Yukie Nakao

Almost Flat Functional Semantics for Speech Translation

We introduce a novel semantic representation formalism, Almost Flat Functional semantics (AFF), which is designed as an intelligent compromise between linguistically motivated predicate/argument semantics and ad hoc engineering solutions based on flat feature/value lists; the central idea is to tag each semantic element with the functional marking which most closely surrounds it. We argue that AFF is well-suited for medium-vocabulary speech translation applications, and describe simple and general algorithms for parsing, generating and performing transfer using AFF representations. The formalism has been fully implemented within a medium-vocabulary interlingua-based Open Source speech translation system which translates between English, French, Japanese and Arabic.

Japanese Speech Understanding using Grammar Specialization

The most common speech understanding architecture for spoken dialogue systems is a combination of speech recognition based on a class N-gram language model, and robust parsing. For many types of applications, however, grammar-based recognition can offer concrete advantages. Training a good class N-gram language model requires substantial quantities of corpus data, which is generally not available at the start of a new project. Head-to-head comparisons of class N-gram/robust and grammar-based systems also suggest that users who are familiar with system coverage get better results from grammar-based architectures (Knight et al., 2001). As a consequence, deployed spoken dialogue systems for real-world applications frequently use grammar-based methods. This is particularly the case for speech translation systems. Although leading research systems like Verbmobil and NE-SPOLE! (Wahlster, 2000; Lavie et al., 2001) usually employ complex architectures combining statistical and rule-based methods, successful practical examples like Phraselator and S-MINDS (Phraselator, 2005; Sehda, 2005) are typically phrasal translators with grammar-based recognizers.

MedSLT: A Limited-Domain Unidirectional Grammar-Based Medical Speech Translator

MedSLT is a unidirectional medical speech translation system intended for use in doctor-patient diagnosis dialogues, which provides coverage of several different language pairs and subdomains. Vocabulary ranges from about 350 to 1000 surface words, depending on the language and subdomain. We will demo both the system itself and the development environment, which uses a combination of rule-based and data-driven methods to construct efficient recognisers, generators and transfer rule sets from small corpora.

Using Artificially Generated Data to Evaluate Statistical Machine Translation

Although Statistical Machine Translation (SMT) is now the dominant paradigm within Machine Translation, we argue that it is far from clear that it can outperform Rule-Based Machine Translation (RBMT) on small- to medium-vocabulary applications where high precision is more important than recall. A particularly important practical example is medical speech translation. We report the results of experiments where we configured the various grammars and rule-sets in an Open Source medium-vocabulary multi-lingual medical speech translation system to generate large aligned bilingual corpora for English → French and English → Japanese, which were then used to train SMT models based on the common combination of Giza++, Moses and SRILM. The resulting SMTs were unable fully to reproduce the performance of the RBMT, with performance topping out, even for English → French, with less than 70% of the SMT translations of previously unseen sentences agreeing with RBMT translations. When the outputs of the two systems differed, human judges reported the SMT result as frequently being worse than the RBMT result, and hardly ever better; moreover, the added robustness of the SMT only yielded a small improvement in recall, with a large penalty in precision.

A Model of Discourse Segmentation and Segment Title Assignment for Lecture Speech Indexing

Dividing a lecture speech into segments and providing those segments as learning objects are quite general and convenient way to construct e-learning resources. However it is difficult to assign an appropriate title to each object that reflects its content. Since there are various aspects of analyzing discourse segments, it is inevitable that researchers will face the diversity when describing the “meanings” of discourse segments. In this paper, we propose the assignment of discourse segment titles from the representation of their “meanings.” In this assigning procedure, we focus on the speakers evaluation for the event or the speech object. To verify the effectiveness of our idea, we examined identification of the segment boundaries from the titles that were described in our procedure. We confirmed that the result of the identification was more accurate than that of intuitive identification.