面向任务的基于深度学习的多轮对话系统与技术

自然语言是人类智慧的结晶,以自然语言的形式与计算机进行交互是人们长久以来的期待。随着自然语言处理技术的发展与深度学习方法的兴起,人机对话系统成为了新的研究热点。人机对话系统按照功能可以分为任务导向型对话系统、闲聊型对话系统、问答型对话系统。任务导向型对话系统是一种典型的人机对话系统,旨在帮助用户完成某些特定的任务,有着十分重要的学术意义和应用价值。文中系统地阐述了一种在实际工程应用中的任务导向型对话系统的通用框架,主要包括自然语言理解、对话管理以及自然语言生成3个部分;介绍了上述各部分所采用的经典深度学习和机器学习方法。最后,对自然语言理解任务进行了实证性的实验验证与分析,结果表明文中内容可以为任务导向型对话系统的构建提供有效指导。     

The TRAINS project: a case study in building a conversational planning agent

Abstract

The TRAINS project is an effort to build a conversationally proficient planning assistant. A key part of the project is the construction of the TRAINS system, which provides the research platform for a wide range of issues in natural language understanding, mixed-initiative planning systems, and representing and reasoning about time, actions and events. Four years have now passed since the beginning of the project. Each year a demonstration system has been produced that focused on a dialogue that illustrates particular aspects of the research. The commitment to building complete integrated systems is a significant overhead on the research, but it is considered essential to guarantee that the results constitute real progress in the field. This paper describes the goals of the project, and the experience with the effort so far.     

Usefulness,localizability, humanness,and language-benefit: additional evaluation criteria for natural language dialogue systems

Human–computer dialogue systems interact with human users using natural language. We used the ALICE/AIML chatbot architecture as a platform to develop a range of chatbots covering different languages, genres, text-types, and user-groups, to illustrate qualitative aspects of natural language dialogue system evaluation. We present some of the different evaluation techniques used in natural language dialogue systems, including black box and glass box, comparative, quantitative, and qualitative evaluation. Four aspects of NLP dialogue system evaluation are often overlooked: “usefulness” in terms of a user’s qualitative needs, “localizability” to new genres and languages, “humanness” or “naturalness” compared to human–human dialogues, and “language benefit” compared to alternative interfaces. We illustrated these aspects with respect to our work on machine-learnt chatbot dialogue systems; we believe these aspects are worthwhile in impressing potential new users and customers.     

Combining language models in the input interface of a spoken dialogue system

This paper presents a new technique to enhance the performance of the input interface of spoken dialogue systems based on a procedure that combines during speech recognition the advantages of using prompt-dependent language models with those of using a language model independent of the prompts generated by the dialogue system. The technique proposes to create a new speech recognizer, termed contextual speech recognizer, that uses a prompt-independent language model to allow recognizing any kind of sentence permitted in the application domain, and at the same time, uses contextual information (in the form of prompt-dependent language models) to take into account that some sentences are more likely to be uttered than others at a particular moment of the dialogue. The experiments show the technique allows enhancing clearly the performance of the input interface of a previously developed dialogue system based exclusively on prompt-dependent language models. But most important, in comparison with a standard speech recognizer that uses just one prompt-independent language model without contextual information, the proposed recognizer allows increasing the word accuracy and sentence understanding rates by 4.09% and 4.19% absolute, respectively. These scores are slightly better than those obtained using linear interpolation of the prompt-independent and prompt-dependent language models used in the experiments.     

Approaches to natural language discourse processing

One of the most difficult problems within the field of Artificial Intelligence (AI) is that of processing language by computer, or natural-language processing. A major problem in natural-language processing is to build theories and models of how individual utterances cling together into a coherent discourse. The problem is important because, to properly understand natural language, a computer should have some sense of what it means for a discourse to be coherent and rational. Theories, models and implementations of natural-language processing argue for a measure of coherence based on three themes: meaning, structure, and intention. Most approaches stress one theme over all the others. Their future lies in the integration of components of all approaches. A theory of intention analysis solves, in part, the problem of natural-language dialogue processing. A central principle of the theory is that coherence of natural-language dialogue can be modelled by analysing sequences of intention. The theory of intention analysis has been incorporated within a computational model, called Operating System CONsultant (OSCON), implemented in Quintus Prolog, which understands, and answers in English, English questions about computer operating systems. Theories and implementations of discourse processing will not only enable people to communicate better with computers, but also enable computers to better communicate with people.     

