Section: New Results

Discourse dynamics

Discourse Structure Modeling

It is usually assumed that the internal structure of a text, typically characterized by discourse or rhetorical relations, plays an important role in its overall interpretation. In order to build such a structure, some approaches rely on discourse grammars. The key idea is to consider the structural regularities in discourse structure similarly as syntactic regularities. A particular trend relies on tree grammars. This trend has been further developed by integrating the modeling of both clausal syntax and semantics, and discourse syntax and semantics within the frameworks of Tree-Adjoining Grammar (TAG)  [48] , [49] and TAG for Discourse (D-LTAG)  [79] , [41] , [80] , [42] .

Two important features characterize these approaches. First, while they use a single grammatical formalism, two different grammars are used for syntactic parsing and then for discourse parsing. In addition to adding an intermediate processing step, this two-tiered treatment both complicates the modeling of connectives that are ambiguous in their syntactic and discourse use, and prevents using standard disambiguation techniques. Second, some discourse structures better represented by directed acyclic graphs (DAG) than by trees are not accounted for.

In order to address the second issue of building DAG structures,  [36] , [37] have proposed Discourse Synchronous TAG (D-STAG), a TAG based approach together with a higher-order interpretation of sentences using Synchronous Tree-Adjoining Grammar (STAG)  [67] , [77] .

We developed a method to interface a sentential grammar and a discourse grammar without resorting to an intermediate processing step. The method is general enough to build discourse structures that are DAG and not only trees. Our analysis is based on D-STAG. We also use an encoding of TAG into ACG. This encoding allows us to express a higher-order semantic interpretation that enables building DAG discourse structures on the one hand, and to smoothly integrate the sentential and the discourse grammar thanks to the modular capability of ACG. The results has been published [13] and all the examples of the article have been implemented and may be run and tested with the ACGtk software (see  6.1 ).

Effects and Handlers

We made the argument that pragmatics are to semantics what side effects are to calculations in a programming language. We demonstrated this parallel on two aspects.

First off, both pragmatics and side effects serve the same function. Side effects in programming languages account for the effects of expressions that reach beyond their scope and for the way a language interacts with the world of its users. Pragmatics is concerned with phenomena that also involve the non-immediate effects of expressions (e.g., discourse anaphora, presupposition accommodation) and with the way language interacts with the world of its users. Secondly, we pointed out that very similar formal theories are being used to treat the both of them (i.e. monads and continuations).

Having established this parallel, we then put forward a preliminary proposal of integrating semantics and pragmatics while keeping them separate by assigning effectful computations of truth values as meanings of linguistic expressions. In this way, we can implement the pragmatics at the level of the side effects and then focus on pure semantics at the level of values.