4:00 PM - 4:20 PM
[1J2-03] Neural Ranking with Pretraining for Curation of Important Sentences in Business Operations
Keywords:Learning to Rank, Deep Learning
We have developed a machine learning method for ranking sentences based on their importance, to assist decision making in business operations. The proposed method is based on deep learning that is trained end-to-end from user feedbacks to capture semantic importance of sentences without relying on domain knowledge of business operations.
We employ pretraining to avoid needs for large training data, which we cannot obtain easily because there are numerous groups of people with different interest in business operations and the each group tends to be small. The proposed method outperformed conventional ranking methods in an automated evaluation. We validated the quality of the ranking with a subjective impression evaluation.
We employ pretraining to avoid needs for large training data, which we cannot obtain easily because there are numerous groups of people with different interest in business operations and the each group tends to be small. The proposed method outperformed conventional ranking methods in an automated evaluation. We validated the quality of the ranking with a subjective impression evaluation.