Fact Verification for Automatic Summarization of Political Minutes by Generating Pseudo-Fake Data Using Round-Trip Translation

Abstract

With the widespread adoption of social media, the proliferation of fake news and false rumors about politics has become a social issue. To detect and verify them, it is essential to extract and summarize information from political minutes automatically. In recent years, automatic summarization by AI and large language models (LLMs) has been actively researched. However, incomplete automatic summarization poses a risk of generating new fake news. To address these issues, the Answer Verification subtask was conducted in the QA Lab-PoliInfo-4 at the NTCIR-17 conference. The purpose of this subtask is to perform fact-checking of automatically summarized answers in the minutes. We used Round-Trip translation to generate pseudo-fake data to augment the training data. Retraining the baseline model with these data improves the fact-checking performance. The result shows that pseudo-fake data generation using Round-Trip translation is also effective for the automatic summarization data.