用於商品識別與檢索的無監督式表徵之學習

Abstract

解決物件的相似性是數據科學中的一個持續存在的問題。在電子商務中，識別替代商品和搜尋相似商品都涉及相似度計算。收集淨化過的帶標記的數據集是非常昂貴。然而，無標籤的多模態數據一直在電子商務網站上生成，並且經常會出現相同類別的商品具有相似圖像或標題的情況。在本文中，我們利用商品的文本標題和圖像來訓練可將商品標題轉換為區分性表徵形式的無監督式模型。我們提出的模型集成了多任務學習和自注意力功能，可以融合文本和圖像特徵並提取商品類別資訊。我們還提出了一套新穎的自我監督機制和注意力規範，以增強學習商品類別表徵的能力。我們在兩個任務上評估我們的模型：商品識別任務和相似的商品檢索任務。結果表明，在這兩項任務上我們的模型優於近幾年提出的單模態和多模態方法。

Solving object similarity is an ongoing problem in data science. In the e-commerce retail, identifying alternative product and finding similar products both involve calculating similarity. Producing a dataset with clean labeled data is expensive. Nevertheless, unlabeled multimodal data is being generated on e-commerce websites all the time, and it is often the case that products of the same category have similar images or titles. In this paper, we leverage the aligned text and image data of e-commerce products to train an unsupervised model that transform product titles to discriminative representations. The model we proposed integrates multi-task learning and self-attention which can fuse text and image features and extract product category information. We also propose a self-supervision mechanism and a novel attention regularization to strengthen the ability to learn category-based product representations. We evaluate our model on two tasks: product identification and similar product retrieval. Results show that our model outperforms state-of-the-art unimodal methods and multi-modal methods recently proposed on both tasks.