This paper describes the response planning and generation cneration of the MERCURY flight reservation system, a mixed-initiative spoken dialogue system that supports bothvoicP5z)u interac)u1 and multimodalinterac; ;u augmenting spoken inputs with typing orcuP;5P) at a displayed Web page. MERCURY iscu;P0;0u using the Galaxy CommunicB2u arcunicB2u (Sene#, Hurley, Lau,Sc,u;N & Zue, 1998), where a suite of servers interac via program cogram mediated by ac:)B0P hub. Language generation is performed in two steps: response planning, ordeep-struc))u generation, iscu;B)P out by the dialogue manager, and is well-integrated with otheraspec; of dialogue calogue calogu flow isspecz5B by a dialogue control table (Sene# & Polifroni, 2000a). Response generation, or surfacu1;BN generation, isexec552 by a separate language generation server, under theguidanc of a set of recBNN)u generation rules and anassoc5u1; lexic (Baptist & Sene#, 2000). The generation of the textual string for thegraphic: interfac and the marked-up synthesis string for spoken outputs arecuP;P25u1 by a shared set of generation rules (Sene# & Polifroni, 2000b). Thus there is adirec meaning-to-speec mapping that eliminates the need to analyzelinguistic strucist for synthesis. To date, we havecveuP;N) over 25 000 utteranc1 from usersinterac5:u with the MERCURY system. We report here on both the results of usersatisfacu1: studies cudiesuP by the National Institute of Standards andTec::)2u1 (NIST), and on our own tabulation of a number of di#erent measures of dialogue success.

Statistical NLG has largely meant n-gram modelling which has the considerable advantages of lending robustness to NLG systems, and of making automatic adaptation to new domains from raw corpora possible. On the downside, n-gram models are expensive to use as selection mechanisms and have a built-in bias towards shorter realisations. This paper looks at treebank-training of generators, an alternative method for building statistical models for NLG from raw corpora, and two different ways of using treebank-trained models during generation. Results show that the treebank-trained generators achieve improvements similar to a 2-gram generator over a baseline of random selection. However, the treebank-trained generators achieve this at a much lower cost than the 2-gram generator, and without its strong preference for shorter realisations. 1

Recent advances in Automatic Speech Recognition technology have put the goal of naturally sounding dialog systems within reach. However, the improved speech recognition has brought to light a new problem: as dialog systems understand more of what the user tells them, they need to be more sophisticated at responding to the user. The issue of system response to users has been extensively studied by the natural language generation community, though rarely in the context of dialog systems. We show how research in generation can be adapted to dialog systems, and how the high cost of hand-crafting knowledge-based generation systems can be overcome by employing machine learning techniques.

After many successes, statistical approaches that have been popular in the parsing community are now making headway into Natural Language Generation (NLG). These systems are aimed mainly at surface realization, and promise the same advantages that make statistics valuable for parsing: robustness, wide coverage and domain independence. A recent experiment aimed to empirically verify the linguistic coverage for such a statistical surface realization component by generating transformed sentences from the Penn TreeBank corpus. This article presents the empirical results of a similar experiment to evaluate the coverage of a purely symbolic surface realizer. We present the problems facing a symbolic approach on the same task, describe the results of its evaluation, and contrast them with the results of the statistical method to help quantitatively determine the level of coverage currently obtained by NLG surface realizers. 1

A module for generation of referring expressions (GRE) derives descriptions that identify specified entities in context. In the implemented system I describe here for specifying simple cases of GRE by example, system-builders pair entities with descriptions of them that would be satisfactory for a system to use in context. Automatic methods then construct a suitable knowledge base and context set for a knowledge-based GRE module for the system. These resources will always account for the sample descriptions the designer has supplied, but can also generalize to other possible referring expressions in other possible contexts in the application. I discuss the results in the perspective of knowledge-acquisition methodology for NLG for dialogue, draw contrasts with other uses of examples in NL technology, and use the results to argue for constrained models of the generation process founded on declarative links between resources and generator output.

Abstract. This paper describes the use of a statistical structural N-gram model in the natural language generation component of a Spanish-English generationheavy hybrid machine translation system. A structural N-gram model captures the relationship between words in a dependency representation without taking into account the overall structure at the phrase level. The model is used together with other components in the system for lexical and structural selection. An evaluation of the machine translation system shows that the use of structural N-grams decreases runtime by 60 % with no loss in translation quality. 1

Abstract not found

This paper presents Segue, a hybrid surface natural language generator that employs case-based paradigm but performs rulebased adaptations. It uses an annotated corpus as its knowledge source and employs grammatical rules to construct new sentences. By using adaptation-guided retrieval to select cases that can be adapted easily to the desired output, Segue simplifies the process and avoids generating ungrammatical sentences. The evaluation results show the system generates grammatically correct sentences (91%), but disfluency is still an issue.

In this paper, we discuss techniques to combine an interlingua translation framework with phrase-based statistical methods, for translation from Chinese into English. Our goal is to achieve high-quality translation, suitable for use in language tutoring applications. We explore these ideas in the context of a flight domain, for which we have a large corpus of English queries, obtained from users interacting with a dialogue system. Our techniques exploit a pre-existing English-to-Chinese translation system to automatically produce a synthetic bilingual corpus. Several experiments were conducted combining linguistic and statistical methods, and manual evaluation was conducted for a set of 460 Chinese sentences. The best performance achieved an “adequate ” or better analysis (3 or above rating) on nearly 94% of the 409 parsable subset. Using a Rover scheme to combine four systems resulted in an “adequate or better ” rating for 88% of all the utterances.

The main focus of this thesis is on improving the quality of concatenative speech synthesis by taking advantage of the natural (allowable) variability in spoken language, namely, the fact that there are multiple ways of uttering a given sentence and there are several word sequences that can represent a given concept. An architecture for speech generation for constrained domain applications is proposed that tightly integrates language generation and speech synthesis, allowing the choice of words and desired intonation in the system's response to be optimized jointly with the speech output quality. Experiments with a travel planning dialog system have demonstrated that by expanding the space of candidate responses and possible prosodic realizations we achieve higher quality speech output.