This paper describes the practice of knowledge construction using a domain ontology and procedural knowledge in guitar performance. Domain ontologies are created to understand and reason the domain knowledge. However, most of them are difficult to understand ontological logic and languages for domain experts. It is important to consider the different types of knowledge representation to connect ontology experts and domain experts. In this study, we verified the effect of the knowledge construction of a domain using a domain ontology and procedural knowledge. We focused on musical instrument performance and used the knowledge of classical guitar which requires many techniques. We practiced the following processes: (1)Constructed procedural knowledge described actions of each guitar rendition in detail, (2)Developed the Guitar Rendition Ontology based on the procedural knowledge, and (3)Reconstructed the procedural knowledge with seeing the ontology by domain experts, and carried out questionnaires to them. As a result, the knowledge was increased and modified appropriately, and we received positive answers about the usefulness of the method. Constructing the domain knowledge using the two types of knowledge representation enables following aspects: (1) Domain experts can construct procedural knowledge intuitively by defining description contents and format, (2) Ontology experts can efficiently construct a domain ontology based on the procedural knowledge and (3) Domain experts can re-construct procedural knowledge with higher consistency and knowledge processing by using domain ontology together.
In the area of the Semantic Web, the expressive description logic SROIQ corresponding to OWL2 provides us rich reasoning and learning tasks for ontologies, e.g., inference engine, query-answering system, and concept learning. However, unlike simple ontologies in RDF graphs, it is not easy for users to build ontologies using the logical and complex expressions of SROIQ. In this paper, we propose (i) minimal model reasoning in the description logic SROIQ for RDF graphs and (ii) a SROIQ-concept constructing algorithm for the classes, properties and individuals in each RDF graph. In the minimal models of RDF graphs based on the closed world assumption (CWA), we prove the completeness, soundness and complexity of the minimal model reasoning in the description logic SROIQ. We define decidable SROIQ-concept constructing in a unique interpretation of SROIQ-concepts based on the minimal model reasoning. For infinite SROIQ-concept combinations constructed by classes, properties and individuals (even less expressive description logic concepts), our constructing method removes semantically identifying concepts, e.g., A⊓A, A⊓A⊓A, . . . if concept name A exists, in the minimal models. As a main theoretical result, we show the decidability and complexity of the concept constructing algorithm. We formalize two applications to the concept constructing algorithm as a SROIQ-concept query system and SROIQ-concept learning for RDF graphs. The query system for RDF graphs returns the answers of expressive SROIQ queries including concept variables. The concept learning enables us to logically induce SROIQ-concepts from positive and negative examples in knowledge bases.
From the standpoint of Applied Ontology, we have challenged the problem of causality on the premise that it is real. As a result, it was shown in the former paper the following three items: (1) causation can be mapped to function, and hence it is possible to talk about causation in terms of function, (2) Cause can be divided into four subfunctions such as Achieve, Prevent, Allow and Disallow by introducing two dimensions of direct/indirect and positive/negative, and (3) using Achieve the other three subfunctions are defined and the essence of causality lies in Achieve. However, elucidation of Achieve itself was left incomplete. In this paper, we first restructure the results obtained in the former paper and summarize it so that the results will become clearer than the original paper. Then, the nature of Achieve is revealed to give a world first solution to the problem of what causation is.
A collaborative filtering technique with transparency using linear regression is proposed. In this study, the transparency is assumed as a role which users can know how the score is calculated. The proposed linear regression model presents regression coefficients to users. This study evaluates application of regularization and dimensionality reduction to estimate regression coefficient, and performs a score prediction experiment which analyze prediction accuracy, computing time, and obtained regression coefficients using five benchmark datasets for collaborative filtering. According to the experimental results, the proposed linear regression model with L2 regularization achieves the best prediction accuracy in the nine estimation methods considered for the regression coefficients. In addition, its prediction accuracy is similar level to Factorization Machines. The learning time of the proposed linear regression model is 24.9 to 1584 times faster than Factorization Machines. In the analysis of regression coefficients, variety in regression coefficient values are found in without regression and with L2 regression cases, and this suggests possibility of which the learned models are personalized.
