Can Multi‑Task Learning Shorten E‑Commerce Titles Without Losing Sales?

Abstract

In e‑commerce platforms such as Taobao and Tmall, merchants often create overly long product titles for SEO, which cannot be fully displayed on mobile apps, harming user experience and conversion. This work introduces a multi‑task learning method that compresses long titles into short ones without reducing transaction conversion rates.

Research Background

Product titles are a key communication medium between sellers and buyers. With the rapid shift of online shopping to mobile devices, limited screen space makes long titles problematic. Overly redundant titles are frequently truncated, preventing users from seeing the full information.

Existing Methods

Traditional text summarization methods fall into extractive and generative categories. Extractive methods select words directly from the source, while generative methods can produce new words. Prior e‑commerce title compression approaches rely on heavy manual preprocessing or focus solely on length reduction, ignoring click‑through and conversion metrics.

Method Introduction

The proposed framework consists of two Sequence‑to‑Sequence tasks sharing a common encoder. The primary task uses a Pointer Network to generate a short title from the original title (extractive summarization). The auxiliary task employs an attention‑based encoder‑decoder to generate the corresponding user search query. Joint optimization aligns the attention distributions of both tasks, encouraging the short title to retain core information that drives conversions.

Main Contributions

The multi‑task learning method outperforms traditional extractive summarization in offline automatic, human, and online evaluations.

An end‑to‑end training pipeline eliminates the need for extensive manual preprocessing and feature engineering.

Attention distribution consistency ensures that generated short titles highlight important keywords that can drive transactions.

Experimental Results

Experiments were conducted on Taobao women's clothing titles, comparing five methods: Truncation, Integer Linear Programming (ILP), Pointer Network (Ptr‑Net), vanilla multi‑task learning (Vanilla‑MTL), and the proposed attention‑consistent multi‑task learning (Agree‑MTL).

Automatic Evaluation

ROUGE scores show that Agree‑MTL significantly surpasses other methods.

Human Evaluation

Human judges rated readability, informativeness, and core product word accuracy. Agree‑MTL achieved the highest scores across all three metrics.

Online A/B Test

Compared with the baseline ILP method, the proposed multi‑task learning approach improved CTR by 2.58% and CVR by 1.32% in a live environment.

Conclusion

By leveraging user search logs and multi‑task learning, the proposed method compresses overly long e‑commerce product titles while preserving core information and enhancing click‑through and conversion rates, outperforming traditional summarization techniques in both offline and online evaluations.

This paper was published in AAAI 2018.