Question Answering System untuk Topik Krisis Iklim dengan Model Sentence-BERT

Abstract

In obtaining information among the many documents with various topics stored on the internet, it is a challenge for users to choose the right information from each search result. This is one of the triggers for users to be exposed to disinformation on important topics such as the climate crisis, which will have an impact on hampering climate change mitigation efforts. In this research, the author builds a question answering system that can receive question sentence text with semantics and overcome differences in sentence structures that have the same meaning, synonyms and word variations to provide relevant answers using the Sentence-BERT model. In this system, the knowledge base used is FAQ from several websites such as ipcc.ch, climate.nasa.gov, nature.org, imperial.ac.uk, natgeokids.com, theguardian.com. The results of experiments fine-tuning the indoSBERT model with 6461 data for training, 1384 validation data, and 1385 test data, as well as with hyperparameters, namely batch size 16, learning rate 2e-5, warmup steps 0.1, and with epoch trials 3, 5, and 10 respectively resulted in MRR evaluation values of 0.878, 0.882, and 0.880. Epoch 5 with the best evaluation results was selected for use in the system. The result of testing the system using 10 queries is that the model can understand the context of the question so that it can provide relevant answers at the top position for each question, with an MRR evaluation value of 0.9.