類似した質問と表に基づく表検索及び大規模言語モデルを用いたOne-shot表質問応答
キーワード:
Table retrieval,
Table QA,
RoBERTa,
Language model,
Large Language Models,
One-shot Prompting
2025 年 2025 巻 FIN-034 号 p. 29-33
詳細
2025 年 2025 巻 FIN-034 号 p. 29-33
This paper proposes a table retrieval method based on similarity and a one-shot question-answering (QA) method using a large language model for the massive volume of tabular data found in securities reports. In the table retrieval task, we first utilize distributed representations obtained from a pre-trained model to retrieve question texts similar to the query. Then, we compute their similarity to associated tabular data, identifying the table with the highest similarity. In the table QA task, we achieve one-shot prompting by providing the retrieved similar question text and its corresponding tabular data as a sample to the large language model. This method aims to enable flexible QA in the financial domain, where various tables may be contained.
