TY - THES T1 - Aportación a la predicción de palabras en castellano y su integración en sistemas de ayuda a personas con discapacidad física (Contribution to word prediction in Spanish and its integration in technical aids for people with physical disabilities) T2 - Electronic Engineering Y1 - 2001 A1 - Sira E. Palazuelos ED - Aguilera Navarro, Santiago KW - augmentative and alternative comunication KW - ayudas a la comunicación KW - communication aids KW - comunicación aumentativa y alternativa KW - people with disabilities KW - personas con discapacidad KW - predicción de palabras KW - word prediction in Spanish AB - This Ph.D. thesis is aimed at the study of including linguistic information in word prediction for Spanish, with the main objective of improving the writing aids available for people with different kind of disabilities. In order to include linguistic information, we propose a novel architecture that allows the development of an original methodology in order to combine the different sources of information we have explored (mainly in the lexical, morphological and syntactic levels), thanks to the inclusion of a management module, able to deal with and combine the different information ows used, and to the strict separation between the lexicons (main, custom and subject) and the prediction methods themselves. The prediction methods included use two main modeling strategies for the linguistic information: stochastic modeling (unigrams, bigrams, bipos and tripos) and formal modelling (using a probabilistic context free grammar strengthened with additional characteristics). In every module including linguistic knowledge, we have made specific contributions, both in the design and organization of the information (mainly oriented to be used in the formal grammar) and in the particular methodology of using this information when facing word prediction and the adequate cooperation with other modules. We also consider a significant contribution of this thesis the design criterion and the definition of the grammatical parts-of-speech (pos) used, in order to better connect with the observed syntactic behavior, along with the design of a feature set towards which we have shifted part of the expressive content. In order to deal with both pos and features, we also propose some original mechanisms included in the design of the formal grammar. With respect to the formal model, the detailed study of linguistic phenomena (both theoretically and empirically) has led us to design a probabilistic context free grammar that uses an original interweaving of different mechanisms (terminal symbol feature concordance, imposition and prohibition; powerful feature management also in non terminal symbols; lemma and word imposition and prohibition; and the possibility of dealing with optional symbols) that endow it with a significant descriptive power of the language, while keeping the number of rules and the search process computationally tractable. Our work is not only limited to a theoretical study. We have also implemented and evaluated a working system, built following the proposed architecture in which, additionally, we have taken into account specific considerations on the user interface design. We also contribute a detailed study on the different factors that affect the quantitative evaluation (where a normalization effort should be done, given the lack of defined standards on this topic), proposing metrics able to analyze the power of the information sources that allows us to select the best combination strategy leading to actual improvements for the users of this technology. In this combination, we prioritize the words coming from the subject and custom lexicons using a bigram model. After this, we use the stochastic pos models, applied first to the subject lexicon and afterwards, with an adequate weighting, to the custom and main lexicons. With respect to the word prediction method based in the formal grammar, the overall set of contributions allowed us to get results close to those obtained with the stochastic pos models, leaving for future research the completion of its descriptive capabilities. The modularity and flexibility of the architecture will allow us to carry out this research work taking great advantage of the effort already invested here. JF - Electronic Engineering PB - Politecnica de Madrid CY - Madrid U1 - PhD. in Telecommunication Engineering ER -