Detecting Inconsistencies in Narrative Elements of Cross Lingual Nakba Texts

AbstractThis paper suggests a methodology for contradiction detection in cross lingual texts about the Nakba. We propose a pipeline that includes text translation using Google’s Gemini for context-aware translations, followed by a fact extraction task using either Gemini or the TextRank algorithm. We then apply Natural Language Inference (NLI) by using models trained for this task, such as XLM-RoBERTa and BART to detect contradictions from different texts about the Nakba. We also describe how the performance of such NLI models is affected by the complexity of some sentences as well as the unique syntactic and semantic characteristics of the Arabic language. Additionally, we introduce a method using cosine similarity of vector embeddings of facts for identifying missing or underrepresented topics among historical narrative texts. The approach we propose in this paper provides insights into biases, contradictions, and gaps in narratives surrounding the Nakba, offering a deeper understanding of historical perspectives.