اميرحسين درخشان

The detection of contradictions in texts has significant applications in identifying fake news, detecting rumors on social networks, an‎d recognizing scientifically incorrect statements. Detecting contradictions between two sentences, as one of the subfields of Natural Language Inference (NLI), has received considerable attention, an‎d numerous models have been developed for this task. However, the use of these models for extracting contradictory sentence pairs from a given text corpus has not been extensively studied. In this study, a method named CETaD is proposed, which leverages pre-trained models on NLI datasets to extract contradictory sentence pairs from a text corpus. The approach first trains a Siamese neural network on NLI datasets. In the second step, sentences from the target text corpus are converted into vector embeddings using the pre-trained language model embedded within the Siamese network. Finally, by defining representative points in the embedding space an‎d comparing sentence embeddings against these representative points, CETaD efficiently extracts contradictory sentence pairs in an optimized manner. Given the potential to generate a large number of can‎didate pairs, several ranking criteria have also been developed. These criteria are based on the similarity between sentences an‎d aim to reduce the impact of biases present in the training data. Results analysis indicates that over 85% of the extracted sentence pairs are classified as contradictory by the trained Siamese neural network with high confidence. Furthermore, CETaD was eva‎luated on synthetic corpora constructed from factchecking datasets, enabling a more accurate assessment of its performance. To assess its effectiveness on real-world data, CETaD was applied to abstracts of medical articles from PubMed. The method achieved a precision of up to 60% an‎d successfully identified several contradictory statements within this domain, demonstrating its practical utility in detecting inconsistencies in scientific literature.

تناقض‌يابي , استنتاج زبان طبيعي , شبكه‌هاي عصبي , تعبيه‌سازي جملات , شبكه‌هاي عصبي سيامي , پردازش زبان طبيعي

Contradiction Detection , Natural Language Inference , Neural Networks , Sentence Embedding , Siamese Neural Networks , Natural Language Processing

Amirhossein Derakhshan

Dr. Hossein Rahmani