Automatic Text Summarization

• Autor(a) principal: Saraiva, João Pedro Abrantes
• Data de Publicação: 2021
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10400.6/12682
• topic: Abstractive Summarization:deep Learning; Automatic Text Summarization; Extractive Summarization; Natural Language Processing; Domínio/Área Científica::Engenharia e Tecnologia::Engenharia Eletrotécnica, Eletrónica e Informática

Writing text was one of the first ever methods used by humans to represent their knowledge. Text can be of different types and have different purposes. Due to the evolution of information systems and the Internet, the amount of textual information available has increased exponentially in a worldwide scale, and many documents tend to have a percentage of unnecessary information. Due to this event, most readers have difficulty in digesting all the extensive information contained in multiple documents, produced on a daily basis. A simple solution to the excessive irrelevant information in texts is to create summaries, in which we keep the subject’s related parts and remove the unnecessary ones. In Natural Language Processing, the goal of automatic text summarization is to create systems that process text and keep only the most important data. Since its creation several approaches have been designed to create better text summaries, which can be divided in two separate groups: extractive approaches and abstractive approaches. In the first group, the summarizers decide what text elements should be in the summary. The criteria by which they are selected is diverse. After they are selected, they are combined into the summary. In the second group, the text elements are generated from scratch. Abstractive summarizers are much more complex so they still need a lot of research, in order to represent good results. During this thesis, we have investigated the state of the art approaches, implemented our own versions and tested them in conventional datasets, like the DUC dataset. Our first approach was a frequency­based approach, since it analyses the frequency in which the text’s words/sentences appear in the text. Higher frequency words/sentences automatically receive higher scores which are then filtered with a compression rate and combined in a summary. Moving on to our second approach, we have improved the original TextRank algorithm by combining it with word embedding vectors. The goal was to represent the text’s sentences as nodes from a graph and with the help of word embeddings, determine how similar are pairs of sentences and rank them by their similarity scores. The highest ranking sentences were filtered with a compression rate and picked for the summary. In the third approach, we combined feature analysis with deep learning. By analysing certain characteristics of the text sentences, one can assign scores that represent the importance of a given sentence for the summary. With these computed values, we have created a dataset for training a deep neural network that is capable of deciding if a certain sentence must be or not in the summary. An abstractive encoder­decoder summarizer was created with the purpose of generating words related to the document subject and combining them into a summary. Finally, every single summarizer was combined into a full system. Each one of our approaches was evaluated with several evaluation metrics, such as ROUGE. We used the DUC dataset for this purpose and the results were fairly similar to the ones in the scientific community. As for our encoder­decode, we got promising results.