Motivations and Goals

The ability of dealing with odd (i.e. ill-formed or simply partial) sentences is largely exhibited by humans. This allows to rather easily manage unknown words (e.g. proper nouns never encountered before), to tackle odd grammatical constructions, to force the interpretation of illegal syntactic structures (e.g. gaps in the information streams as in remote/telephonic dialogue) as well as the ability of resorting always to partial information during the interpretation of uncomplete (or erroneous) input.

All of the above phenomena are interesting aspects of what has been recently called robustness in NLP processing (Menzel,1997). The modeling of such phenomena within computational devices is thus more than a relevant research area either for linguistic research as well as for the design of real NLP systems. Robustness has been traditionally stressed as a general desirable property of any computational model and system. NLP engineering methods and NLP systems are crucially faced by problems caused by the noise found in the "real" target texts. However, the nature of these problems and their interactions with the different levels of the language analysis process exhibits specific properties that are hardly approached by exisisting software engineering criteria and practice. Moreover, the above research area is also central from a linguistic point of view. Effective models of robusteness, tha are able to fill gaps or to recover from deficiencies against wrong or poor input streams, pose challanges to the expresiveness of any underlying explanatory language theory. Cognitive aspects of robustness are here also playing the role of experimental evidence as well as definitory knowledge.

The interdisciplinary nature of these research theme is even more critical as without a systematic validation within "real" NLP systems no linguistic or psycholinguistic definition of robustness is possible, that is objectively captured and assessed. The success of a recent Special Issue of the Journal of Natural Language Engineering (Cambridge University Press) is a further evidence of the relevance of these problems within the current research trends. ROMAND 2002 is the second of a series of workshop that aims at bringing together researchers working on robust methods in natural language processing. The term natural language is here intended as all possible modalities of human communication and it is not restricted to written or spoken language.

The main goal of the workshop will be to bring together researchers working in fields like artificial intelligence, computational linguistics, language engineering, human-computer interaction, cognitive science who are facing with the problem of feasible and reliable systems implementation. Theoretical aspects of robustness in NLP are welcome as well as engineering and industrial experiences.