Word Filtering Approach for Next Product Title Generation
Keywords: KDD Cup, Multilingual Session Recommendation, Text Generation
TL;DR: Introducing 6th place approach for KDD Cup 2023 Task 3.
Abstract: The rise of e-commerce has prompted e-commerce platform providers to make efforts toward improving the shopping experience of their users. One approach involves the optimization of product recommendations tailored to user preferences and the presentation of product titles based on historical session data, which encompass past instances of browsing various products. In considering these, the organizers of KDD Cup 2023 launched a competition focused on session-based recommendations, accompanied by a comprehensive dataset known as the Amazon Multilingual Multi-locale Session Dataset (Amazon-M2). In this paper, we, NTT-DOCOMO-LABS-RED, present our solution that achieved a 6th place ranking on the public leaderboard for Task 3 (Next Product Title Generation) in the competition. The proposed solution centers around the key notion that words appearing infrequently in product titles are likely extraneous and introduce noise to the prediction. By identifying and removing such words, we aim to enhance the overall quality of the solution. Furthermore, we introduce a second solution that attained a ranking equivalent to 10th place on the public leaderboard. The core idea behind the second solution involves incorporating a prediction for the BLEU score by submitting the last title to the first solution. This prediction is then utilized to determine whether to retain the best solution as it is or to apply further modifications to the title.
Submission Number: 2