dialogue responses as part of interactive human- machine conversation. The high degree requirement of linguistic input specifications in grammar-based natural language generation is the classic problem of knowledge engineering bottleneck. Statistical generation through language models, although robust, is expensive because alternative realizations and their probabilities have to be calculated individually. Language models also have built- in bias to short strings because likelihood of a string of words is determined by joint probability of words. This is not desirable for generation in dialogues because utterances render intentions; hence all realizations should be treated as equally good regardless of length, in fact, regardless of grammar. The main focus of this paper is to propose a new architecture for response generations based on intentions. The remainder of this paper will be organized as follows. Section 2 will begin with discussion of related works in natural language generation. Section 3 will introduce intentions; the basic building block to our response generation architecture. Section 4 will present the two- staged classification-and-ranking architecture while Section 5 will present validation experiments to compare the proposed architecture with the existing overgeneration-and-ranking architecture. Finally, in section 6 we will draw some concluding remarks. 2. Related Works Existing architectures for natural language generation in dialogue systems mainly concern with generation of words into sentences, either by means of grammar or some statistical distribution. Grammar-based approach requires specification of fine-grained meaning representations as input to guide the generation process. Because this process expects a large number of explicit features, it leads to knowledge acquisition bottleneck in both constructing and maintaining the hand-crafted rule systems [1]. To alleviate the knowledge engineering load in grammar-based approach, the statistical approach of overgeneration-and- ranking architecture [1-6] provides the necessary linguistic decisions through statistical models trained on corpus to furnish semantically related utterances. Overgeneration-and-ranking architecture combines rule- based overgeneration with ranking based on statistical or language models. The principle objective is to help