High-Speed ​​Personalization of Conversational AI Speech Plans with Quantum Computing Technology / Presentation of Research Findings at the 28th Annual Meeting of the Association for Natural Language Processing held in March | Hebei Shinpo Online News / ONLINE NEWS

~ AI that conveys the topic of your interest in a conversational format in an easy to understand way ~

 Naigai Cutout Communication Co.,Ltd. https://www.naigaipc.co.jp/ (Location: Shinjuku-ku, Tokyo, President: Yoshiaki Kondo) is Waseda University’s Green Computing System Research Organization (GCS Research). Perceptual Information Systems Research Institute (Location: Shinjuku-ku, Tokyo, Director: Tetsunori Kobayashi, Professor, Faculty of Science and Engineering) Hiroaki Takatsu, Research Group (Head of Research: Yoichi Matsuyama), Fujitsu Limited (Location : Tokyo) Conversation AI speech plan in collaboration with Minato Ward, president: Takahito Tokita)(※1)We have developed a technology to optimize at high speed.

* 1 speech plane = prepared sentences for the AI ​​to speak

The Dialogue Systems Research Group of the research organization GCS conducts research and development of a voice dialogue system that summarizes a large amount of information represented by news articles according to individual interests and conveys effectively in a conversational format. .. We cooperated to increase the robustness of machine learning models by combining knowledge of data construction and media expression for AI learning.

 With the increase in the number of users and the scale of the source information, this voice dialogue system requires a huge computation time to find the optimal solution. Therefore, it is important to speed up the optimization computation to generate a personalized utterance plan for many users in a convenient time.

In this joint research, with the cooperation of Fujitsu Limited, Fujitsu Digital Annealer(※2)High-speed optimization achieved using.

Going forward, we will encourage research and development as a practical effort to design new services such as newsreading services and voice chat advertisements and experience value.

The results of this research were presented at the 28th Annual Meeting of the Association for Natural Language Processing held in March.

[28e conférence annuelle de l’Association pour le traitement du langage naturel tenue en mars 2022]
https://www.anlp.jp/nlp2022/
https://www.anlp.jp/nlp2022/program.html

  • Title: Quick solution of custom extractive summaries using digital annealing
  • Authors: Hiroaki Takatsu, Takahiro Kashiwakawa, Koichi Kimura, Ryota Ando, ​​Yoichi Matsuyama
  • Proceedings: Proceedings of the 28th Annual Meeting of the Association for Natural Language Processing, pp. 758-763, 2022.

From left to right, Researcher Ryota Ando (Doshisha University), SCHEMA (Waseda University), Researcher Hiroaki Takatsu (Doshisha University)

Adapt to an ever-increasing amount of information and changing personal needs
The amount of information generated in the ever-changing digital economy continues to grow, and this information is the source of value creation. For companies, the importance of collecting, analyzing and using information in business is increasing.
In addition, the spread of personal computers and smart phones has made it easier for consumers to access information on the web. As a result, contact time with the media is increasing year by year.
(※3).. On the other hand, there is a growing awareness that “the amount of information in the world is too much”, and there is a growing need to acquire information more efficiently in a limited time.
Starting from this social milieu, this joint research aims to create ideas for enriching people’s media experiences by posing problems on the topic of ever-increasing information.

*3 Hakuhodo DY Media Partners Co., Ltd. (ed.), Media Guide 2020 for those involved in the advertising industry, Sendenkaigi Co., Ltd. (editing).

[Annonce des résultats pour 2021]
The results of the joint research were announced at the following international workshops, national study groups and EXPO on artificial intelligence.

① IWSDS (announced November 2021)
IWSDS (International Workshop on Spoken Dialog System Technology) is an international workshop on speech dialog system technology, and the 12th time in 2021 was held in Singapore.
https://www.colips.org/conferences/iwsds2021/wp/accepted-papers/

  • Title: Personalized Extractive Summarization with Discourse Structure Constraints for Efficient and Coherent Dissemination of Dialogue-Based Information
  • Authors: Hiroaki Takatsu, Ryota Ando, ​​Hiroshi Honda, Yoichi Matsuyama and Tetsunori Kobayashi
  • Proceedings: Proceedings of the 12th International Workshop on Speech Dialogue Systems Technology, 2021.

