Section: New Results

Modelling and extracting discourse structures

Participants : Laurence Danlos, Charlotte Roze.

Cross-lingual lexical semantics of discourse connectives

Discourse connectives are words or phrases that indicate senses holding between two spans of text. The theoretical approaches accounting for these senses, such as text coherence, cohesion, or rhetorical structure theory, share at least one common feature: they acknowledge that many connectives can indicate different senses depending on their context. Depending on its sense, the translation of a connective into another language can vary greatly, either using an equivalent connective, or using a different construction or even no explicit connective at all .

On the basis of data provided by the bilingual concordancer TransSearch which propose statistical word alignment [64] , [53] made a semi-manual annotation of the English translation of two French connectives ("en effet" and "alors que"). The results of this annotation show that the translations of these connectives do not correspond to the “transpots” identified by TransSearch and even less to the translations proposed in bilingual dictionaries.

The conclusions of this work were presented at an European workshop organized by the project COMTIS(http://www.idiap.ch/project/comtis ), and some members decide to use our technic for other connectives and other aligned corpora (e.g. Europarl).

Discourse relations inference rules

In 2011 we have developed a new methodology for building discourse relations inference rules, to be integrated into an algebra of these relations [54] , [38] . The construction of such an algebra has as main objective the improvement of the comparison of discourse structures within the evaluation of discourse annotations and the creation of a gold-standard corpus. The inference rules can also help detecting inconsistencies in discourse structures, in order to improve human or machine annotation. The premises of rules already studied lead to the formulation of inference rules, established by the theoretical definition of discourse relations, manually constructed data and extracted data. By manually annotating discourses, we also compute inference probabilities. We have illustrated the adopted methodology taking as theoretical background the Segmented Discourse Representation Theory [60] .

Discourse structure and factivity

Discursive annotations proposed in theories of discourse such as RST (Rhetorical Structure Theory) or SDRT (Segmented Representation Theory Dicourse) have the advatange of building a global discourse structure linking all the information in a text. Discursive annotations proposed in PDTB (Penn Discourse Tree Bank) have the advatange of identifying the "source" of each information – thereby answering to questions such as who says or thinks what?

In collaboration with Owen Rambow (Columbia University), we have proposed [26] , [28] a unified approach for discursive annotations combining the strengths of these two streams of research. This unified approach relies crucially on factivity information, as encoded in the English corpus FactBank. We intend to pursue this avenue of research by initiating in 2012 the development of a French FactBank.