Stochastic Finite-State Models for Spoken Language Machine Translation

The problem of machine translation can be viewed as consisting of twosubproblems (a) lexical selection and (b) lexical reordering. In thispaper, we propose stochastic finite-state models for these two subproblems. Stochastic finite-state models are efficiently learnablefrom data, effective for decoding and are associated with a calculusfor composing models which allows for tight integration of constraintsfrom various levels of language processing. We present a method forlearning stochastic finite-state models for lexical selection andlexical reordering that are trained automatically from pairs of sourceand target utterances. We use this method to develop models forEnglish–Japanese and English–SPANISH translation and present the performance of these models for translation on speech and text. We also evaluate the efficacy of such a translation model in the context of a call routing task of unconstrained speech utterances.     

Head-Transducer Models for Speech Translation and Their Automatic Acquisition from Bilingual Data

This article presents statistical language translation models,called dependency transduction models, based on collectionsof head transducers. Head transducers are middle-out finite-state transducers which translate a head word in a source stringinto its corresponding head in the target language, and furthertranslate sequences of dependents of the source head into sequencesof dependents of the target head. The models are intended to capturethe lexical sensitivity of direct statistical translation models,while at the same time taking account of the hierarchical phrasalstructure of language. Head transducers are suitable for directrecursive lexical translation, and are simple enough to be trainedfully automatically. We present a method for fully automatictraining of dependency transduction models for which the only inputis transcribed and translated speech utterances. The method has beenapplied to create English–Spanish and English–Japanese translationmodels for speech translation applications. The dependencytransduction model gives around 75% accuracy for an English–Spanishtranslation task (using a simple string edit-distance measure) and70% for an English–Japanese translation task. Enhanced with targetn-grams and a case-based component, English–Spanish accuracy is over76%; for English–Japanese it is 73% for transcribed speech, and60% for translation from recognition word lattices.     

Learning finite-state models for machine translation

In formal language theory, finite-state transducers are well-know models for simple “input-output” mappings between two languages. Even if more powerful, recursive models can be used to account for more complex mappings, it has been argued that the input-output relations underlying most usual natural language pairs can essentially be modeled by finite-state devices. Moreover, the relative simplicity of these mappings has recently led to the development of techniques for learning finite-state transducers from a training set of input-output sentence pairs of the languages considered. In the last years, these techniques have lead to the development of a number of machine translation systems. Under the statistical statement of machine translation, we overview here how modeling, learning and search problems can be solved by using stochastic finite-state transducers. We also review the results achieved by the systems we have developed under this paradigm. As a main conclusion of this review we argue that, as task complexity and training data scarcity increase, those systems which rely more on statistical techniques tend produce the best results. This work was partially supported by the European Union project TT2 (IST-2001-32091) and by the Spanish project ITEFTE (TIC 2003-08681-C02-02). Editor: Georgios Paliouras and Yasubumi Sakakibara     

口语自动翻译系统技术评析

近几年来,随着信息技术的发展,口语自动翻译技术成为新的研究热点。目前国际上一些著名大学和研究机构甚至企业,都纷纷加入这一高技术的竞争行列,我国在相关技术方面也进行了卓有成效的研究。本文对目前自动口语翻译研究的技术现状进行了全面综述和分析,并对一些具体问题作了深入探讨。作者希望本文作出的分析和讨论的问题,能够对我国的自动口语翻译研究提供有益的参考。     

A Context-Aware Language Model for Spoken Query Retrieval

This paper concentrates on the problem of designing and developing a spoken query retrieval (SQR) system to access large document databases via voice. The main challenge is to identify and address issues related to the adaptation and scalability of integrating automatic speech recognition (ASR) and information retrieval (IR). In this paper, a Context Aware Language Model (CALM) framework allowing information retrieval to large document databases via voice is presented and findings from a research study using the framework will be discussed as well.     

人机互助的交互式口语翻译方法

基于短语的统计翻译模型是目前机器翻译领域广泛使用的模型之一。但是,由于在解码时采用短语精确匹配的策略,造成了严重的数据稀疏问题,短语表中的大量短语无法得到充分利用。为此,该文提出了人机互助的交互式翻译方法。对于翻译短语表中找不到的短语,首先通过模糊匹配的方法,在短语表中寻找与其相似的短语。然后利用组合分类器,判断哪些相似短语可能提高句子的翻译质量。最后,通过人机交互的方法,选择可能提高翻译质量且保持原句语义的短语。在口语语料上的实验结果证明,这种方法可以有效地提高翻译系统的译文质量。     

Inference of finite-state transducers from regular languages

Finite-state transducers are models that are being used in different areas of pattern recognition and computational linguistics. One of these areas is machine translation, where the approaches that are based on building models automatically from training examples are becoming more and more attractive. Finite-state transducers are very adequate to be used in constrained tasks where training samples of pairs of sentences are available. A technique to infer finite-state transducers is proposed in this work. This technique is based on formal relations between finite-state transducers and finite-state grammars. Given a training corpus of input-output pairs of sentences, the proposed approach uses statistical alignment methods to produce a set of conventional strings from which a stochastic finite-state grammar is inferred. This grammar is finally transformed into a resulting finite-state transducer. The proposed methods are assessed through series of machine translation experiments within the framework of the EUTRANS project.     

