Recursive autoencoders (RAEs) for compositionality of a vector space model were applied to utterance intent classification of a smartphone-based Japanese-language spoken dialogue system. Though the RAEs express a nonlinear operation on the vectors of child nodes, the operation is considered to be different intrinsically depending on types of child nodes. To relax the difference, a data-driven untying of autoencoders (AEs) is proposed. The experimental result of the utterance intent classification showed an improved accuracy with the proposed method compared with the basic tied RAE and untied RAE based on a manual rule.

We demonstrate a bilingual robot application, WikiTalk, that can talk fluently in both English and Japanese about almost any topic using information from English and Japanese Wikipedias. The English version of the system has been demonstrated previously, but we now present a live demo with a Nao robot that speaks English and Japanese and switches language on request. The robot supports the verbal interaction with face-tracking, nodding and communicative gesturing. One of the key features of the WikiTalk system is that the robot can switch from the current topic to related topics during the interaction in order to navigate around Wikipedia following the user’s individual interests.

Dialogue robots are attractive to people, and in language learning systems, they motivate learners and let them practice conversational skills in more realistic environment. However, automatic speech recognition (ASR) of the second language (L2) learners is still a challenge, because their speech contains not just pronouncing, lexical, grammatical errors, but is sometimes totally disordered. Hence, we propose a novel robot assisted language learning (RALL) system using two robots, one as a teacher and the other as an advanced learner. The system is designed to simulate multiparty conversation, expecting implicit learning and enhancement of predictability of learners’ utterance through an alignment similar to “interactive alignment”, which is observed in human-human conversation. We collected a database with the prototypes, and measured how much the alignment phenomenon observed in the database with initial analysis.

The listener’s gazing activities during utterances were analyzed in a face-to-face three-party conversation setting. The function of each utterance was categorized according to the Grounding Acts defined by Traum (Traum, 1994) so that gazes during utterances could be analyzed from the viewpoint of grounding in communication (Clark, 1996). Quantitative analysis showed that the listeners were gazing at the speakers more in the second language (L2) conversation than in the native language (L1) conversation during the utterances that added new pieces of information, suggesting that they are using visual information to compensate for their lack of linguistic proficiency in L2 conversation.

This paper describes a method of generating a reduced phoneme set for dialogue-based computer assisted language learning (CALL)systems. We designed a reduced phoneme set consisting of classified phonemes more aligned with the learners’ speech characteristics than the canonical set of a target language. This reduced phoneme set provides an inherently more appropriate model for dealing with mispronunciation by second language speakers. In this study, we used a phonetic decision tree (PDT)-based top-down sequential splitting method to generate the reduced phoneme set and then applied this method to a translation-game type English CALL system for Japanese to determine its effectiveness. Experimental results showed that the proposed method improves the performance of recognizing non-native speech.

We developed a dialogue-based tutoring system for teaching English to Japanese students and plan to transfer the current software tutoring agent into an embodied robot in the hope that the robot will enrich conversation by allowing more natural interactions in small group learning situations. To enable smooth communication between an intelligent agent and the user, the agent must have realistic models on when to take turns, when to interrupt, and how to catch the partner's attention. For developing the realistic models applicable for computer assisted language learning systems, we also need to consider the differences between the mother tongue and second language that affect communication style. We collected a multimodal corpus of multi-party conversations in English as the second language to investigate the differences in communication styles. We describe our multimodal corpus and explore features of communication style e.g. filled pauses, and non-verbal information, such as eye-gaze, which show different characteristics between the mother tongue and second language.

Some big languages like English are spoken by a lot of people whose mother tongues are different from. Their second languages often have not only distinct accent but also different lexical and syntactic characteristics. Speech recognition performance is severely affected when the lexical, syntactic, or semantic characteristics in the training and recognition tasks differ. Language model of a speech recognition system is usually trained with transcribed speech data or text data collected in English native countries, therefore, speech recognition performance is expected to be degraded by mismatch of lexical and syntactic characteristics between native speakers and second language speakers as well as the distinction between their accents. The aim of language model adaptation is to exploit specific, albeit limited, knowledge about the recognition task to compensate for mismatch of the lexical, syntactic, or semantic characteristics. This paper describes whether the language model adaptation is effective for compensating for the mismatch between the lexical, syntactic, or semantic characteristics of native speakers and second language speakers.

This paper proposes a method for integrating example-based and rule-based machine translation systems with statistical methods. It extends a greedy decoder for statistical machine translation (SMT), which searches for an optimal translation by using SMT models starting from a decoder seed, i.e., the source language input paired with an initial translation hypothesis. In order to reduce local optima problems inherent in the search, the outputs generated by multiple translation engines, such as rule-based (RBMT) and example-based (EBMT) systems, are utilized as the initial translation hypotheses. This method outperforms conventional greedy decoding approaches using initial translation hypotheses based on translation examples retrieved from a parallel text corpus. However, the decoding of multiple initial translation hypotheses is computationally expensive. This paper proposes a method to select a single initial translation hypothesis before decoding based on a machine learning approach that judges the appropriateness of multiple initial translation hypotheses and selects the most confident one for decoding. Our approach is evaluated for the translation of dialogues in the travel domain, and the results show that it drastically reduces computational costs without a loss in translation quality.

This paper experimentally compares two automatic evaluators, RED and BLEU, to determine how close the evaluation results of each automatic evaluator are to average evaluation results by human evaluators, following the ATR standard of MT evaluation. This paper gives several cautionary remarks intended to prevent MT developers from drawing misleading conclusions when using the automatic evaluators. In addition, this paper reports a way of using the automatic evaluators so that their results agree with those of human evaluators.

The main goal of the present paper is to propose new schemes for the overall evaluation of a speech translation system. These schemes are expected to support and improve the design of the target application system, and precisely determine its performance. Experiments are conducted on the Japanese-to-English speech translation system ATR-MATRIX, which was developed at ATR Interpreting Telecommunications Research Laboratories. In the proposed schemes, the system’s translations are compared with those of a native Japanese taking the Test of English for International Communication (TOEIC), which is used as a measure of one’s speech translation capability. Subjective and automatic comparisons are made and the results are compared. A regression analysis on the subjective results shows that the speech translation capability of ATR-MATRIX matches a Japanese person scoring around 500 on the TOEIC. The automatic comparisons also show promising results.

An automatic translation quality evaluation method is proposed. In the proposed method, a parallel corpus is used to query translation answer candidates. The translation output is evaluated by measuring the similarity between the translation output and translation answer candidates with DP matching. This method evaluates a language translation subsystem of the Japanese-to-English ATR-MATRIX speech translation system developed at ATR Interpreting Telecommunications Research Laboratories. Discriminant analysis is then carried out to examine the evaluation performance of the proposed method. Experimental results show the effectiveness of the proposed method. The discriminant ratio is 83.5% for 2-class discrimination between absolutely correct and less appropriate translations classified subjectively. Also discussed are issues of the proposed method when it is applied to speech translation systems which inevitably make recognition errors.

ATR-MATRIX is a multi-lingual speech-to-speech translation system designed to facilitate communications between two parties of different languages engaged in a spontaneous conversation in a travel arrangement domain. In this paper, we propose a new evaluation method for speech translation systems. Our current focus is on measuring the robustness of a language translation sub-system, with quick calculation and low cost. Therefore, we calculate the difference between the translation output from transcription texts and the translation output from input speech by a dynamic programming method. We present the first trial experiment of this method applied to our Japanese-to-English speech translation system. We also provide related discussions on such points as error analysis and the relationship between the proposed method and translation quality evaluation manually done by humans.