Representation and Processing of Composition, Variation and Approximation in Language Resources and Tools - TEL - Thèses en ligne Accéder directement au contenu
Hdr Année : 2014

Representation and Processing of Composition, Variation and Approximation in Language Resources and Tools

Résumé

In my habilitation dissertation, meant to validate my capacity of and maturity for directing research activities, I present a panorama of several topics in computational linguistics, linguistics and computer science. Over the past decade, I was notably concerned with the phenomena of compositionality and variability of linguistic objects. I illustrate the advantages of a compositional approach to the language in the domain of emotion detection and I explain how some linguistic objects, most prominently multi-word expressions, defy the compositionality principles. I demonstrate that the complex properties of MWEs, notably variability, are partially regular and partially idiosyncratic. This fact places the MWEs on the frontiers between different levels of linguistic processing, such as lexicon and syntax. I show the highly heterogeneous nature of MWEs by citing their two existing taxonomies. After an extensive state-of-the art study of MWE description and processing, I summarize Multiflex, a formalism and a tool for lexical high-quality morphosyntactic description of MWUs. It uses a graph-based approach in which the inflection of a MWU is expressed in function of the morphology of its components, and of morphosyntactic transformation patterns. Due to unification the inflection paradigms are represented compactly. Orthographic, inflectional and syntactic variants are treated within the same framework. The proposal is multilingual: it has been tested on six European languages of three different origins (Germanic, Romance and Slavic), I believe that many others can also be successfully covered. Multiflex proves interoperable. It adapts to different morphological language models, token boundary definitions, and underlying modules for the morphology of single words. It has been applied to the creation and enrichment of linguistic resources, as well as to morphosyntactic analysis and generation. It can be integrated into other NLP applications requiring the conflation of different surface realizations of the same concept. Another chapter of my activity concerns named entities, most of which are particular types of MWEs. Their rich semantic load turned them into a hot topic in the NLP community, which is documented in my state-of-the art survey. I present the main assumptions, processes and results issued from large annotation tasks at two levels (for named entities and for coreference), parts of the National Corpus of Polish construction. I have also contributed to the development of both rule-based and probabilistic named entity recognition tools, and to an automated enrichment of Prolexbase, a large multilingual database of proper names, from open sources. With respect to multi-word expressions, named entities and coreference mentions, I pay a special attention to nested structures. This problem sheds new light on the treatment of complex linguistic units in NLP. When these units start being modeled as trees (or, more generally, as acyclic graphs) rather than as flat sequences of tokens, long-distance dependencies, discontinu- ities, overlapping and other frequent linguistic properties become easier to represent. This calls for more complex processing methods which control larger contexts than what usually happens in sequential processing. Thus, both named entity recognition and coreference resolution comes very close to parsing, and named entities or mentions with their nested structures are analogous 3to multi-word expressions with embedded complements. My parallel activity concerns finite-state methods for natural language and XML processing. My main contribution in this field, co-authored with 2 colleagues, is the first full-fledged method for tree-to-language correction, and more precisely for correcting XML documents with respect to a DTD. We have also produced interesting results in incremental finite-state algorithmics, particularly relevant to data evolution contexts such as dynamic vocabularies or user updates. Multilingualism is the leitmotif of my research. I have applied my methods to several natural languages, most importantly to Polish, Serbian, English and French. I have been among the initiators of a highly multilingual European scientific network dedicated to parsing and multi- word expressions. I have used multilingual linguistic data in experimental studies. I believe that it is particularly worthwhile to design NLP solutions taking declension-rich (e.g. Slavic) languages into account, since this leads to more universal solutions, at least as far as nominal constructions (MWUs, NEs, mentions) are concerned. For instance, when Multiflex had been developed with Polish in mind it could be applied as such to French, English, Serbian and Greek. Also, a French-Serbian collaboration led to substantial modifications in morphological modeling in Prolexbase in its early development stages. This allowed for its later application to Polish with very few adaptations of the existing model. Other researchers also stress the advantages of NLP studies on highly inflected languages since their morphology encodes much more syntactic information than is the case e.g. in English. In this dissertation I am also supposed to demonstrate my ability of playing an active role in shaping the scientific landscape, on a local, national and international scale. I describe my: (i) various scientific collaborations and supervision activities, (ii) roles in over 10 regional, national and international projects, (iii) responsibilities in collective bodies such as program and organizing committees of conferences and workshops, PhD juries, and the National University Council (CNU), (iv) activity as an evaluator and a reviewer of European collaborative projects. The issues addressed in this dissertation open interesting scientific perspectives, in which a special impact is put on links among various domains and communities. These perspectives include: (i) integrating fine-grained language data into the linked open data, (ii) deep parsing of multi-word expressions, (iii) modeling multi-word expression identification in a treebank as a tree-to-language correction problem, and (iv) a taxonomy and an experimental benchmark for tree-to-language correction approaches.
Fichier principal
Vignette du fichier
savary-hdr-2013.pdf (2.5 Mo) Télécharger le fichier
Loading...

Dates et versions

tel-01322052 , version 1 (26-05-2016)

Identifiants

  • HAL Id : tel-01322052 , version 1

Citer

Agata Savary. Representation and Processing of Composition, Variation and Approximation in Language Resources and Tools. Computation and Language [cs.CL]. Université François Rabelais Tours, 2014. ⟨tel-01322052⟩
250 Consultations
1435 Téléchargements

Partager

Gmail Facebook X LinkedIn More