Thesis or Dissertation A Study on Dialogue Agent Adapting to Various Situations

Jordan, Arnaud

In the present world, machines and computers are used everywhere all days long by many different users. Therefore, many different kinds of interface have been set up in order to communicate with machines, such as for example buttons, mouse, touchscreen or keyboard. On the other hand, humans used to use natural languages such as Japanese or French in order to communicate each other. That is why it is commonly accepted that for humans it is more natural to speak, use gestures and so forth to communicate. In this context, interfaces using natural languages gained considerable attention in recent years. Especially with the democratization of smartphones, voice interfaces and dialogue agents attract evenmore attention. However, even if speech recognition precision increased continuously, computers are still not able to fully understand all the aspects of natural languages, and as a result users feel they are hard to talk and to use. Moreover, speech recognition precision already reached a ceiling and cannot be largely improved, as a result, the malaise felt by the user is considered due to the deficiency of the responses generated by the system more than to voice processing. That is why in order to solve this problem I consider new interactive processing have to be developed to achieve natural conversation between humans and machines. In order to achieve a smooth dialogue with humans, several faculties such as casualchatting, question answering, emotional reaction, joke or gesture are required. As a first step, I considered developing a dialogue agent focusing on casual dialogue to handle daily life conversation. Concretely, the proposed agent does not achieve particular task, butfocuses on providing natural responses corresponding to the user input. The present thesis proposes several innovative methods based on the human’s behavior and presents the results of the developed dialogue agents.For many years, diverse dialogue agents have been developed. However, they generally use a limited static knowledge in order to give priority to experiment results. Contrary, the present thesis does not only focus on the systems’ results, but also on the future possible developments of the system and attempts to propose flexible and adaptive systems. Concretely, the research converges on dialogue agents constantly adapting to the environment where they are used. Therefore, several methods and dialogue agents are proposed. In addition, experiments have been carried out to prove their reliability and to discuss their advantages and drawbacks. The present thesis is composed of four chapters. In chapter one, the introduction presents the background of the research and the state of the art of dialogue agents. In chapter two, as a first step, a dialogue agent handling uestion answering for casual dialogue in Japanese and the related experimental performance evaluation are presented. Namely, the proposed method uses rule acquisition and interrogative pronoun identifica-tion to generate responses. This method uses a morphological analyzer to identify interrogative pronouns present in the input sentence, and then selects the most adequate word among all the knowledge it acquired to substitute it, and then generate the output. In addition, several methods of output generation using knowledge acquired from different users are presented, tested and compared. Moreover, in order to facilitate the use of the developed systems several interfaces have been developed such as a web page and smartphone applications providing vocal dialogue. However, the proposed systems depend on the language. Furthermore, implementing all the possible inputs or input structures using this kind of language-limited methods is complicated. Especially, the system cannot be enough flexible to handle unknown wordsor linguistic phenomena since predicting all the inputs is impossible. That is why the proposed first methods are considered insufficient to achieve the target precision. In chapter three, the feasibility of language-independent dialogue agent, i.e. a systemwhich does not limit the language of the input, is studied and a method is proposed. Concretely, the proposed method accepts any language as input. Contrary to methods which require huge external resources and tools, the proposed method attempts to limit them as much as possible in order to focus on improving the flexibility and the knowledge acquisition processing and finally obtain an improvement of the precision of the system. Therefore, dictionaries and morphological analyzers are not used since they are language-dependent. Basically, the newly proposed dialogue agent represents all the knowledge it acquires during dialogues into a graph, and uses the latter to generate outputs. As a consequence, previously proposed graph algorithm can be adapted to the system. For example, results prove that the proposed method can infer new knowledge using graph clustering. In addition, parallel simultaneous graph traversal has been efficiently imple-mented in the system due to the hierarchical construction of the graph. Concretely, the proposed dialogue agent represents all the acquired knowledge into a graph and accepts any kind of input at any time, i.e. even during the output generation, and then generates one or several outputs in function of its internal state. Experiments in five languages; Chi-nese, English, French, Korean and Japanese has been carried out and the results exceed a more relative multilingual dialogue agent. Moreover, since the proposed method handles any kind of input, a system upgrade enabling emotional input was proposed and a result improvement compared to the previous agent was observed. Therefore, since language-independent methods can handle any tag representing emotion or environment similarly to textual input, they are considered well-adapted to the development of multimodal dialogueagent. Chapter four concludes the present thesis. Especially, details about the research progress, the evolution of the objectives and the final results are presented. In addition, advantages and drawbacks of each method are discussed and possible improvementsare proposed. Moreover, future evolutions and challenges of dialogue agents, as well as new objectives are broached.
v, 101p
Hokkaido University(北海道大学). 博士(情報科学)

Number of accesses :  

Other information