L0-The first five years of an automated language acquisition project

The L₀ project at ICSI and UC Berkeley attempts to combine not only vision and natural language modelling, but also learning. The original task was put forward in (Feldman et al. 1990a) as a touchstone task for AI and cognitive science. The task is to build a system that can learn the appropriate fragment of any natural language from sentence-picture pairs. We have not succeeded in building such a system, but we have made considerable progress on component subtasks and this has led in a number of productive and surprising directions.     

Deduction, Strategies, and Rewriting

Automated deduction methods should be specified not procedurally, but declaratively, as inference systems which are proved correct regardless of implementation details. Then, different algorithms to implement a given inference system should be specified as strategies to apply the inference rules. The inference rules themselves can be naturally specified as (possibly conditional) rewrite rules. Using a high-performance rewriting language implementation and a strategy language to guide rewriting computations, we can obtain in a modular way implementations of both the inference rules of automated deduction procedures and of algorithms controling their application. This paper presents the design of a strategy language for the Maude rewriting language that supports this modular decomposition: inference systems are specified in system modules, and strategies in strategy modules. We give a set-theoretic semantics for this strategy language, present its different combinators, illustrate its main ideas with several examples, and describe both a reflective prototype in Maude and an ongoing C++ implementation.     

Shaping user input: a strategy for natural language dialogue design

Traditional approaches to natural language dialogue interface design have adopted ordinary human-human conversation as the model for online human-computer interaction. The attempt to deal with all the subtleties of natural dialogues, such as topic focus, coherence, ellipsis, pronominal reference, etc. has resulted in prototype systems that are enormously complex and computationally expensive.

In a series of experiments^†, we explored ways of minimizing the processing burden of a dialogue system by channeling user input towards a more tractable, though still natural, form of Englishlanguage questions. Through linking a pair of terminals, we presented subjects with two different dialogue styles as a framework for online help in the domain of word-processing. The first dialogue style involved ordinary conversational format. The second style involved a simulation of an automated dialogue system, including apparent processing restrictions and ‘system process messages’ to inform the subject of the steps taken by the system during query analysis. In both cases human tutors played the role of the help system. After each dialogue session, subjects were interviewed to determine their assessments of the naturalness and usability of the dialogue interface.

We found that user input became more tractable to parsing and query analysis as the dialogue style became more formalized, yet the subjective assessment of naturalness and usability remained fairly constant. This suggests that techniques for channeling user input in a dialogue system may be effectively employed to reduce processing demands without compromising the benefits of a natural language interface. Theoretically, this data lends support to the hypothesis that unrestricted human-human conversation is not the most appropriate model for the design of human-computer dialogue interfaces.     

SWIM: A “Natural” Interface for the scientifically minded language learner

We describe a system under development, whose goal is to provide a “natural” environment for students learning to produce sentences in French. The learning objective is personal pronouns, the method is inductive (learning through exploration). Input of the learning component are conceptual structures (meanings) and the corresponding linguistic forms (sentences), its outputs are rules characterizing these data. The learning is dialogue based, that is to say, the student may ask certain kinds of questions such as:How does one say 〈idea〉?,Can one say 〈linguistic form〉?,Why does one say 〈linguistic form〉?, and the system answers them. By integrating the student into the process, that is, by encouraging him to build and explore a search space we hope to enhance not only his learning efficiency (what and how to learn), but also our understanding of the underlying processes. By analyzing the trace of the dialogue (what questions have been asked at what moment), we may infer the strategies a student put to use. Although the system covers far more than what is discussed here, we will restrict our discussion to a small subset of grammar, personal pronouns, which are known to be a notorious problem both in first and second language learning.     

MeteoBayes: Effective Plan Recognition in a Weather Dialogue System

When you've called a voice portal for any kind of information retrieval, chances are that an automated system guided the entire interaction. It might have correctly identified your goal, but probably only after asking too many questions. MeteoBayes is a meteorological information dialogue system that lets you use natural language to direct the interaction. Based on Bayesian networks, MeteoBayes' inference engine attempts to identify user intentions by consulting its past dialogue repository. For unfamiliar words, MeteoBayes has an unknown-term disambiguation module that learns word similarities from texts to avoid unnecessary system inquiries, thus speeding up the understanding process     

Can We Talk? Methods for Evaluation and Training of Spoken Dialogue Systems

There is a strong relationship between evaluation and methods for automatically training language processing systems, where generally the same resource and metrics are used both to train system components and to evaluate them. To date, in dialogue systems research, this general methodology is not typically applied to the dialogue manager and spoken language generator. However, any metric for evaluating system performance can be used as a feedback function for automatically training the system. This approach is motivated with examples of the application of reinforcement learning to dialogue manager optimization, and the use of boosting to train the spoken language generator.     