Users’ attributes, such as home location, are necessary for various applications, such as news recommendations and event detections. However, most real user attributes (e.g., home location) are not open to the public. Therefore, their attributes are estimated by relationships between users. A social graph constructed from relationships between users can help estimate home locations, but it is difficult to collect many relationships, such as followers’ relationships. We focus on users whose home locations are difficult to estimate, so that we can select users whose locations can be accurately estimated before collecting relationships. In this paper, we use their profiles which can be collected before collecting relationships. Then, we analyze difficult users with their profiles. As a result, we found that users whose home locations incorrectly estimated had a longer duration since the date their account was created, longer name, and longer description. In addition, the results indicated that the users whose home locations were incorrectly estimated differed from those whose home locations could not be estimated.
In this study, we address the problem of detecting effective installation sites of signboards and identifying their influential zones for a given road network, under the setting that many residents can view them on their shortest paths to the destinations. To this end, based on a notion of group-betweenness centrality measure, we newly formalize this problem as a k-betweens problem and propose a community extraction method of road networks to identify the influential zones. In an existing method, the influential zone of each signboard is extracted by Voronoi tessellation against its installation site, which assumes that residents view the nearest signboard. In our proposed method, it is extracted by the proportion including the signboard on the shortest paths from the resident’s departure point to various destinations. From experimental evaluations using artificial and real road networks, we confirmed that our method can extract effective installation sites of signboards such as intersection nodes near the entrance of the express highway, and their influential zones as communities according to the positional relationships with arterial roadways. Furthermore, by computing the degree of antagonism among communities using the entropy of betweenness contribution rates, we can quantify the effectiveness of installing the same signboards for residents of areas where the influential zones of multiple signboards overlap.
Though a reinforcement learning framework has numerous achievements, it requires a careful shaping of a reward function that represents the objective of a task. There is a class of task in which an expert could demonstrate the optimal way of doing, but it is difficult to design a proper reward function. For these tasks, an inverse reinforcement learning approach seems useful because it makes it possible to estimates a reward function from expert’s demonstrations. Most existing inverse reinforcement learning algorithms assume that an expert gives demonstrations in a unique environment. However, an expert also could provide demonstrations of tasks within other environments of which have a specific objective function. For example, though it is hard to represent objective explicitly for a driving task, the driver could give demonstrations under multiple situations. In such cases, it is natural to utilize these demonstrations in multiple environments to estimate expert’s reward functions. We formulate this problem as Bayesian Inverse Reinforcement Learning problem and propose a Markov Chain Monte Carlo method for the problem. Experimental results show that the proposed method quantitatively overperforms existing methods.
We propose a method that assists legislation drafters in finding inappropriate use of Japanese legal terms and their corrections from Japanese statutory sentences. In particular, we focus on sets of similar legal terms whose usages are strictly defined in legislation drafting rules that have been established over the years. In this paper, we first define input and output of legal term correction task. We regard it as a special case of sentence completion test with multiple choices. Next, we describe a legal term correction method for Japanese statutory sentences. Our method predicts suitable legal terms using Random Forest classifiers. The classifiers in our method use adjacent words to a target legal term as input features, and are optimized in various parameters including the number of adjacent words to be used for each legal term set. We conduct an experiment using actual statutory sentences from 3,983 existing acts and cabinet orders that consist of approximately 47M words in total. As for legal term sets, we pick 27 sets from legislation drafting manuals. The experimental result shows that our method outperformed existing modern word prediction methods using neural language models and that each Random Forest classifier utilizes characteristics of its corresponding legal term set.
In most role-playing games (RPGs), non-player characters (NPCs) tend to have most initiative in conversation between the players and the NPCs. The NPCs are mainly talking to lead the players along the storylines of games, and sometimes ask the player to answer simple questions. Such a unilateral conversation is unnatural and unrealistic, and consequently, it is concerned that the player’s sense of immersion into the game will decrease. In order to improve the reality of the conversation, this study proposed a conversation method whereby an NPC adaptively switch the initiative according to the players’ preference of which of the player and the NPC has the initiative. To confirm the effect of this method, we conducted an experiment in which the participants talked with an NPC in a game scenario. This experiment compared the proposed method with the existing method that an NPC has the most initiative and the method that an NPC randomly switches the initiative. The result of the experiment suggested that the proposed method can decrease the unilateral impression compared with the existing method. Furthermore, the result also showed a probability that the proposed method motivates the players to talk to the NPC.
A chat-oriented dialogue system can become more likeable if it can remember information about users and use that information during a dialogue. We propose a chat-oriented dialogue system that can use user information acquired during a dialogue and discuss its effectiveness on the interaction over multiple days. In our subjective evaluation over five consecutive days, we compared three systems: A system that can remember and use user information over multiple days (proposed system), one that can only remember user information within a single dialogue session, and another that does not remember any user information. We found that users were significantly more satisfied with our proposed system than with the other two. This paper is the first to verify the effectiveness of remembering on the interaction over multiple days with a fully automated chat-oriented dialogue system.
