Section: New Results

Advances in symbolic and hybrid parsing with DyALog and FRMG

Participants : Éric Villemonte de La Clergerie, François Barthélemy, Julien Martin.

Within the team is developed a wide-coverage French meta-grammar (FRMG) and a efficient hybrid TAG/TIG parser based on the DyALog logic programming environment [120] and on the Lefff morphological and syntactic lexicon [105] . It relies on the notion of factorized grammar, themselves generated from a representation that lies at a higher level of abstraction, named Meta-Grammars  [122] . At that level, linguistic generalizations can be expressed, which in turn makes it possible to transfer meta-grammars from one language to a closely related one. The hybrid TAG/TIG parser generator itself implements all kinds of parsing optimizations: lexicalization (in particular via hypertags), left-corner guiding, top/bottom feature analysis, TIG analysis (with multiple adjoining), and others. The recent evolutions go towards an hybridization with statistical approaches.

Tuning FRMG's disambiguation mechanism

Continuing works initiated in 2011 on the exploitation of the dependency version of the French TreeBank (FTB), Éric de La Clergerie has explored the tuning of FRMG's rule base disambiguation mechanism using a larger set of features and weight learned from the FTB. In 2011, this approach led to on improvement from 82.31% to 84.54% in terms of accuracy (LAS - Labelled Attachment Score) on the test part of the FTB. By increasing the set of features, in particularly using higher-order dependency features (on parent edge and sibling edges), and a better understanding of the iterative tuning mechanism, it was possible to reach 85.95% LAS. This tuning mechanism is based on the idea of adding or subtracting some weight to a disambiguation rule given some specific contexts (provided by the features), where the delta is progressively learned from the accuracy of the disambiguation rule in terms of edge selection or rejection. The learning algorithm presents some relationships with the perceptron approach, but the use of a more standard implementation of the perceptron led to less interesting gains.

During the same time, the coverage of FRMG was improved (to reach for instance 94% of full parses on the FTB).

Synchronous Tree-Adjoining Grammars

A preliminary work has been done to implement Synchronous Tree-Adjoining Grammars (STAGs) in DyALog , relying on the notion of Thread Automata [119] . Synchronous Tree Adjoining Grammars is an instance of formalism where the order of the components of a tree structure is not fully determined. This leads to combinatorial alternatives when parsing, while a tree-structure corresponding to the input string has to be build. A specific front-end has been written to implement STAGs. The work on the back-end is still in progress, with the goal to have a common intermediate representation for several mildly context-sensitive formalisms where some node operations non-deterministically pick a node out of a finite set of nodes. STAGs are an instance of such formalisms, Multi-Component Tree Adjoining Grammars (MCTAGs) are another instance. The intermediate representation consists in Thread Automata (TA), an extension of Push-Down Automata where several threads of computations are considered and only one is active at any time.

Adding weights and probabilities to DyALog

Weights can already be used during the disambiguation phase of the FRMG parser, implemented in DyALog . However, a deeper implementation of weights and probabilities in DyALog was initiated in 2012 by Julien Martin during his Master internship. By enriching the structure of the backpointers (relating the items to their parent items), it is now possible to maintain an ordered weighted list of derivations, to update the scheduling of items wrt their weight, to update the weights of all the descendants of an item $I$ when updating $I$'s weight. The motivation is of course to be able to favor the best analysis first during parsing. A second objective (which has been implemented) is the possibility to extract the $n$-best parses after parsing (but keeping a shared derivation forest). A third objective, remaining to be done, is related to the use of beam search techniques to prune the search space during parsing. A longer-term objective is the abstraction of this work to be able to work on semi-rings.