|
|
English grammar checker and corrector: the determiners
Auersperger, Michal ; Pecina, Pavel (advisor) ; Straňák, Pavel (referee)
Correction of the articles in English texts is approached as an article generation task, i.e. each noun phrase is assigned with a class corresponding to the definite, indefinite or zero article. Supervised machine learning methods are used to first replicate and then improve upon the best reported result in the literature known to the author. By feature engineering and a different choice of the learning method, about 34% drop in error is achieved. The resulting model is further compared to the performance of expert annotators. Although the comparison is not straightforward due to the differences in the data, the results indicate the performance of the trained model is comparable to the human-level performance when measured on the in-domain data. On the other hand, the model does not generalize well to different types of data. Using a large-scale language model to predict an article (or no article) for each word of the text has not proved successful. 1
|
|
|
Natural Language Correction
Náplava, Jakub ; Straka, Milan (advisor) ; Straňák, Pavel (referee)
The goal of this thesis is to explore the area of natural language correction and to design and implement neural network models for a range of tasks ranging from general grammar correction to the specific task of diacritization. The thesis opens with a description of existing approaches to natural language correction. Existing datasets are reviewed and two new datasets are introduced: a manually annotated dataset for grammatical error correction based on CzeSL (Czech as a Second Language) and an automatically created spelling correction dataset. The main part of the thesis then presents design and implementation of three models, and evaluates them on several natural language correction datasets. In comparison to existing statistical systems, the proposed models learn all knowledge from training data; therefore, they do not require an error model or a candidate generation mechanism to be manually set, neither they need any additional language information such as a part of speech tags. Our models significantly outperform existing systems on the diacritization task. Considering the spelling and basic grammar correction tasks for Czech, our models achieve the best results for two out of the three datasets. Finally, considering the general grammatical correction for English, our models achieve results which are...
|
guest ::
