Machine Learning for Indigenous and Local Knowledge of Sustainable Management of Nature

Abstract

Biodiversity is an important foundation for human survival. Traditional knowledge, known as Indigenous and Local Knowledge (ILK), is critical to conserving biodiversity. However, the inheritance of ILK is getting difficult all in Japan. Against this trend, machine learning is expected to have the potential to facilitate the challenge of inheriting ILK. The purpose of this study is to develop a dialogue system to support the succession of Japanese ILK for managing nature. As study data, 529 articles from "Kikigaki Koshien" were collected, which is a knowledge base related to nature's contributions and inheritance and revival of tradition. The two dialogue systems were built: a retrieval system and a generative system. For retrieval system, the articles were preprocessed and the natural language data frame with 52360 triplets was built. And ILK was learned from the triplets using Sentence-BERT, which was pre-trained on Japanese Wikipedia. Next, By the fine-tuned BERT, articles are retrieved and answers to questions are synthesized. And for generative system, ILK was learned from the bodies of the articles using GPT-2, which was pre-trained on Japanese web crawl data. Next, By the fine-tuned GPT-2, texts on traditional knowledge following the given prompt were generated. Finally, an experiment with 19 subjects to evaluate the naturalness and usefulness of the 10 synthesized conversations by retrieval system was conducted and future development perspectives were discussed.