基于神经主题模型的对话情感分析

对话情感分析旨在识别出一段对话中每个句子的情感倾向,其在电商客服数据分析中发挥着关键作用。不同于对单个句子的情感分析,对话中句子的情感倾向依赖于其在对话中的上下文。目前已有的方法主要采用循环神经网络和注意力机制建模句子之间的关系,但是忽略了对话作为一个整体所呈现的特点。建立在多任务学习的框架下,该文提出了一个新颖的方法,同时推测一段对话的主题分布和每个句子的情感倾向。对话的主题分布,作为一种全局信息,被嵌入到每个词以及句子的表示中。通过这种方法,每个词和句子被赋予了在特定对话主题下的含义。在电商客服对话数据上的实验结果表明,该文提出的模型能充分利用对话主题信息,与不考虑主题信息的基线模型相比,Macro-F1值均有明显提升。     

基于BERT的强化语境与语义信息的对话幽默识别模型

幽默识别是自然语言处理的新兴研究领域之一。对话的特殊结构使得在对话中的幽默识别相较于短文本幽默识别更具有挑战性。在对话中,除了当前话语以外,上下文语境信息对于幽默的识别也至关重要。因此,该文在已有研究的基础上结合对话的结构特征,提出基于BERT的强化语境与语义信息的对话幽默识别模型。模型首先使用BERT对发言人信息和话语信息进行编码,其次分别使用句级别的BiLSTM、CNN和Attention机制强化语境信息,使用词级别的BiLSTM和Attention机制强化语义信息。实验结果表明,该文方法能有效提升机器识别对话中幽默的能力。     

基于对话结构和联合学习的情感和意图分类

在社交媒体中存在大量的对话文本,而在这些对话中,说话人的情感和意图通常是相关的。不仅如此,对话的整体结构也会影响对话的情感和意图,因此,需要对对话中的情感和意图进行联合学习。为此,该文提出了基于对话结构的情感、意图联合学习模型,考虑对话内潜在的情感与意图的关联性,并且利用对话的内在结构与说话人的情感和意图之间的关系,提升多轮对话文本的每一子句情感及其意图的分类性能。同时,通过使用注意力机制,利用对话的前后联系来综合考虑上下文对对话情感的影响。实验表明,联合学习模型能有效地提高对话子句情感及意图分类的性能。     

Dimensionality of dialogue act tagsets

This article compares one-dimensional and multi-dimensional dialogue act tagsets used for automatic labeling of utterances. The influence of tagset dimensionality on tagging accuracy is first discussed theoretically, then based on empirical data from human and automatic annotations of large scale resources, using four existing tagsets: damsl, swbd-damsl, icsi-mrda and maltus. The Dominant Function Approximation proposes that automatic dialogue act taggers could focus initially on finding the main dialogue function of each utterance, which is empirically acceptable and has significant practical relevance.     

Pragmatic modeling: toward a robust natural language interface

One of the most important ways in which an information-provider can assimilate an information-seeking dialogue is by inferring the underlying task-related plan motivating the information-seeker's queries. This paper presents a strategy for hypothesizing and tracking the changing task-level goals of an information-seeker and building a model of his task-related plan as the dialogue progresses.
Naturally occurring utterances are often imperfect. The information-provider often appears to use acquired knowledge about the information-seeker's underlying task-related plan to remedy many of the information-seeker's faulty utterances and enable the dialogue to continue without interruption. This paper presents a strategy for understanding one kind of defective utterance. Our approach relies on the information-seeker's inferred task-related plan as the primary mechanism for suggesting how an utterance should be understood, thereby considering only interpretations that are relevant to what the information-seeker is trying to accomplish. If multiple interpretations are suggested, relevance to the current focus of attention in the dialogue and similarity to the information-seeker's actual utterance are used to select the interpretation that is most likely to represent his intended meaning or satisfy his needs.     

面向任务口语对话系统中不含槽信息话语的端到端对话控制

端到端(end-to-end)模型因其能有效避免传统管道式设计存在的错误传递与累积问题,成为了近年来口语对话系统(spoken dialogue system, SDS)的研究热点。在面向任务SDS的end-to-end对话控制中,处理携带任务领域语义信息(槽信息)的话语可以结合命名实体识别、数据库查询结果等语义特征,而不含槽信息的话语,由于缺乏领域语义信息以及表达多样,其有效对话控制仍然是一个挑战。该文提出一种融合“显式”话语特征和“隐式”上下文信息的end-to-end混合编码网络用于处理不含槽信息话语。具体地,在应用卷积神经网络(convolutional neural network, CNN)对“显式”话语序列提取得到的特征表达的基础上,通过构造和捕获对话序列中“隐式”的系统后台上下文信息,进一步丰富了系统动作分类模型的特征表达。在限定领域面向中文任务SDS中的评估结果表明,与传统的管道式SDS和经典的end-to-end SDS相比,该文的方案在不含槽信息话语的单回合处理以及对话段整体性能上都得到了显著提升。     