The Role of Evaluation in the Development of Spoken Language Systems

In this article, several criteria and paradigms are described tomeasure the performance of spoken language systems developed in theframework of national and international research projects. Theseevaluations are carried out in the domain of spontaneous human-humaninteraction as supported by machine translation systems. They are alsoapplied in the domain of spontaneous human-machine interactiontypically used in information retrieval applications. Some evaluationparadigms are discussed in more detail. It is also shown that officialperformance tests and site-specific evaluation criteria arecomplementary in use.     

对话行为信息在口语翻译中的应用

在口语翻译中,如何融入语义及语用信息一直是目前研究的难点之一。对话行为作为浅层话语结构描述的特征,近年来陆续应用于不同类型的翻译系统中。该文在介绍对话行为理论和口语标注语料的基础上,以基于短语的统计翻译系统为应用对象,提出了对话行为应用于翻译过程的三种方式。该方法通过对对话行为的自动分类,使训练语料—测试语料、开发集—测试集、源语言—目标语言的一致性得到提高,提高了翻译系统的性能,使最终的翻译结果可以更准确地反映源语言所要表达的对话意图。在汉英口语翻译评测数据上的实验证明,对话行为信息的加入使翻译系统的性能得到了有效的提高。     

基于双语对齐口语语料的翻译词典的自动生成

提出了一个基于英汉双语口语对齐语料库的翻译词典的自动生成算法，首先利用释义词典过渡双语文本，得到“过滤词典”，继而通过统计共现概率，计算出所有词对的相互关联值，并且生成“汉英（英汉）相互关联值表”，对于每个源语词汇选取相互关联值最大的若干项目标误作为候选词对，分别赋予信任值1，然后统计每个候选词对人信任值作为翻译词典的分级标准，得到4个不同级别的词典，其中“过滤词典＋4级词典”在召回率为93．5％的情况下，正确率达到93．389％。     

Toward Practical Spoken Language Translation

This paper argues that the time is now right to field practical Spoken Language Translation (SLT) systems. Several sorts of practical systems can be built over the next few years if system builders recognize that, at the present state of the art, users must cooperate and compromise with the programs. Further, SLT systems can be arranged on a scale, in terms of the degree of cooperation or compromise they require from users. In general, the broader the intended linguistic or topical coverage of a system, the more user cooperation or compromise it will presently require. The paper briefly discusses the component technologies of SLT systems as they relate to user cooperation and accommodation (“human factors engineering”), with examples from the authors’ work. It describes three classes of “cooperative” SLT systems which could be put into practical use during the next few years.All trademarks are hereby acknowledged. All URLs last accessed between 6th and 25th January 2006.     

The Janus-III Translation System: Speech-to-Speech Translation in Multiple Domains

The Janus-III system translates spoken languages in limiteddomains. The current research focus is on expanding beyond tasksinvolving a single limited semantic domain to significantly broaderand richer domains. To achieve this goal, The MT components of oursystem have been engineered to build and manipulate multi-domain parselattices that are based on modular grammars for multiple semanticdomains. This approach yields solutions to several problems includingmulti-domain disambiguation, segmentation of spoken utterances intosentence units, modularity of system design, and re-use of earliersystems with incompatible output.     

跨模态信息融合的端到端语音翻译

语音翻译旨在将一种语言的语音翻译成另一种语言的语音或文本. 相比于级联式翻译系统, 端到端的语音翻译方法具有时间延迟低、错误累积少和存储空间小等优势, 因此越来越多地受到研究者们的关注. 但是, 端到端的语音翻译方法不仅需要处理较长的语音序列, 提取其中的声学信息, 而且需要学习源语言语音和目标语言文本之间的对齐关系, 从而导致建模困难, 且性能欠佳. 提出一种跨模态信息融合的端到端的语音翻译方法, 该方法将文本机器翻译与语音翻译模型深度结合, 针对语音序列长度与文本序列长度不一致的问题, 通过过滤声学表示中的冗余信息, 使过滤后的声学状态序列长度与对应的文本序列尽可能一致; 针对对齐关系难学习的问题, 采用基于参数共享的方法将文本机器翻译模型嵌入到语音翻译模型中, 并通过多任务训练方法学习源语言语音与目标语言文本之间的对齐关系. 在公开的语音翻译数据集上进行的实验表明, 所提方法可以显著提升语音翻译的性能.     

