Research on the Generation of Tibetan Word-levelAdversarial Samples Based on Morphosyntactic Linkage

Abstract Tibetan word substitution is a common technical strategy in text adversarial defense and data augmentation. This paper addresses the issue of neglecting Tibetan grammatical norms for function words during the core keyword substitution process in the field of Tibetan word-level text adversarial attacks and defense, proposing a novel text adversarial sample generation method TWLASBMSI that conforms to Tibetan grammatical rules through morphosyntactic linkage. The method consists of four core steps: First, using a Tibetan sememe knowledge base, the content in the original Tibetan text is randomly matched and replaced once with words of higher semantic similarity; second, employing a function word attachment detection and correction algorithm based on Tibetan grammatical categories to ensure automatic and consistent adaptation and updating of function words such as genitive, ergative, locative, and personal noun suffixes in the replaced text, thereby guaranteeing the generation effectiveness and text quality of Tibetan adversarial samples. Finally, the generated Tibetan adversarial samples are input into the target model to carry out attacks. Experimental results show that the TWLASBMSL attack method leads to an average accuracy drop (ADV) of 2.69%, an average attack success rate (ASR) of 28.83%, and an average Levenshtein distance (LD) of 6.80% on Tibetan sentiment analysis and text classification tasks using TiBERT, CINO-base-v2, and CINO-large target models. In the task of automatic adaptation and error correction for Tibetan function words, the function word error detection rate, overall accuracy, and complete error sentence correction success rate all reached 100%. Overall, the experimental results indicate that the proposed morphosyntactically linked text adversarial sample generation method TWLASBMSL not only effectively misleads models into making incorrect classification decisions but also ensures the grammatical logical rigor of Tibetan adversarial samples.