Outlook and conclusion

Organizing thoughts and expressions in such a way that discourse flows is a challenge for everyone. The difficulties are of various sorts. We have to acquire knowledge concerning linguistic resources (lexicon, grammar, textual devices: anaphora, etc.) and learn how our mind interprets the use of each one of them or the combination of conceptual entities (propositions). Once we have acquired this, effects can be used as goals, the normal starting point in communication.

A lot of work has been done on language production, yet, most of it from an engineering perspective [REI 00, BAT 16], i.e. fully automated text generation. Unfortunately, next to nothing has been done to assist the writer (our goal), though there is clearly a need for it. Note that the conditions under which an author works are far more multifarious and open-ended than the ones computers have to struggle with [AND 96]. Hence, instead of presenting a fully automatic solution as most computational linguists do, we propose here an interactive one. More precisely, we try to build an authoring tool where the computer helps the writer to organize his/her thoughts. Given a set of messages (user’s knowledge state) and possibly a goal, the system structures the data in such a way that, conceptually speaking, the output is coherent, yielding one or several conceptual trees. Since this is quite a complex task, we have started with a very simple set of data. Nevertheless, we have tried to go beyond the obvious (co-references), to reveal some of the hidden and distributed information.

Our next steps will involve dealing with other conceptual relations and with the labeling of topical categories. We would also like to explore the idea of the information associated with the seed words. This should allow us to see on what basis (origin of the information on which) topical clustering is performed. As data can be grouped in many ways (depending on the chosen criteria or view points), and as various orders are likely to yield different effects, it would be nice to show the relationship between ordering and rhetorical effects.

Concerning the method, we could have also considered the following strategy: take a set of well-written texts of a specific type (description), extract its sentences, normalize and scramble them and have the computer try to reorganize them to produce a coherent whole, matching, as well as possible, the initial document (gold standard). Yet, having thought about this strategy too late (see, e.g., [BAR 08, BOL 10, LAP 06, WAN 06]), we used a different approach.