hrvatski

Duboko učenje za ekstraktivno sažimanje teksta

Extractive text summarization is tasked with the automatic creation of a summary by extracting the most salient sentences from the original text. In this thesis, two extractive summarization methods are trained and tested: multilayer perceptron and gated recurrent unit. Both methods are trained as a binary classifier capable of assigning the class summary or not-summary to input sentences. In the multilayer perceptron method, each sentence is represented by Doc2Vec embedding. The embedding layer of GRU is initialized with Word2Vec vectors of the words in the vocabulary and the input vectors contain sequences of the words from the original text. The output is a class probability assigned by the logistic function classifier. The CNN/DailyMail dataset is used to train and evaluate extractive summarization models using the ROUGE-1, ROUGE-2, and ROUGE-LCS measures to assess the performance. Generally, GRU achieves better extractive summarization results when precision is considered, while perceptron performs better according to the recall metrics, regardless of the used ROUGE measures. The results are indicating that both methods are capable of performing extractive summarization task.

sažimanje teksta, neuronska mreža, duboko učenje, višeslojni perceptron, rekurentna neuronska mreža, upravljačka rekurentna jedinica

nije evidentirano