On the calculus of positively constructed formulas for automated theorem proving

The paper deals with an expressive logic language LF and its calculus. Formulas of this language consist of some large-scale structural elements, such as type quantifiers. The language LF contains only two logic symbols—∀ and ∃, which form the set of logic connectives of the language. The logic calculus JF and complete strategy for automated proof search based on a single unary rule of inference are considered. This calculus has a number of other features which lead to the reduction of the combinatorial complexity of finding the deductions in comparison to the known systems for automated theorem proving as the Resolution method and Genzen calculus. Problems of effective implementation of JF as a program system for automated theorem proving are considered.     

Case-based adaptation of workflows

This paper presents on a Case-based Reasoning approach for automated workflow adaptation by reuse of experience. Agile workflow technology allows structural adaptations of workflow instances at build time or at run time. The approach supports the expert in performing such adaptations by an automated method. The method employs workflow adaptation cases that record adaptation episodes from the past. The recorded changes can be automatically transferred to a new workflow that is in a similar situation of change. First, the notion of workflow adaptation cases is introduced. The sample workflow modeling language CFCN is presented, which has been developed by the University of Trier as a part of the agile workflow management system Cake. Then, the retrieval of adaptation cases is briefly discussed. The case-based adaptation method is explained including the so-called anchor mapping algorithm which identifies the parts of the target workflow where to apply the changes. A formative evaluation in two application domains compares different variants of the anchor mapping algorithm by means of experts assessing the results of the automated adaptation.     

基于WebKit浏览器的中间件系统架构

基于浏览器来构建中间件系统模型,可以使业务系统的应用部署更加灵活方便,同时也带来应用系统开发难度的降低。基于WebKit浏览器,通过增加平台访问的JavaScript接口,使应用系统构建在统一的脚本语言基础上,新的应用代码不需要编译就可以控制整个系统的行为和逻辑,是中间件应用系统模型的较好解决方案。     

Evaluation of Methods for the Determination of Factors Inducing Fatigue in Man at Work

The aim of the study was to determine the influence of textual feedback on the content and outcome of spoken interaction with a natural language dialogue system. More specifically, the assumption that textual feedback could disrupt spoken interaction was tested in a human–computer dialogue situation. In total, 48 adult participants, familiar with the system, had to find restaurants based on simple or difficult scenarios using a real natural language service system in a speech-only (phone), speech plus textual dialogue history (multimodal) or text-only (web) modality. The linguistic contents of the dialogues differed as a function of modality, but were similar whether the textual feedback was included in the spoken condition or not. These results add to burgeoning research efforts on multimodal feedback, in suggesting that textual feedback may have little or no detrimental effect on information searching with a real system.

Statement of Relevance: The results suggest that adding textual feedback to interfaces for human–computer dialogue could enhance spoken interaction rather than create interference. The literature currently suggests that adding textual feedback to tasks that depend on the visual sense benefits human–computer interaction. The addition of textual output when the spoken modality is heavily taxed by the task was investigated.     

XML queries and algebra in the Enosys integration platform

We describe the Enosys XML integration platform, focusing on the query language, algebra, and architecture of its query processor. The platform enables the development of eBusiness applications in customer relationship management, e-commerce, supply chain management, and decision support. These applications often require that data be integrated dynamically from multiple information sources. The Enosys platform allows one to build (virtual and/or materialized) integrated XML views of multiple sources, using XML queries as view definitions. During run-time, the application issues XML queries against the views. Queries and views are translated into the XCQL algebra and are combined into a single algebra expression/plan. Query plan composition and query plan decomposition challenges are faced in this process. Finally, the query processor lazily evaluates the result, using an appropriate adaptation of relational database iterator models to XML. The paper describes the platform architecture and components, the supported XML query language and the query processor architecture. It focuses on the underlying XML query algebra, which differs from the algebras that have been considered by W3C in that it is particularly tuned to semistructured data and to optimization and efficient evaluation in a system that follows the conventional architecture of database systems.