In the past few years, there has been an increasing number of works on negotiation dialog. These studies mainly focus on situations where interlocutors work cooperatively to agree on a mutual objective that can fulfill each of their own requirements. However, in real-life negotiation, such situations do not happen all the time, and participants can tell lies to gain an advantage. In this research, we propose a negotiation dialog management system that detects when a user is lying and a dialog behavior for how the system should react when faced with a lie. We design our system for a living habits consultation scenario, where the system tries to persuade users to adopt healthy living habits. We show that we can use the partially observable Markov decision process (POMDP) to model this conversation and use reinforcement learning to train the system’s policy. Our experimental results demonstrate that the dialog manager considering deceptive states outperformed a dialog manager without this consideration in terms of the accuracy of action selection, and improved the true success rate of the negotiation in the healthcare consultation domain.
Argumentation-based dialogue systems, which can handle and exchange arguments through dialogue, have been widely researched. It is required that these systems have sufficient supporting information to argue their claims rationally; however, the systems do not often have enough information in realistic situations. One way to fill in the gap is acquiring such missing information from dialogue partners (information-seeking dialogue). Existing informationseeking dialogue systems were based on handcrafted dialogue strategies that exhaustively examine missing information. However, these strategies were not specialized in collecting information for constructing rational arguments. Moreover, the number of system’s inquiry candidates grows in accordance with the size of the argument set that the system deal with. In this paper, we formalize the process of information-seeking dialogue as Markov decision processes (MDPs) and apply deep reinforcement learning (DRL) for automatic optimization of a dialogue strategy. By utilizing DRL, our dialogue strategy can successfully minimize objective functions: the number of turns it takes for our system to collect necessary information in a dialogue. We also proposed another dialogue strategy optimization based on the knowledge existence. We modeled the knowledge of the dialogue partner by using Bernoulli mixture distribution. We conducted dialogue experiments using two datasets from different domains of argumentative dialogue. Experimental results show that the proposed dialogue strategy optimization outperformed existing heuristic dialogue strategies.
Understanding the various information from user utterances is important for chat-oriented dialogue systems. However, no study has yet clarified the types of information that should be understood by such systems. With this purpose in mind, we first collected information that humans perceive from each utterance (perceived information) in chat-oriented dialogue. We then categorized the types of perceived information. The types were evaluated on the basis of inter-annotator agreement, which showed substantial agreement and demonstrated the validity of our categorization. To the best of our knowledge, this study is the first attempt to clarify the types of information that a chat-oriented dialogue system should understand from varied user utterances.
Recently, end-to-end learning is frequently used to implement dialogue systems. However, existing systems still suffer from issues to handle complex dialogues. In this paper, we target on the conversation game “Mafia”, which requires players to make consistent and complex communications. We propose a middle language expression and a converter from natural language input. We implemented our dialogue system to play the Mafia game with humans and other automatic agents. Our evaluation on the play shows that our middle language increases conversion coverage.
The task of detecting dialogue breakdown, the aim of which is to detect whether a system utterance causes dialogue breakdown in a given dialogue context, has been actively researched in recent years. However, currently, it is not clear which evaluation metrics should be used to evaluate dialogue breakdown detectors, hindering progress in dialogue breakdown detection. In this paper, we propose finding appropriate metrics for evaluating the detectors in dialogue breakdown detection challenge 3. In our approach, we first enumerate possible evaluation metrics and then rank them on the basis of system ranking stability and discriminative power. By using the submitted runs (results of dialogue breakdown detection of participants) of dialogue breakdown detection challenge 3, we experimentally found that RSNOD(NB,PB,B) is an appropriate metric for dialogue breakdown detection in dialogue breakdown detection challenge 3 for English and Japanese, although NMD(NB,PB,B) and MSE(NB,PB,B) were found appropriate specifically for English and Japanese, respectively.
Interactive tutoring systems is expected to contribute self-learning in various type of topics, such as machine learning. In this study, we develop an interactive tutoring system that performs tutoring in a three-party dialogue form using multiple virtual agents. It would reduce the psychological burden of the learner compared with one-to-one tense dialogue. To adapt this system to various levels of users, question-answering functionality is essential. The proposed system generates an answer to the question using a knowledge graph automatically constructed from the textbook.