A Semantic Analysis of Instructional Texts

Texts as well as dialogues originated a number of analysis at discourse level from various perspectives, such as: modelling of nominal, temporal or spatial reference resolution [1, 5], of rhetorical structure [6], argumentative structure, and of cooperative discourse structure [3]. So far, little has been done to formally represent the structure of instructional (also called procedural) texts, besides [4, 6]. However, from an applied perspective, the study of instructional text structure is most useful for answering questions in How to ?. While research in questionanswering (QA) mainly focussed on responding to factoid-like questions using stochastic processes, a more recent trend in QA aims at responding to other types of questions of much importance in everyday life or in professional environments such as procedural, causal, comparative, evaluative or opinion questions. These have not yet been studied in depth, they require different types of methodologies and formalisms, particularly at the level of linguistic models, knowledge representation and reasoning procedures. The overall structure of instructional texts can be described on three levels with their internal coherence and independence, however with numerous interactions. The first level, that we call (the task structure), is mainly analyzed as an ordered sequence of instructions operating over a specific set of entities in order to reach a goal. Instructions may be sequential, or may have a more complex structure including, for example, options, alternatives or operations to realize in parallel with others. This level also includes the analysis of markers that connect instructions. A more refined analysis of instructional texts, including some forms of coherence checking, is then accounted for by our second level based on the rhetorical structure (RST). Finally, a third level deals with a more refined instruction typology that characterizes their facets: imperative character (including forms of illocutionary force), modalities and manners, internal motivations, etc. In this short paper, we briefly investigate the two first levels. The work presented here is integrated into COOPML, which is a generic framework for representing the structure of almost any type of cooperative QA discourse [3].