Section: New Results

Discourse Dynamics

Participants : Maxime Amblard, Timothée Bernard, Clément Beysson, Maria Boritchev, Philippe de Groote, Bruno Guillaume.

Dynamic Generalized Quantifiers

We have started a classification of the (French) determiners according to the dynamic properties of the generalized quantifiers they denote [12], [17].

Following Groenendijk and Stokhof [43], we say that a generalized quantifier is internally dynamic in case the dynamic binders occurring in its restriction have the capacity of binding material that occurs in their scopes. We also say that a generalized quantifier is externally dynamic in case the dynamic binders occurring in both its arguments have the capacity of binding material that occur in the continuation of the discourse. In addition to these notions of internal and external dynamicity, we consider a third notion that we call intrinsic dynamicity. We say that a generalized quantifier is intrinsically dynamic in case it introduces new referent markers and makes them available to the continuation of the discourse.

Using these three notions, we have defined three classes of dynamic generalized quantifiers, which fairly correspond to the notions of specific (e.g., the, this, his), general (e.g., a, some, another), and quantificational determiners (e.g., every, no). We then have shown how the dynamic generalized quantifiers belonging to these three classes may be formalized using the continuation-based approach introduced in [5].

Dialogue Modeling

Studying dialogical interactions is a major subject in natural language processing, since dialogues represent the basis of human communication. Addressing this problem requires relating approaches from fields such as semantics, pragmatics, and, more generally, logic, and cognition. We have presented a compositional dynamic model of questions and answers mechanisms in a dialogical setting. We address dialogical and lexical issues starting from the formal definitions of frame semantics given in [7]. We achieve compositionality and dynamicity in our model by constructing it on top of concepts inherited from Type Theoretical Dynamic Logic [5]. We introduce control in the common (accessible to all participants of a dialogue) context of a conversation by formulating the concept of dialogical context and elaborating corresponding storage operations. We apply our model to real non-controlled examples of dialogical interactions provided by the Schizophrenia and Language, Analysis and Modeling corpus [29]. The linguistic analysis of dialogues between patients with schizophrenia and psychologists has revealed specific language-driven manifestations of cognitive dysfunction. This approach to dialogue modeling in a dynamic framework allowed us to develop tools to handle specifics of dialogical interactions on top of already existing methods for general discourse.

Discourse Structure

A text as a whole must exhibit some coherence that makes it more than just a bag of sentences. This coherence hinges on the discourse relations (DRs). The latter express the articulations between the different pieces of information of the text. There is still debate about the number and the nature of these DRs. Yet, typical DRs include Contrast , Consequence , or Explanation . Using a discourse connective (because, instead, although) is usually the most direct and reliable way to express a DR. These lexical items have specific syntactic, semantic, and pragmatic properties. In particular, one can often observe a mismatch between the arguments of a DR and the (syntactic) ones of the connective lexicalizing it. It happens in configurations in which the argument of the DR does not directly correspond to syntactic argument of the discourse marker. In (1), for instance, the second argument of the Explanation relation is not the whole conditional, its antecedent, nor its consequent. But it is the possibility of the conditional, paraphrasable by she might miss her train. The discourse argument is here presupposed by the conditional (i.e., the syntactic argument).

1. Mary is worried because if there is too much delay, she will miss her train.

2. John did not come to the party although Mary said he was already back in town.

Another common case occurs when an attitude verb (think, believe) or a verb of report (say, tell) is used evidentially as in (2). In such cases, the contrast expressed by the writer holds between John did not come to the party and he was already back in town. The main function of the evidential (Mary said ...) is to introduce the argument of a DR without being itself part of the discourse structure.

Whereas DRs have two arguments, some discourse markers, such as adverbial connectives (so, otherwise), have only one syntactic argument. It then seems natural to use an anaphoric mechanisms to describe how the other argument of the DR they lexicalise is determined from the context. We extended this idea to all connectives and showed how this view can explain most usual cases of mismatch. Additionally, considering that discourse arguments are implicit semantic objects akin to the events introduced in the Davidsonian theory, it is possible to implement this proposal in Type Theoretic Dynamic Logic, without the need of a syntactic parse above the sentence level, and in a strictly compositional way, using continuations.