Unsupervised syntactic disambiguation for turkish

In natural languages, a sentence can be represented by more than one syntax tree, each one corresponding to different structural interpretations. This is called syntactic ambiguity. To put it simply, in syntactic disambiguation, the syntactic trees obtained from the sentence are ranked from the most appropriate to the least appropriate based on the context. In this dissertation, the problem of syntactic disambiguation is addressed for Turkish and a solution based on an unsupervised method is proposed. The reason for naming the proposed method as unsupervised is that the probability models used for sorting syntax trees are derived from an unannotated text collection. Within the scope of the dissertation, in order to realize the syntactic disambiguation process, novel infrastructure items including a syntactic parser, a morphologic analyzer called Morfolog, a lexicon called TrLex are designed and a system named TMoST that manages them in a coordinated manner is constituted. Besides, a new sentence representation based on phrase structure grammar is proposed and a new concept called syntheme, which allows morphological and syntactic structures to work together, is introduced. In the study, 24 probabilistic models, some of which are novel, are used. In order to measure the performance of the models over the problem, a treebank called AUT is constituted as well. In the literature, the performance for syntactic disambiguation is commonly measured by the position of the best tree in the ranking or by the similarity of the first tree to the best one. In the dissertation, two new performance measures are proposed and it is revealed that the measure called correlation is more stable. When the probabilistic models are used individiually, the best performance is obtained with the morpheme trigram language model. When the models are combined, the best correlation value is achieved as 0.41 approximately.