Transfer Learning for Relationship Classification using Instance Based Convolutional Neural Network

• Main Author: Syaifullah, Robby
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/43889
• Institution: Institut Teknologi Bandung

Information extraction is a task in natural language processing which aims to extract information from unstructured text into structured text. The extracted information is in the form of an entity and the relationship between the entities of the text. At this time, the development of relationship classification with machine learning approaches especially deep neural networks (DNN). DNN has a barrier if the data available for the training model for the classification of relations in the domain is limited. Transfer learning is one solution to the problem of the small amount of data by helping to model training in new domains with existing models from similar source domains. The DNN that was tested in this final project experiment was Convolutional Neural Network (CNN) with instanced-based transfer learning techniques. Instance based transfer learning is a transfer learning method that adds instances from the source domain to the target domain with weight adjustments. In building the model, the selection of available datasets and manual annotations for the domain of the datataset is needed. The dataset selected is the SemEval2010 dataset task 8. The number of relation types in this dataset is 9. The dataset is divided into 12 domains based on the topic of each sentence entry. After that, the dataset is converted into a numeric vector as a word representation of each instance by using word embedding and position embedding. The final research baseline is classification of relations without transfer learning Bidirectional Long-short Term Memory (BiLSTM). The source domain used is the Story and News domain while the target domain is Techonology with 500 training data and 100 test data, and Politics domain with 240 training data and 60 test data. From the experimental results, a model with an instance based transfer learning approach succeeded in improving the performance of the target domain model namely 34.36% accuracy score and increasing F1 score 96.21% higher than the training model with the target domain only.