| Summary: | The automatic answer to questions in natural language is an area that has been studied for many years. However, based on the existing question answering systems, the percentage of correct answers over a set of questions, generated from a dataset, we can see that the performance it is still far away from to 100%, which is many times the value achieved when the questions are tested by humans. This work addresses the idea of a book-oriented Chatbot, more precisely a question answering system directed to answer to questions in which the dataset is one or more books. This way, we intend to adopt a new system, incorporating two existent projects, the OpenBookQA and the Question-Generation. We have used two Domain Specific Datasets that were not studied in both project, that were the QA4MRE and RACE. To these we have applied the main approach: enrich them with automatic generated questions. We have run many experiments, training neural network models. This way, we intended to study the impact of those questions and obtain good accuracy results for both datasets. The obtained results suggest that having a significant representation of generated questions in a dataset, leads to a higher test accuracy results of correct answers. Becoming clear that, enrich a dataset, based on a book, with generated questions about that book, is giving to the dataset the content of the book. This dissertation presents promising results, through the datasets with automatic generated questions.
|
|---|