A probabilistic model of early argument structure acquisition

Developing computational algorithms that capture the complex structure\nof natural language is an open problem. In particular, learning the\nabstract properties of language only from usage data remains a\nchallenge. In this dissertation, we present a probabilistic\nusage-based model of verb argument structure acquisition that can\nsuccessfully learn abstract knowledge of language from instances of\nverb usage, and use this knowledge in various language tasks. The\nmodel demonstrates the feasibility of a usage-based account of\nlanguage learning, and provides concrete explanation for the\nobserved patterns in child language acquisition.\n\nWe propose a novel representation for the general constructions of\nlanguage as probabilistic associations between syntactic and semantic\nfeatures of a verb usage; these associations generalize over the\nsyntactic patterns and the fine-grained semantics of both the verb and\nits arguments. The probabilistic nature of argument structure\nconstructions in the model enables it to capture both statistical\neffects in language learning, and adaptability in language use. The\nacquisition of constructions is modeled as detecting similar usages\nand grouping them together. We use a probabilistic measure of\nsimilarity between verb usages, and a Bayesian framework for\nclustering them. Language use, on the other hand, is modeled as a\nprediction problem: each language task is viewed as finding the best\nvalue for a missing feature in a usage, based on the available\nfeatures in that same usage and the acquired knowledge of language so\nfar. In formulating prediction, we use the same Bayesian framework as\nused for learning, a formulation which takes into account both the\ngeneral knowledge of language (i.e., constructions) and the specific\nbehaviour of each verb. We show through computational simulation that\nthe behaviour of the model mirrors that of young children in some\nrelevant aspects. The model goes through the same learning stages as\nchildren do: the conservative use of the more frequent usages for each\nindividual verb at the beginning, followed by a phase when general\npatterns are grasped and applied overtly, which leads to occasional\novergeneralization errors. Such errors cease to be made over time as\nthe model processes more input.\n\nWe also investigate the learnability of verb semantic roles, a\ncritical aspect of linking the syntax and semantics of verbs. In\ncontrary to many existing linguistic theories and computational models\nwhich assume that semantic roles are innate and fixed, we show that\ngeneral conceptions of semantic roles can be learned from the semantic\nproperties of the verb arguments in the input usages. We represent\neach role as a semantic profile for an argument position in a general\nconstruction, where a profile is a probability distribution over a set\nof semantic properties that verb arguments can take. We extend this\nview to model the learning and use of verb selectional preferences, a\nphenomenon usually viewed as separate from verb semantic roles. Our\nexperimental results show that the model learns intuitive profiles for\nboth semantic roles and selectional preferences. Moreover, the learned\nprofiles are shown to be useful in various language tasks as observed\nin reported experimental data on human subjects, such as resolving\nambiguity in language comprehension and simulating human plausibility\njudgements.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.