Naslov
Dynamic Language Models

Autori
Ćavar, Damir

Izvornik
PPT prezentacija, JOTA, Ljubljana

Vrsta, podvrsta
Ostale vrste radova, ostalo

Godina
2007

Ključne riječi
statistical language processing; grammar induction

Sažetak
The last decades of computational linguistics were characterized by a shift from rule-based and linguistically motivated language modeling for various types of applications towards a statistical paradigm, a paradigm that might be said to have its significant foundations in the empiricist linguistic in the 50's and 60's. The main motivation for the paradigm shift seems to have been a lack of coverage, the complexity of the models (rules and grammars), and the expert effort that is required to generate wide coverage rule- based models with enough robustness and efficiency. The apparent advantage of statistical models appeared to be elegancy, efficiency, and the independents of linguistic experts in the development of natural language processing applications. However, looking at the performance of rule-based and statistical models, there does not seem to be a serious advantage in any of those. Furthermore, while statistical models can usually better cope with deviations from rules or grammars, rule- or grammar- based models could contain rare, but peculiar constraints of natural language, that statistical models could not cope with, or they had to make use of e.g. sophisticated smoothing techniques that model events outside the scope of the available language data. Statistical models rely on large corpora that need to be annotated and checked by experts, so the apparent reduction of effort is also rather relative. Furthermore, probably less than 1% of all natural languages are documented sufficiently enough in form of language corpora (textual or acoustic), so they appear outside the scope of plain statistical methods that rely on training on enriched linguistic data. There seems to be a general problem with both approach types. It seems that they both presuppose a static language model, i.e. static sets of rules or grammars, or once trained and maybe adapted and smoothed statistical models that do not extend or (automatically) adjust to specific applications and potential changes of natural language data over time, domain, or scenario. In this talk we will discuss possibilities to make rule-based or statistical language models adaptive, and in fact incrementally extensible, using machine learning strategies. We will discuss such approaches, their psycholinguistic and cognitive grounding, applied to different linguistic levels, e.g. morphology, syntax, and (lexical) semantics, and their potential for real NLP applications, as well as their relation to psycholinguistic models of language acquisition and change.

Izvorni jezik
Engleski

Znanstvena područja
Računarstvo, Informacijske i komunikacijske znanosti, Filologija

POVEZANOST RADA