Generated Narratives For Computer-Aided Language Teaching

VINCI is a Natural Language Generation environment designed for use in computer-aided second language instruction. It dynamically generates multiple parallel trees representing an initial text, questions on this text, and expected answers, and either orthographic or phonetic output. Analyses of a learner’s answers to questions are used to diagnose comprehension and language skills and to adaptively control subsequent generation. The paper traces stages in the generation of short texts in English and French, and discusses issues of architecture, textual enrichment, and planning. 1 VINCI: Architecture, Implementation and Output Ideally, a language teaching system should both “encourage the creative use of language in communicatively relevant settings” (Menzel and Schroeder, 1998) and also provide detailed and adaptive feedback (cf. (Michaud, 2002)). Many systems resolve the issue by means of complex parsing. In what follows, we describe VINCI, a multilingual generation environment which represents a complementary approach in that it assumes conversational control, dynamically producing more or less complex texts and asking information of users. VINCI is based on a collection of metalanguages which define the semantics, syntax, lexicon and morphology of a language. Text files defining the language are read by an interpreter (written in C) and output is presented either orthographically or phonetically. When used in a teaching context, VINCI creates an utterance and presents it to a learner. It also creates a series of questions based on each utterance, together with some hidden answers. The learner is prompted to respond to each question, and his or her response is compared with the hidden one (or ones) and a detailed report is produced on the relation between the two. This report provides information on a learner’s comprehension and language skills, as well as guidance for subsequent generation. VINCI is capable of generation in many languages, and has been tested on such diverse languages as Spanish, Italian, Russian and Chinese. Our work to date has been carried out in both English and French, predominantly the latter. For the generation of simple utterances, VINCI constructs a syntax tree using contextfree rules and syntactic transformations. The nodes of the tree may be decorated with attributes, whose role is to maintain grammatical and perhaps semantic agreement by restricting the choice of lexical items and controlling morphology. Once a tree is formed, its leaves are assigned suitable lexical choices which are inflected by morphology rules. In the most basic situation, choices among syntactic alternatives and among possible lexical entries are made at random. In such a model, the semantic control exercised by attributes is minimal. They can, for example, ensure that the subject of a verb such as eat is animate and its object edible, but they cannot influence the overall meaning of the utterance produced. To achieve this, VINCI makes use of preselections.