PACIFIC

为了促进金融混合环境中的对话问题回答 (CQA)，我们提出了一个名为PACIFIC的新数据集。与现有的CQA数据集相比，PACIFIC具有三个主要特征 :( i) 主动性，(ii) 数值推理以及 (iii) 表格和文本的混合上下文。相应地定义了一个新任务来研究主动对话问题回答 (PCQA)，该任务结合了澄清问题生成和CQA。此外，我们提出了一种新颖的方法，即UniPCQA，将PCQA中的输入和输出内容的混合格式改编为Seq2Seq问题，包括将数值推理过程重新表述为代码生成。UniPCQA对PCQA中的所有子任务执行多任务学习，并结合了一个简单的集成策略，通过交叉验证top-k采样Seq2Seq输出来缓解多任务学习中的错误传播问题。我们用广泛的基线对太平洋数据集进行基准测试，并对PCQA的每个子任务进行全面评估。

To advance conversational question answering (CQA) in financial hybrid scenarios, we propose a novel dataset named PACIFIC. Compared with existing CQA datasets, PACIFIC features three core characteristics: (i) proactivity, (ii) numerical reasoning, and (iii) hybrid context of tables and text. Accordingly, we define a novel task to investigate proactive conversational question answering (PCQA), which integrates clarification question generation and CQA. Furthermore, we propose a novel method called UniPCQA that adapts the mixed formats of inputs and outputs in PCQA into Seq2Seq problems, including reformulating the numerical reasoning process into code generation. UniPCQA performs multi-task learning across all subtasks in PCQA, and incorporates a simple ensemble strategy to mitigate error propagation in multi-task learning by cross-validating the top-k sampled Seq2Seq outputs. We benchmark the PACIFIC dataset with extensive baselines and conduct comprehensive evaluations on each subtask of PCQA.

搜集汇总

背景与挑战

背景概述

PACIFIC是一个用于金融混合环境中对话问题回答的数据集，具有主动性、数值推理和混合上下文的特点，并提出了主动对话问题回答（PCQA）任务和UniPCQA方法。