Um sistema gerador de respostas enganadoras para questões de exame

• Autor(a) principal: Fontes, João Rui Macedo
• Data de Publicação: 2019
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10400.22/15613
• topic: Natural Language Processing (NLP); Golang; Automatic Question Generation (AQG); Neural Networks (NN)

University professors are responsible for teaching and grading their students in each semester. Normally, in order to evaluate the students progress, professors create exams that are composed of questions regarding the subjects taught in the teaching period. Each year, professors need to develop new questions for their exams since students are free to discuss and register the correct answers to the various questions on prior exams. Professors want to be able to grade students based on their knowledge and not on their memorization skills. Each year, as discovered by our research, professors spend over roughtly 2:30 hours each year for a single course only on multiple answer questions sections. This solution will have at its core a misleading answer generator that would reduce the time and effort when creating a Fill Gap Type Questions through the merger of highly biased lexical model towards a specific subject with a generalist model. To help the most amount of professors with this task a web-server was implemented that served as an access to a exam creator interface with the misleading answer generator feature. To implement the misleading answer generator feature, several accessory programs had to be created as well as manually edditing textbooks pertaining to the question base topic. To evaluate the effectiveness of our implementation, several evaluation methods were proposed composed of objective measurements of the misleading answers generator, as well as subjective methods of evaluation by expert input. The development of the misleading answer suggestion function required us to build a lexical model composed from a highly biased corpus in a specific curricular subject. A highly biased model is probable to give good in-context misleading answers but their variance would most likely be limited. To counteract this the model was merged with a generalist model, in hopes of improving its overall performance. With the development of the custom lexical model and the server the professor can receive misleading answers suggestions to a newly formed question reducing the time spent on creating new exams questions each year to assess students’ knowledge.