Dialogue Games for Inconsistent and Biased Information

In this article, a dialogue game is presented in which coherent conversational sequences with inconsistent and biased information are described at the speech act level. Inconsistent and biased information is represented with bilattice structures, and based on these bilattice structures, a multi-valued logic is defined that makes it possible to describe a dialogue game in which agents can communicate about their cognitive states with inconsistent and biased information. A dialogue game is formalized by, first, defining the agent's cognitive state as a set of multi-valued theories, second, by defining the dialogue rules that prescribe permissible communicative acts based on the agent's cognitive state, and last, by defining update rules that change the agent's cognitive state as a result of communicative acts. We show that an example dialogue with inconsistent and biased information can be derived from our dialogue game.     

Evaluating Plan Recognition Systems: Three Properties of a Good Explanation

Plan recognition in a dialogue system is the process of explaining why an utterance was made, in terms of the plans and goals that its speaker was pursuing in making the utterance. I present a theory of how such an explanation of an utterance may be judged as to its merits as an explanation. I propose three criteria for making such judgments: applicability, grounding, and completeness. The first criterion is the applicability of the explanation to the needs of the system that will use it. The second criterion is the grounding of the explanation in what is already known of the speaker and of the dialogue. Finally, the third criterion is the completeness of the explanation's coverage of the goals that motivated the production of the utterance. An explanation of an utterance is a good explanation of that utterance to the extent that it meets these three criteria. In addition to forming the basis of a method for evaluating the merit of an explanation, these criteria are useful in designing and evaluating a plan recognition algorithm and its associated knowledge base.     

Dynamic Bayesian networks and variable length genetic algorithm for designing cue-based model for dialogue act recognition

The automatic recognition of dialogue act is a task of crucial importance for the processing of natural language dialogue at discourse level. It is also one of the most challenging problems as most often the dialogue act is not expressed directly in speaker’s utterance. In this paper, a new cue-based model for dialogue act recognition is presented. The model is, essentially, a dynamic Bayesian network induced from manually annotated dialogue corpus via dynamic Bayesian machine learning algorithms. Furthermore, the dynamic Bayesian network’s random variables are constituted from sets of lexical cues selected automatically by means of a variable length genetic algorithm, developed specifically for this purpose. To evaluate the proposed approaches of design, three stages of experiments have been conducted. In the initial stage, the dynamic Bayesian network model is constructed using sets of lexical cues selected manually from the dialogue corpus. The model is evaluated against two previously proposed models and the results confirm the potentiality of dynamic Bayesian networks for dialogue act recognition. In the second stage, the developed variable length genetic algorithm is used to select different sets of lexical cues to constitute the dynamic Bayesian networks’ random variables. The developed approach is evaluated against some of the previously used ranking approaches and the results provide experimental evidences on its ability to avoid the drawbacks of the ranking approaches. In the third stage, the dynamic Bayesian networks model is constructed using random variables constituted from the sets of lexical cues generated in the second stage and the results confirm the effectiveness of the proposed approaches for designing dialogue act recognition model.     

Accent Patterning on Domain-Related Information in Swedish Travel Dialogues

Investigation of travel-domain dialogues reveals travel-agent (System) utterances with intonational contours characterized by late-timed focal accents on given information. These accents occur on content words in utterance-initial position. The accentuation can be assumed to be related to the interactive nature of the dialogue in which the travel agent links back to a domain-related concept introduced by the client (User) and comments on it in an engaged manner. A perception test using constructed human-machine dialogues in which the machine (synthesized) responses vary as to the type of accent pattern on the initial words was developed to test listeners' preference for accent type. Results indicate that i) focal accents on domain-related utterance-initial given concepts are indeed preferred to nonfocal accents and that ii) late-timed focal accents are preferred to early-timed focal accents. These results have implications for the design of the prosody-generating component of human-machine dialogue systems.     

一种基于弱监督学习的论坛帖子对话行为分类方法

论坛帖子对话行为分类可以明确每个帖子在当前线索中的角色,有助于重构论坛线索中的对话关系,提高论坛信息检索的效果。该文提出了一种基于弱监督学习的论坛帖子对话行为分类方法,把帖子的对话行为分类作为线索的序列标注问题来解决。该方法的特点是只要指定合理的特征约束,就可以训练对话行为分类模型。方法在CNET和edX数据集上的分类精确率分别达到75.6%和60.7%,优于有监督的条件随机域方法。     

