Based on our previous work on example-based chat-oriented dialog systems that utilize a human-to-human conversation. Though promising, our previous simple retrieval techniques resulting a weakness on handling an out of vocabulary (OOV) database queries. In this paper we discuss an approach to increase the robustness of example-based dialog response retrieval. We employ a recursive neural network paraphrase identification technique to achieve a good performance on finding a good response in dialog-pair database. To achieve that, we remodel our previous dialog-pair database into distributed word representation. We apply an recursive autoencoders and dynamic pooling algorithm between user utterance and database to decide whether they have a same meaning or not, even when they composed in different word and structure. These distribute representations and recusive autoencoders have the potential to enhance our retrieval techniques and reduce confusion in example matching, especially when handling an OOV cases. We also present the system performance evaluation based on objective and subjective metrics.

In text and speech, there are various features that express the individuality of the writer or speaker. We proposed a method for transforming individuality using a technique inspired by statistical machine translation (SMT), and showed the effectiveness. In previous work, we proposed a method for paraphrasing for characteristic words using n-gram clustering. However, the method can be improved, because it considers only short context. In this paper, we propose a model of transforming individuality that considers longer contexts. To achieve this, we suggest adaptation of the language models and expansion of paraphrasing for characteristic words.

Rapport has been identi ed as an important function of human interaction to increase task performances. In this paper we leverage existing literature and a corpus of peer tutoring data to develop a framework able to explain how humans in dyadic interactions build, maintain, and break rapport through the use of specific conversational strategies that function to fulfil specific social goals, and that are instantiated in particular verbal and nonverbal behaviors.

When humans attempt to detect deception, they perform two actions: looking for telltale signs of deception, and asking questions to attempt to unveil a deceptive conversational partner. There has been signi cant prior work on automatic deception detection that attempts to learn signs of deception. On the other hand, we focus on the second action, envisioning a dialogue systems that asks questions to attempt to catch a potential liar. In this paper, we describe the results of an initial analysis towards this goal, attempting to make clear which questions make the features of deception more salient. In order to do so, we collect a deceptive corpus in Japanese, our target language, perform an analysis of this corpus comparing with a similar English corpus, and perform an analysis of what kinds of questions result in a higher deception detection accuracy.

In this paper, we create a non-task-oriented dialogue agent \KELDIC" on Twitter, and anlyze communication between the agent and twitter users. On twitter, users can react to tweets from others in three ways, reply, adding to favorite and retweet. From the point of view of user reactions to KELDIC's tweets, we demonstrate statistical features of users' behaviour. The result of analysis indicates the possibility of quantitative evaluation of dialogue agents using users' reaction.

The evaluation measures for chat-oriented dialogue systems are required in order to effectively improve such systems. Some studies have evaluated systems with several arbitrarily de ned measures; however, it is not examined whether their measures are appropriate. We analyze evaluation measures for chat-oriented dialogue systems through the semantic differential. Our analysis shows that evaluation measures are clustered into four factors for each evaluator. The factors consist of two common factors, one resemble factor between evaluators, and one personal factor. We also develop an automatic evaluation system that estimates each evaluation measure defined in the semantic differential. Our experiment shows that the developed system estimates most of the scores with the similar correlation coefficients as between human evaluators.

We are developing a framework where a dialogue system asks questions for users and thus acquires information (cuisine of unknown restaurants) through dialogue. Such a question would be better if it is more concise and concrete. We propose a method to select an appropriate question on the basis of expected utility calculated for four question types: Yes/No, binary, ternary, and Wh- questions. We rst de ne utility for the four types by considering their appropriateness for users. We then de ne a probability that represents how likely a question contains a correct cuisine by using con dence measure (CM) of cuisine estimation. Then the sum totals of products of the utility and the probability are calculated as expected utility. As experiments, we compare several ways to integrate two basic CMs we proposed before. We also discuss adequacy of the utility values by using questionnaire results for 15 subjects.