The Syrinx Spoken Language System

This paper describes the Syrinx Spoken Language System (Sylan), an automated dialogue system that is fully integrated with the Syrinx Large Vocabulary Speech Recogniser (Sycon) into the Syrinx SpeechMaster platform. This platform combines speech recognition, natural language processing, dialogue management, telephony and database integration into a robust and flexible Voice User Interface that permits the deployment of natural language dialogue systems in automated call centres. We first describe the architecture of Sylan which, being modular, allows us to build a system whose domain-independent components are reusable from application to application. We then present those components from the point of view of application developers, describing the data structures used by the system and the utilities to build them. The two prototypes that have already been developed using Sylan are briefly presented, and we conclude by drawing the lessons learned along the way and pointing to further research directions.     

System Combination for Machine Translation of Spoken and Written Language

This paper describes an approach for computing a consensus translation from the outputs of multiple machine translation (MT) systems. The consensus translation is computed by weighted majority voting on a confusion network, similarly to the well-established ROVER approach of Fiscus for combining speech recognition hypotheses. To create the confusion network, pairwise word alignments of the original MT hypotheses are learned using an enhanced statistical alignment algorithm that explicitly models word reordering. The context of a whole corpus of automatic translations rather than a single sentence is taken into account in order to achieve high alignment quality. The confusion network is rescored with a special language model, and the consensus translation is extracted as the best path. The proposed system combination approach was evaluated in the framework of the TC-STAR speech translation project. Up to six state-of-the-art statistical phrase-based translation systems from different project partners were combined in the experiments. Significant improvements in translation quality from Spanish to English and from English to Spanish in comparison with the best of the individual MT systems were achieved under official evaluation conditions.      

Translation Disambiguation in Mixed Language Queries

Code-switching is very common among bilingual speakers. Spoken queries by these speakers are typically in mixed language. In this paper, we propose an unsupervised method for mixed-language query understanding, using only a monolingual corpus and a bilingual dictionary. Secondary-language words mixed in a primary-language query are translated into words in the primary language. We found that using a single disambiguation feature for translation is more effective than using multiple features, provided this feature is based on the most salient seed-word, chosen automatically by confidence scoring. We propose and compare four types of disambiguation features that are based on context seed-words. A baseline method uses the nearest neighboring seed-word as disambiguation feature. Multiple-context seed-word voting is also proposed in order to enlarge the context window. On the other hand, merely using the inverse-distance as weights on context words degrades the performance as it runs counter to the potential underlying syntactic relations between words. Our final proposal is a solution that uses multiple-context seed-words and the translation candidates of all mixed language words to select a single most salient seed-word for translation disambiguation. The translation disambiguation accuracy for this feature is at 83.7% for all words in the ATIS spontaneous speech query database, and 66.7% for content words.     

Interactive Speech Translation in the Diplomat Project

The Diplomat rapid-deployment speech-translation systemis intended to allow naï ve users to communicate across a languagebarrier, without strong domain restrictions, despite the error-pronenature of current speech and translation technologies. In addition,it should be deployable for new languages an order of magnitude morequickly than traditional technologies. Achieving this ambitious setof goals depends in large part on allowing the users to correct recognition and translation errors interactively. We present the Multi-Engine Machine Translation (MEMT) architecture, describing how it is well suited for such an application. We then discuss ourapproaches to rapid-deployment speech recognition and synthesis.Finally we describe our incorporation of interactive error correctionthroughout the system design. We have already developed workingbidirectional Croatian English and Spanish English systems, and have Haitian Creole English and Korean English versions under development.     

Unit Completion for a Computer-aided Translation Typing System

This work is in the context of, a system thatwatches over the users as they type a translation andrepeatedly suggests completions for the text already entered.The users may either accept, modify, or ignore these suggestions. Wedescribe the design, implementation, and performance of aprototype which suggests completions of units of texts that arelonger than one word.     

基于知识库的语言翻译系统研究

语言翻译是工作和学习中经常要做的事,传统的语言翻译是基于词法结构分析,然而自然语言并不是那么的规范,这导致了这种翻译方式存在本质上的缺陷,不管算法如何改进,翻译结果和人工翻译都会相差很远。本文以英汉翻译为例介绍了基于知识库的语言翻译,详细讲解了基于知识库翻译的原理,以及这种翻译方式与传统的基于词法结构分析的翻译方式相比具有的优越性,并实现了一个简单的基于知识库的翻译软件。