一种基于窗口机制的口语理解异构图网络

口语理解(spoken language understanding, SLU)是面向任务的对话系统的核心组成部分,旨在提取用户查询的语义框架.在对话系统中,口语理解组件(SLU)负责识别用户的请求,并创建总结用户需求的语义框架, SLU通常包括两个子任务:意图检测(intent detection, ID)和槽位填充(slot filling, SF).意图检测是一个语义话语分类问题,在句子层面分析话语的语义;槽位填充是一个序列标注任务,在词级层面分析话语的语义.由于意图和槽之间的密切相关性,主流的工作采用联合模型来利用跨任务的共享知识.但是ID和SF是两个具有强相关性的不同任务,它们分别表征了话语的句级语义信息和词级信息,这意味着两个任务的信息是异构的,同时具有不同的粒度.提出一种用于联合意图检测和槽位填充的异构交互结构,采用自注意力和图注意力网络的联合形式充分地捕捉两个相关任务中异构信息的句级语义信息和词级信息之间的关系.不同于普通的同构结构,所提模型是一个包含不同类型节点和连接的异构图架构,因为异构图涉及更全面的信息和丰富的语义,同时可以更好地交互表征不同粒度节点之间的信息.此...     

COMODA: a conversation model for database access

As electronic database technology becomes less expensive, people will want to access information without undergoing special training. These people could use their native language if databases could be accessed through natural language conversations. The approach of the current research is that in order for the computer to be controlled by natural language, the computer does not have to understand it, only respond correctly. The conversation model for database access (COMODA) describes information retrieval as a dialogue. The dialogue is modelled by a series of states, where each state has an utterance that provides some information. The states are linked by transitions that are followed if a parse template matches the input sentence. Provisions are made for backtracking to earlier states, and for changes in topic. A small database of general information about one division of the Federal Government was implemented on an IBM-PC using these principles. When ten untrained people were allowed to converse with this database, 59% of their queries were answered correctly. All but one person said that they would use this type of database if more information was available. It was concluded that it is feasible to create a database of general information which can be accessed with natural language conversations by untrained users.     

基于回复生成的对话意图预测

随着人机对话系统的不断发展,让计算机能够准确理解对话者的对话意图,并根据对话的历史信息对回复进行意图预测,对于人机对话系统有着十分重要的意义。已有研究重点关注根据对话文本和已有标签对回复进行意图预测,但是,在很多场景下回复可能并没有生成。因此,文中提出了一种结合回复生成的对话意图预测模型。在生成部分,使用Seq2Seq结构,根据对话历史信息生成文本,作为对话中未来回复的文本信息;在分类部分,利用LSTM模型,将生成的回复文本与已有的对话信息转变为子句级别的表示,并结合注意力机制突出同一轮次对话句与生成回复的联系。实验结果表明,所提出的模型相比简单基线模型取得了2.54%的F1-score提升,并且联合训练的方式有助于提升模型性能。     

Multifunctionality in dialogue

This paper studies the multifunctionality of dialogue utterances, i.e. the phenomenon that utterances in dialogue often have more than one communicative function. It is argued that this phenomenon can be explained by analyzing the participation in dialogue as involving the performance of several types of activity in parallel, relating to different dimensions of communication. The multifunctionality of dialogue utterances is studied by (1) redefining the notion of ‘utterance’ in a rigorous manner (calling the revised notion ‘functional segment’), and (2) empirically investigating the multifunctionality of functional segments in a corpus of dialogues, annotated with a rich, multidimensional annotation schema. It is shown that, when communicative functions are assigned to functional segments, thereby eliminating every form of segmentation-related multifunctionality, an average multifunctionality is found between 1.8 and 3.6, depending on what is considered to count as a segment's communicative function. Moreover, a good understanding of the nature of the relations among the various multiple functions that a segment may have, and of the relations between functional segments and other units in dialogue segmentation, opens the way for defining a multidimensional computational update semantics for dialogue interpretation.     

Cooperative Understanding of Utterances and Gestures in a Dialogue-Based Problem Solving System

A dialogue plays an important role in learning how to solve a problem and form a concept. We are developing a problem solving and knowledge acquisition system based on co-reference between drill texts and dialogue with a teacher, focusing on first-grade mathematics. This paper presents a method of cooperative understanding of utterances and gestures within dialogue. We first describe our system design principles, which provide the basis for the integration of multimodal information during a dialogue. We define a principle of complementarity, explain its implementation, and describe the architecture of the problem solving system. We then show how to integrate our algorithms for utterance and gesture analysis within that software architecture. A feature-based approach is used for gesture recognition, derived from a sequence of images arising during the cooperative analysis of utterances. We conclude with an evaluation of the system against the design principles.     

口语对话中的语句主题分析

本文研究如何根据浅层的语义分析确定自然口语对话中的语句主题。首先将对话中的语句主题定义为说话者所关注的显著语义实体,并讨论了这样的语句主题所具有的两个特点(即话语性和连续性) 以及语句主题跟(扩展) 句子类型的关系(因而也介绍了句子类型及其扩展和扩展句子类型的识别) 。然后根据这些建立了语句主题分析算法,并在实际的对话语料中进行分析。实验结果表明,语句主题的分析正确率可达到6111～8716 % ,取决于不同的扩展句子类型和不同的正确率定义。