Glasgow Representation and Information Learning Lab (GRILL) at the Conversational Assistance Track 2020
Glasgow Representation and Information Learning Lab (GRILL) at the Conversational Assistance Track 2020
复制标题
2020 年对话协助赛道上的格拉斯哥表示和信息学习实验室 (GRILL)
DOI：
--
复制
发表时间：
2020
期刊：
影响因子：
--
通讯作者：
Jeffrey Dalton
中科院分区：
文献类型：
--
作者：
Carlos Gemmell;Jeffrey Dalton
在本文中，我们介绍了我们的方法，实验设置和结果的会话辅助跟踪（CAST）在TREC 2020。我们提出了一种新的神经查询重写对象，用于会话消歧，通过调整预训练和微调目标来最大限度地提高语义和语法知识转移。当解决查询时，我们的模型重新生成以前的上下文，在填充目标中保持真实。我们的重写器同化查询和上下文来自动回归地解决来自先前话语的查询，从而使性能接近手动结果。当用作利用逐点和成对分数的多阶段检索管道的一部分时，我们的系统允许鲁棒的会话信息搜索。我们通过在赛道上手动和自动运行中显著优于中值结果来展示我们的系统，并通过定性示例展示我们系统的泛化。
In this paper we present our methods, experimental setup and results for the Conversational Assistance Track (CAsT) at TREC 2020. We present a novel neural query re-writing ob-jective for conversational disambiguation that maximises semantic and grammatical knowledge transfer by aligning pre-training and fine tuning objectives. When resolving queries, our model regenerates previous context staying true to original infilling objective. Our re-writer assimilates query and context to auto-regressively resolve queries from previous utterances resulting performance approaching that of manual results. When used as part of a multi-stage retrieval pipeline leveraging point-wise and pair-wise scores, our system allows for robust conversational information seeking. We demonstrate our system by significantly outperforming median results in both manual and automatic runs from the track and show generalisation of our system with qualitative examples.