| 研究実績の概要 |
During the final year of the grant period I concentrated on data creation. As there is a growing number of datasets for various languages and less for Japanese, I have used crowdsourcing for making 1) danger-safety simple sentences dataset and 2) a story dataset based on previously created simple sentences.
The first one contains over 21,000 sentences written by crowdworkers based on "dangerous" nad "safe" verbs. The entries differ in gender or age of agents and patients, which can be used for bias detection. Other group of crowdworkers evaluated how dangerous a given act is and the human evaluation can be used for testing AI’s capability for recognizing potentially dangerous situations.
The story corpus is probably the first such dataset for Japanese and it contains 8,000 five sentence stories, which can be used by researches for various tasks. I have used almost the whole sum planned to be spent for the year to hire crowdworkers with various backgrounds, genders and ages. The acquired data is based on the first dataset and can be used beyond the purpose for the grant topic (being used for discovering thought process to be used in inverse reinforcement learning): natural language processing specialists can utilize it for story understanding and generation, causal and temporal knowledge processing, etc.
I have managed to perform experiments on the first dataset (Japanese BERT language model appeared to have lower error rate than LSTM and BiLSTM) but evaluating the story generation output is problematic.
|
