El próximo viernes 25 de marzo 2022 el Prof. Arief Setyanto de la University of Amikom Yogyakarta (Indonesia), en el marco del programa ERASMUS+ Movilidad Internacional (MI), dará una charla relacionada con el uso de Deep Learning en Natural Language Processing titulada:

 

 

 

"Opinion mining is an important task to understand public opinion polarity towards an issue. Understanding public opinion leads to better decisions in many fields, such as public services or business. Language background plays a vital role in understanding opinion polarity. The sentence is a time series signal; therefore, sequence gives a significant correlation to the meaning of the text. A recurrent neural network (RNN) is a variant of deep learning where the sequence is considered. Long-term short memory (LSTM) is an implementation of RNN with a particular gate to keep or ignore specific word signals during a sequence of input. Text is unstructured data, and it cannot be processed further by machine unless an algorithm transforms the representation into readable machine learning format in a vector of numerical value. The transformation algorithm ranging from Terms Frequency – Invers Text Frequency (TF-IDF) transform text to some advanced word Embedding. Some word embedding methods include GloVe, word2vec, BERTH, and fastText. This research experimented with those sentences to vector transformation of the text dataset. This study implements Word2Vec, as well as comparing GloVe and fastText word embedding algorithms and Long short-term memory (LSTM) implementations in single, double, and triple layers. We propose indonesian sentiment dataset, evaluate LSTM architecture on indonesian sentiment dataset. It evaluates the proposed algorithm with a dataset from traveling site reviews consisting of 25,000 reports in two classes of equal proportion (positive and negative). According to the evaluation results, the model has 95.0% accuracy. Our research also comparing various LSTM architectures and their accuracy for the opinion mining Arabic dataset. It evaluates the proposed algorithm with the ASAD dataset of 55k annotated tweets in three classes. The dataset was augmented to achieve equal proportions of positive, negative, and neutral classes. According to the evaluation results, the triple layers LSTM with fastText word embedding achieves the best testing accuracy at 88,9%, surpassing all other experiment scenarios."

 

Duración: 10:30-11:30 (1h)

 

Lugar: Salón de Grados.

 

Esperamos que sea de vuestro interés.