Dynamic-FACT: A Dynamic Framework for Adaptive Context-Aware Translation

3593

Authors

nan

Abstract

Abstract“Document-level neural machine translation (NMT) has garnered considerable attention sincethe emergence of various context-aware NMT models. However, these static NMT models aretrained on fixed parallel datasets, thus lacking awareness of the target document during infer-ence. In order to alleviate this limitation, we propose a dynamic adapter-translator frameworkfor context-aware NMT, which adapts the trained NMT model to the input document prior totranslation. Specifically, the document adapter reconstructs the scrambled portion of the originaldocument from a deliberately corrupted version, thereby reducing the performance disparity be-tween training and inference. To achieve this, we employ an adaptation process in both the train-ing and inference stages. Our experimental results on document-level translation benchmarksdemonstrate significant enhancements in translation performance, underscoring the necessity ofdynamic adaptation for context-aware translation and the efficacy of our methodologies. Introduction”

Keywords

nan

Full Content

基 基 基于 于 于多 多 多意 意 意图 图 图融 融 融合 合 合框 框 框架 架 架的 的 的联 联 联合 合 合意 意 意图 图 图识 识 识别 别 别和 和 和槽 槽 槽填 填 填充 充 充 尹 尹 尹商 商 商鉴 鉴 鉴, 黄 黄 黄沛 沛 沛杰 杰 杰*, 梁 梁 梁栋 栋 栋柱 柱 柱, 何 何 何卓 卓 卓棋 棋 棋, 黎 黎 黎倩 倩 倩尔 尔 尔, 徐 徐 徐禹 禹 禹洪 洪 华南农业大学, 数学与信息学院, 广东广州, 510642洪 sjy8460@163.com, pjhuang@scau.edu.cn, liang dz@stu.scau.edu.cn, 13428897035@163.com, li@stu.scau.edu.cn, xuyuhong@scau.edu.cn 摘 摘 摘要 要 要 近年来,多意图口语理解(SLU)已经成为自然语言处理领域的研究热点。当前先进 的多意图SLU模型采用图-交互式框架进行联合多意图识别和槽位填充,能够有效地 捕捉到词元级槽位填充任务的细粒度意图信息,取得了良好的性能。但是,它忽略了 联合作用下的意图所包含的丰富信息,没有充分利用多意图信息对槽填充任务进行指 引。为此,本文提出了一种基于多意图融合框架(MIFF)的联合多意图识别和槽填 充框架,使得模型能够在准确地识别不同意图的同时,利用意图信息为槽填充任务提 供更充分的指引。我们在MixATIS和MixSNIPS两个公共数据集上进行了实验,结果 表明,我们的模型在性能和效率方面均超过了当前最先进的方法,同时能够有效从单 领域数据集泛化到多领域数据集上。 关 关 关键 键 键词 词 词: : : 多意图口语理解 ;多意图融合框架 ;联合多意图识别和槽位填充 A Multi-Intent Fusion Framework for Joint Intent Detection and Slot Filling Shangjian yin, Peijie Huang*, Dongzhu Liang, Zhuoqi He, Qianer Li, Yuhong Xu College of Mathematics and Informatics, South China Agricultural University, China sjy8460@163.com, pjhuang@scau.edu.cn, liang dz@stu.scau.edu.cn, 13428897035@163.com, li@stu.scau.edu.cn, xuyuhong@scau.edu.cn Abstract In recent years, multi-intent spoken language understanding (SLU) has become a re- search hotspot in the field of natural language processing. The current state-of-the-art multi-intent SLU model uses a graph-interaction framework for joint multi-intent de- tection and slot-filling, which can effectively capture fine-grained intent information for lexical element-level slot-filling tasks and achieves good performance. However, it ignores the rich information contained in the intent under joint action and does not fully utilize the multi-intent information to guide the slot-filling task. To this end, this paper proposes a joint multi-intent detection and slot-filling approach based on the multi-intent fusion framework (MIFF), which enables the model to accurately identify different intents while using the intent information to provide