We design an attentive listening agent which can generate flexible backchannels. In our previous work, we analyzed the morphological forms (category) of backchannels by focusing on their relationship with the syntactic structure in the preceding utterances. In this paper, based on the analysis, we conduct machine learning to predict the category of backchannels using features of the preceding utterance. At first, we annotated all acceptable backchannel categories for every backchannel occurrence and regard them as a ``correct'' label for the reference in evaluating prediction of the backchannel category. This annotation also gives a good insight on the relationship between backchannel forms. The results of the prediction suggest that we can choose appropriate backchannels depending on the preceding utterance. The proposed model improved prediction accuracy in comparison with the baseline which always outputs the most frequent morphological form of backchannels. Furthermore, evaluations by human subjects show that our method obtained a significantly higher rating than the baseline method.

When a robot interacts with users in public spaces, it receives various sounds such as surrounding noises and users' voices, and furthermore needs to interact with multiple people at the same time. If it incorrectly determines whether it should respond to these sounds, it will erroneously respond to surrounding noises or ignores user utterances toward it. In this paper, we present a machine learning-based method to estimate a response obligation, i.e., whether an input sound should be responded to by the robot or not. This enables the robot to reject monologues and user utterances toward other users as well as noises. Our method uses not only acoustic information but also users' motions and postures during the input sound and user behaviors after the input sound as features. We demonstrate the new features significantly improved the estimation performance. We also investigate performances with various combinations of features and reveal that input sound classification results and a user's whole body motion are helpful for the estimation.

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 a significant amount of prior work on automatic deception detection, which focuses on the former. On the other hand, we focus on the latter, constructing a dialog system for an interview task that acts as an interviewer asking questions to attempt to catch a potentially deceptive interviewee. We propose several dialog strategies for this system, and measure the utterance-level deception detection accuracy of each, finding that a more intelligent dialog strategy results in slightly better deception detection accuracy.

This paper gives an overview on the spoken dialogue system for autonomous android Erica developed in the JST ERATO Ishiguro human robot interaction project. In order to realize human-like behaviors, the system introduces a mixed-initiative dialogue management and generation of natural backchannels and noddings, assuming a part to play in a social interaction. Natural interaction with real users is achieved by quick response generation with suppression of incorrect language understanding and also reacting with non-verbal behavior when the system fails in speech recognition or language understanding.

In this study, we examine the effects of using a game for encouraging the use of a spoken dialogue system. As a case study, we developed a word-chain game, called Shiritori in Japanese, and released the game as a module in a Japanese Android/iOS app, Onsei-Assist, which is a Siri-like personal assistant based on a spoken dialogue technology. We analyzed the log after the release and confirmed that the game can increase the number of user utterances. Furthermore, we discovered a positive side effect, in which users who have played the game tend to begin using non-game modules. This suggests that just adding a game module to the system can improve user engagement with an assistant agent.

Detecting breakdowns in dialogue can be one of the promising techniques in dialogue systems. To propel the research and development for dialogue breakdown detection, we organized the dialogue breakdown detection challenge. This paper describes the background of the challenge and the dataset, and overviews the methods and results of the submitted runs of the participants. We show to what extent dialogue breakdowns can be detected by current techniques and discuss limitations and problems to be tackled in the future.

It is difficult to detect a breakdown phenomena in the dialogue with a chat-oriented dialogue system because of the variety of causes of breakdown. To deal with this problem, we analyzed a chat dialogue corpus and made a taxonomy of the errors that yield the dialogue break-down. In this paper, we propose a breakdown detection method that consists of combinations of the classifiers for the different cause of errors based on the taxonomy.

This paper presents a new method which detects dialogue breakdowns automatically for the dialogue task of Project Next NLP. Our keyword extraction is performed by a morphological analyser using our customized dictionary, regarding dictionary matched morphemes as keywords. We use these keywords as an utterance forcus. We establish three rules which we found sheer variety of dialog breakdowns, based on our observation and genaralization of the given dialog data. We implemented detectors by these rules which tries to exclude possible latent ambiguity of anotations.

For detecting dialog breakdowns, we make it as a classification problem using features of sentences in the dialog. The feature including context information will perform better on the task. Recurrent neural networks (RNN) can encode sentences to fixed-length feature vectors, which are based on previous contexts. In this paper, we propose a multi-task learning method of language modeling and detecting dialog breakdowns for RNN and entended models. We conducted comparative experiments for our various RNN models and learning methods on dataset of Dialog dreakdown detection challenge, and show that our proposed model has very high precision, low recall, high accuracy and very small mean squared error for prediction of dialog breakdown labels compared to other participants.