  • Title: Personalized extractive summary using an Ising machine towards the real-time generation of effective and coherent dialogue scenarios.
  • Authors: Hiroaki Takatsu, Takahiro Kashikawa, Koichi Kimura, Ryota Ando and Yoichi Matsuyama
  • Proceedings: Proceedings of the 3rd Workshop on Natural Language Processing for Conversational AI, pp. 16-29, 2021.

  • Title: Optimization of the museum guide statement plan using the Ising machine
  • Authors: Hiroaki Takatsu, Ryota Ando, ​​Tetsuhei Nakano, Takahiro Kashiwakawa, Koichi Kimura, Yoichi Matsuyama
  • Proceedings: 93rd Study Group on Language and Speech Understanding and Dialogue Processing (12th Dialogue System Symposium), pp. 26-31, 2021.

This voice dialogue system facilitates dialogue based on an utterance script composed of a main plane of explanation of main points and a sub-plane of complementary explanations. The system transmits the content of the main plan while the user listens passively, but switches to the subplan and provides additional explanations, if necessary, in response to the user’s request for information. In this study, we investigated the master plan generation method to achieve efficient and consistent information transmission. In the IWSDS paper, the master plan generation problem is optimally combined by extracting from the document set the set of sentences that maximizes the computed degree of interest for each user under the constraints of the discourse structure. of the document and the total speaking time. Formulated as a conversion problem. As an index of evaluation, we defined the efficiency of the transmission of information, which is the harmonic mean of the ratio which can cover the information of interest and of the ratio which can exclude the information which was not of interest. interest. We evaluated the information transmission efficiency of the proposed method using a dataset with a discourse structure and degree of interest in news articles, and confirmed that a highly consistent and effective can be generated.
On the other hand, the integer linear programming problem to generate the master plan formulated here requires enormous computation time to search for the optimal solution as the scale of the problem increases. In order to generate a custom master plan for a large number of users in a convenient time, it is important to speed up the optimization calculation. Therefore, in the NLP4ConvAI paper, the entire linear programming problem formulated in the IWSDS paper is converted to Quadratic Unconstrained Binary Optimization (QUBO) format and solved using digital annealing, so that a near-optimal solution without constraint violation is convenient. confirmed that it can be obtained within a reasonable time. Moreover, we have implemented this technology in the museum guide system and demonstrated the effectiveness of explaining the exhibits on the virtual museum based on the personalized master plan. The results were presented at the Dialogue System Symposium.

On-site server “Digital Annealer”

FUJITSU Digital Annealing Unit (DAU)

  • Title: Quick solution of custom extractive summaries using digital annealing
  • Authors: Hiroaki Takatsu, Takahiro Kashiwakawa, Koichi Kimura, Ryota Ando, ​​Yoichi Matsuyama
  • Proceedings: Proceedings of the 28th Annual Meeting of the Association for Natural Language Processing, pp. 758-763, 2022.
⑤ NexTech Week 2021 Autumn 2nd AI・Artificial Intelligence EXPO【Autumn】Exhibition and Cooperation
We exhibited at “Academic Forum” special area of ​​”2nd AI/Artificial Intelligence EXPO[Autumn]which was held at Makuhari Messe in October 2021.

The 2nd AI・Artificial Intelligence Expo【Autumn】


Mr. Yoichi Matsuyama, Research Representative, Dialogue Systems Research Group, GCS Research Organization

Contact for any request on this subject
Naigai Cutout Communication Co.,Ltd.
Inquiries https://www.naigaipc.co.jp/ask/

responsible
Ando ando@naigaipc.co.jp
Matsui t.matsui@naigaipc.co.jp

Related links
Naigai Cutout Communication Co.,Ltd.
https://www.naigaipc.co.jp/
Waseda University GCS Research Institute
https://www.waseda.jp/inst/gcs/
Fujitsu Limited
https://www.fujitsu.com/jp/

Leave a Comment