more adequate guidance for the slot-filling task. We conducted experiments on two public datasets, MixATIS and MixSNIPS, and the results show that our model outperforms current state-of- the-art methods in terms of performance and efficiency, while being able to effectively generalize from single-domain dataset to multi-domain dataset. *通讯作者 ©2023 中国计算语言学大会 根据《Creative Commons Attribution 4.0 International License》许可出版 CCL 2023 第 二 十 二 届 中 国 计 算 语 言 学 大 会 论 文 集 , 第 54页 -第 63页 , 哈 尔 滨 , 中 国 , 2023年 8月 3日 至 5日 。 (c) 2023 中 国 中 文 信 息 学 会 计 算 语 言 学 专 业 委 员 会 54 计算语言学 Keywords: Spoken language understanding , Multi-intent fusion framework , Joint intent detection and slot filling. 1 引 引 引言 言 言 口语理解(Spoken Language Understanding, SLU) (Tur et al., 2011; Young et al., 2013)是 面向任务的对话的关键组成部分系统,其目的是创建一个语义框架,总结用户的请求。利用 意图识别来识别用户意图,利用槽填充来提取相关语义成分,构建语义框架。由于意图识别 和槽填充这两个子任务紧密相关,主流SLU系统采用联合模型来建模它们之间的相关性。在 现实场景中,用户通常在话语中表达多个意图,如亚马逊内部数据集中52%的例子是多意图 的(Gangadharaiah et al., 2019)。Figure 1展示了一个两意图示例,其中包含一个分类任务来对 意图标签进行分类(即,预测意图为atis aircraft 和atis city) 和一个序列标记任务来预测槽 标签序列(即,将语句标记为{O, O, O, O, O, O, O, B-aircraft code, O, O, O, O, O, O, B-city name, O })。然而,以往的工作大多只关注简单的单意图场景,无法有效处理原始网络的多意图。最 近,多意图SLU逐渐受到关注,因为它在我们日常生活中具有重要的应用价值 (Gangadharaiah et al., 2019; Qin et al., 2000),它可以处理包含多个意图的语句。为了满足现实生活的需 要,Xu等人 (2013)和Kim等人 (2017)开始探索多意图SLU。然而,他们的模型只考虑了多重意 图识别,而忽略了槽填充任务。最近,Gangadharaiah 和Narayanaswamy (2019) 首次尝试提出 了一个多任务框架来联合建模多重意图识别和槽填充。Qin等 (2020)进一步提出了一种自适应 交互框架(AGIF)来实现细粒度的多意图信息集成。基于图注意网络[7]和非自回归方式的思 想,Qin等人 (2021)提出了一个全局局部图交互网络(global - local graph Interaction network, GL-GIN),该网络建模了多个意图和话语中所有槽之间的槽依赖和交互,获得了先进的性能。 虽然该模型取得了较好的效果,然而,它在模型浅层交互中只是将多意图和槽填充信息简单的 拼接,没有充分挖掘意图中对槽的填充具有重要的指导意义的信息,从而让训练效率和模型性 能大打折扣。 为了进一步增强多意图与槽填充的关联性以及协同优化的效果,我们提出了一个多意图融 合框架,其核心是多意图融合层,该层包含了三种信息的融合过程,实现了在意图识别和槽填 充任务之间建立更加稳定有效的连接。考虑到意图识别任务在联合效应中起着重要作用,我们 提出了意图强化层,获得了更丰富的多意图信息表示,提高了整体框架的鲁棒性。在MixATIS (Hemphill et al., 1990)和MixSNIPS (Coucke et al., 2018)两个公共数据集上的实验结果表明, …

Content truncated to 500 words

Labeled Text

Abstract“Document-level neural machine translation (NMT) has garnered considerable attention sincethe emergence of various context-aware NMT models. However, these static NMT models aretrained on fixed parallel datasets, thus lacking awareness of the target document during infer-ence. In order to alleviate this limitation, we propose a dynamic adapter-translator frameworkfor context-aware NMT, which adapts the trained NMT model to the input document prior totranslation. Specifically, the document adapter reconstructs the scrambled portion of the originaldocument from a deliberately corrupted version, thereby reducing the performance disparity be-tween training and inference. To achieve this, we employ an adaptation process in both the train-ing and inference stages. Our experimental results on document-level translation benchmarksdemonstrate significant enhancements in translation performance, underscoring the necessity ofdynamic adaptation for context-aware translation and the efficacy of our methodologies. Introduction”

Publication Details

Paper ID:
3593
Year:
2023
Publication:
TIPSTER
Publication Type:
conf
DOI:
nan