Good-enough parsing, Whenever possible interpretation: a constraint-based model of sentence comprehension - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 2016

Good-enough parsing, Whenever possible interpretation: a constraint-based model of sentence comprehension

Philippe Blache

Résumé

Natural language processing and psycholinguistics are progressively getting closer, and new language processing architectures, bringing together both computational and cognitive aspects, emerge. We propose in this position paper a brief overview of the basic needs paving the way towards such unified frameworks and show how constraints constitute an adequate answer. What constraints can do: Let's start with some basic recalls about constraint programming. Constraints are not only efficient for ruling out unwanted solutions or ill-formed structures. They are also capable of building approached solutions (or ultimately instantiating values) by restricting the search space. For example, we can limit the definition domain of an integer variable x by means of interval constraints such as [x > 1; x < 4] (note that adding a new constraint [x > 2] leads to the solution x = 3). Moreover, constraints can be of different types (interval, boolean, numerical, etc.), a same variable being possibly involved is many of them. In this sense, constraints form a system that is in itself source of information: a problem can be described with a set of constraints, and this description leads to the solutions (which is the basis of distinction between declarative and procedural approaches in computer science [Colmerauer, 1986, Jaffar and Lassez, 1986]). Solving a problem consists in evaluating the constraint system which leads to instantiate values and more generally provide information about the set of variables forming the problem. In other words, the state of the constraint system after evaluation, for a given set of input values, constitutes a precise description of this input set. We propose in this short note some arguments in favor of considering this computational framework as an efficient cognitive model for language processing. Needs and requirements for sentence processing: Language is comprehended by humans in real time. To be more precise, sentence interpretation is done such efficiently in most of the cases, which means that different types of mechanisms can be at work according to the input. It is important to note that this property is preserved even when the input is not perfectly-formed (errors, disfluencies, unstructured productions, etc.), which occurs frequently in natural situations (typically during conversations). One question is then to explain how does interpretation works under noisy input. A classical solution explains that non-canonical productions are repaired, the difficulty of the interpretation being dependent to the number of repairs [Gibson, 1998]. This noisy-channel approach [Levy, 2008a, Levy, 2008b, Johnson and Charniak, 2004] proposes in particular to introduce the notions of uncertainty and inference. This constitutes a first important requirement: 1
Fichier principal
Vignette du fichier
CSLP-Blache.pdf (129.65 Ko) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)
Loading...

Dates et versions

hal-01907632 , version 1 (29-10-2018)

Identifiants

  • HAL Id : hal-01907632 , version 1

Citer

Philippe Blache. Good-enough parsing, Whenever possible interpretation: a constraint-based model of sentence comprehension. CSLP-2016, Nov 2016, New York, United States. ⟨hal-01907632⟩
80 Consultations
667 Téléchargements

Partager

Gmail Facebook X LinkedIn More