We propose a dialogue breakdown detector based on the Recurrent Neural Netwoek (RNN). RNN is an extension of the neural network, which can consider old context information in sequential input. We construct a RNN network that receives utterance pairs of user and dialogue system as inputs, and outputs the breakdown label for each system's utterance. The input utterances are separated into vectors of words,cooccurrence words,and represented distributional sentence vectors. Finally, we combined these features and detector evaluated the breakdown detection accuracy.

Chat-oriented dialogue systems sometimes generate utterances that are inappropriate as the responses for user utterances and cause dialogue breakdown If a system can predict whether an utterance cause dialogue breakdown, it helps to continue dialogue with suppressing such inappropriate system utterances. In this paper, I develop a dialogue breakingdown detector and analyze the effects of training features, data and algorithms for dialogue breakdown detection performance.

Thip paper describes a method for dialogue breakdown detection using recurrent neural network with long short-term memory cells (LSTM-RNN). The proposed method uses a pair of system's utterance and preceding user's utterance for dialogue breakdown detection. Each utterances are converted into sequences of vector representation of word by word2vec and we use it for the input of the LSTM-RNN. In our model, we build two LSTM-RNNs, for processing user's utterances and system's uttearnces. The sequences of user's utterance and system's utterance are processed by each LSTM-RNNs and our model estimates distributions of annotations of dialogue breakdown by integrating each outputs. Experimenal results show that the proposed methods outperform the baseline method in detection of X and estimation of annotation distribution. However, in detection of [] and X, the performances of our methods are lower then the baseline method.

We have developed a natural-language dialogue platform aiming to enable whoever aims to create interactive communication services. Feature of this platform is that service developers can customize interactive services easily by combining various components. We produced a talking toy(OHaNAS) using this platform with Tomy Company, LTD.

Teneo is a unique platform that allows users to build complete, artificially intelligent natural language solutions that understand and act on what people are saying. It packages up the end-to-end process of understanding requirements, building the solution and analyzing it - to provide quantifiable, data driven optimization and customer insight - in a single, unified solution

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.

Due to the difficulties in voice recognition and natural language processing, it has been difficult for us to keep having a sufficient sense of conversation with a conversation robot. On the other hand, it has been demonstrated that we can make conversation between human and a robot without voice recognition by carefully designing the script for robot's utterances. However, it is difficult for a single robot to escape from collapse of conversation when the prepared replies do not match with the user's utterances. This study focuses on how to escape such collapses by using multiple robots. Through comparing to a single robot that talks to human with the same script, it is suggested that the multiple robots can provide human with more sense of conversation than the single robot. It is argued that the effect originates from subject's experience of observing conversation and switches of speaker role between robots, which is specific in multiple conversation.

We demonstrate a Voice Assistant System for inheritance procedures at banks released in Nov. 2014. This system obtains user's inheritance conditions through dialogue, then provides a list of required documents and prior procedures for inheritance procedures. Our dialogue platform enables the system to answer questions from user at any dialogue states without large-scale dialogue flows.

We propose a standard problem AIWolf game for artificial intelligence. This game is one of communication game around the table. This game called ``Are You a Werewolf?'', generaly. We has been thought this game is useful metrics for evaluating progress artificial intelligence. Moreover, our project held the competision at 2015

In the demonstration, a conversation robot that facilitates four-participant conversation and answers questions from other participants is shown. The participants consist of the robot and three persons, one host and two guests. The topic of conversation is introduction of the robot itself. The demonstration is realized from three key functions: (1) answering the question from the participant, (2) facilitating utterances of a participant who has less turns relatively than others, and (3) managing the addressee of the answer especially to the question which the host asks on behalf of the guests.

Recently, speech dialog systems have been used in many applications, including car navigators. However, because there are still many issues with usability and speech recognition accuracy, interactive speech has not yet become commonplace. One problem with usability is the limit placed on user timing. The user has to push a button or produce a magic word used to trigger recognition. The user must also refrain from speaking while the system is answering, playing music etc. We demonstrate a system which uses echo cancellation to process barge-in, removing constraints on the user's speech timing. For example, the user can interrupt the system in the middle of a lengthy guidance, allowing stress-free operation of the system via interactive speech.

In dialogue system, dialogue modeling is one of the most important factors contributing to user satisfaction. Especially in the Example-Based Dialogue Modeling (EBDM), effective method for selecting response utterances from examples improve dialogue quality. We describe an EBDM framework that predicts user satisfaction to select the best system response for the user from multiple response candidates. This framework select satisfactory response for user preference while prediction using user feedback can be predicts user satisfaction for system responses.