Data Augmentation for Abstractive Query-Focused Multi-Document Summarization
Ramakanth Pasunuru,1 Asli Celikyilmaz,2 Michel Galley,2 Chenyan Xiong,2 Yizhe Zhang,2 Mohit Bansal,1 Jianfeng Gao2
Abstract
The progress in Query-focused Multi-Document Summarization (QMDS) has been limited by the lack of sufficient large-scale high-quality training datasets. We present two QMDS training datasets, which we construct using two data augmentation methods: (1) transferring the commonly used single-document CNN/Daily Mail summarization dataset to create the QmdsCnn dataset, and (2) mining search-query logs to create the QmdsIr dataset. These two datasets have complementary properties, i.e., QmdsCnn has real summaries but queries are simulated, while QmdsIr has real queries but simulated summaries. To cover both these real summary and query aspects, we build abstractive end-to-end neural network models on the combined datasets that yield new state-of-the-art transfer results on DUC datasets. We also introduce new hierarchical encoders that enable a more efficient encoding of the query together with multiple documents. Empirical results demonstrate that our data augmentation and encoding methods outperform baseline models on automatic metrics, as well as on human evaluations along multiple attributes.111Code: https://github.com/ramakanth-pasunuru/QmdsCnnIr
中文速览
查询式多文档摘要（Query-focused Multi-Document Summarization, QMDS）长期面临大规模高质量训练数据稀缺的瓶颈，本文通过两种数据增强方法构建了两个新数据集：QmdsCnn（把单文档CNN/DM数据集改造成多文档形式，查询用文章标题模拟）和QmdsIr（从必应搜索日志中挖掘真实用户查询与网页文档），两者在"真实摘要"与"真实查询"上互补，合并后为模型训练提供了约40万条样本。在此基础上，作者提出了层级化神经摘要模型HEROSumm，通过同时利用局部与全局文档编码、端到端的文档排序模块以及独立的查询编码器，更有效地整合查询与多篇文档的信息。实验显示，该方法在DUC 2006/2007等标准测评集上取得了新的最优迁移学习结果，并在人工评测的多个维度上均优于基线。这项工作的意义在于，它不仅缓解了QMDS领域的数据饥渴问题，还为个性化信息检索、对话式搜索等现实场景提供了更实用的摘要建模方案。
原文 arXiv:2103.01863;中英对照 + 大白话阅读 https://aha.fim.ai/paper/2103.01863v1