Project Next NLP is a project to perform error analyses in the eld of natural language processing. In the dialogue task of the project, we are focusing on the error analysis of dialogue systems. This paper describes the ongoing effort of the dialogue task, including how the participants of the task collected chat dialogues by talking to the dialogue system based on NTT Docomo's chat API and how dialogue breakdowns are annotated for error analysis.

The purpose of this paper is to clarify how the decision making process is built in the discussion for problem solving. We let some university students taking a communication class carry out the group discussion of the fishbowl style to aim for the problem solving on given topic. One group consists of five or six the students. We record the discussion with audio and video as data and analyze it. Specifically, we annotate the three steps which are the cognition of the problem, the design of solution proposal, choice in the decision making process for the recorded data as three kind of labels. For the annotated data, we analyze how three steps constitute the decision making process, and how the transition of each step influence the discussion. Furthermore, we analysed the participants addressed consensus building on the transition of each step to clarify the action of the designing and the control of the participants for the decision making process. .Keyword decision making process, group discussion, problem-solving, annotation, consensus building

The appearance of unknown words often disturbs communication. Most spoken dialog systems deal with unknown words that are uttered by the user, but which are not covered by the system's vocabulary. In this paper, we focus on detecting unknown words from the user side in which the system utterance is unknown to the user. In particular, we develop a classi er based on Electroencephalography (EEG) signal from user's brain waves, including the use of absolute power and Event-Related Desynchronization (ERD) features. The results show that we could detect the characteristics of brain waves at the time of unknown word perception signi cantly better than the chance rate.

This paper introduces a network-based multilingual speech translation and spoken dialog system development tool kit (MCML Spoken Language Communication SDK) developed by NICT. This tool kit provides APIs and sample applications which enable to construct multilingual communication applications connecting to NICT's ASR, TTS, MT and DM server with MCML communication protocol. We are planning to release a new version of SDK including the DM function (DM Builder). In this paper, we illustrate the outline of the MCML SLC SDK and the DM Builder.

iNAGO's platform product netpeople is the launch pad that enables businesses to offer a smart assistant in the cloud for any device and service. netpeople provides actual human-like conversation and interaction with exclusive Content-Aware Goal-Oriented technologies that work together in unison.

A chatter bot has been developed and released on a communication service(LINE). The character of chatter bot is named "パン田一郎"(Panda Ichiro) who is an official character of a job hunting service (fromA). The main function of LINE Panda Ichiro is chat. LINE panda Ichiro responds a static sentence from its database. The number of sentence in the database is above 10,000. Furthermore, some practical functions such as weather report are implemented. LINE Panda Ichiro was registered by eight millions users only for three months. The user log data of LINE Panda Ichiro was analyzed for a revealing the attractive point of the chatter bot for users. We confirmed that a simple system of LINE Panda Ichiro's chat could entertain people.

We demonstrate a spoken dialogue system for information navigation that is adaptive to the information demand of users. The system navigates daily news on Web in a conversational manner. The dialogue management of the system tracks user focus to provide information with accordance of the user interests. The system can be adapted to a variety of domains of news because the modules of the system are trained with an unsupervised manner.

Onsei Assist" by Yahoo Japan is a voice dialog application for mobile devices that enables us to retrieve various information including train route, weather, news, web pages, map, shops and general facts. It also supports mobile device operations such as alarm setting, application launch, dialing, calendar and address book search.

NTT DOCOMO developed a voice agent system which is intended to be used in automobiles. The system monitors vehicle's information such as brake operations, and spontaneously begins voice conversation when it detects the events such as "sudden brake". Moreover, users can begin talk with the system without tapping smartphone or ressing button on steering(Hands-free voice recognition). This paper describers the system's architecture, and process to generate system's responses.

We demonstrate this dialogue system on REGZA Z10X we released in Oct. 2014. This system is allowed to retrieve TV contents such as programs and these scenes through spontaneous speeches and dialogues. One of characteristic points is that it utilizes states of TV such as views and settings to understand a user's intent. Another point is that it has a new words acquisition system to understand frequent-increasing keywords such as TV titles. We demonstrate this dialogue system on REGZA Z10X we released in Oct. 2014.