Persian Natural Language Inference: A Meta-learning Approach

Proceedings of the 29th International Conference on Computational Linguistics, pages 4306–4319 October 12–17, 2022. 4306 Persian Natural Language Inference: A Meta-learning approach Heydar Soudani∗ Sharif University of Technology heydars@ce.sharif.edu Mohammad Hassan Mojab∗ Sharif University of Technology mhmojab@ce.sharif.edu Hamid Beigy Sharif University of Technology beigy@sharif.edu Abstract Incorporating information from other lan- guages can improve the results of tasks in low- resource languages. A powerful method of building functional natural language processing systems for low-resource languages is to com- bine multilingual pre-trained representations with cross-lingual transfer learning. In general, however, shared representations are learned separately, either across tasks or across lan- guages. This paper proposes a meta-learning approach for inferring natural language in Per- sian. Alternately, meta-learning uses different task information (such as QA in Persian) or other language information (such as natural language inference in English). Also, we in- vestigate the role of task augmentation strat- egy in forming additional high-quality tasks. We evaluate the proposed method using four languages and an auxiliary task. Compared to the baseline approach, the proposed model consistently outperforms it, improving accu- racy by roughly six percent. We also exam- ine the effect of finding appropriate initial pa- rameters using zero-shot evaluation and CCA similarity. Our code is publicly available at https://github.com/HassanMojab/MetaNLI. 1 Introduction In natural language processing (NLP), the goal is to improve models for the processing and produc- tion of human languages. As part of NLP, several tasks are defined, each covering a different level of natural language understanding. Meanwhile, natural language inference (NLI) is considered an appropriate and rigorous measure of language com- prehension. This task requires recognizing the con- sequences of natural language sentences, which in- dicates how well it understands the language (Mac- Cartney, 2009). NLI aims to determine the inferential relation- ship between a premise p and a hypothesis h. The ∗Equal contribution problem involves a three-class classification in which every pair (p, h) falls into one of three cat- egories: entailment, contradiction, and neutral. If the hypothesis can be inferred from the premise, pair (p, h) will be assigned to the entailment class. For a hypothesis that contradicts the premise, pair (p, h) will be assigned to the contradiction and neu- tral otherwise (Amirkhani et al., 2020). The Persian language lacks sufficient linguistic resources when it comes to natural language un- derstanding. The lack of data can be addressed by collecting annotated data, but this process is both time-consuming and expensive (Nooralahzadeh et al., 2020). FarsTail (Amirkhani et al., 2020) is currently available for Persian, which is cre- ated using the same method as SciTail (Khot et al., 2018). It contains 10,367 samples. Also, ParsiNLU (Khashabi et al., 2021) is created for high-level tasks in Persian, and for NLI, it con- sists of 2700 samples. As it turns out, this amount of data is too small compared with resource- rich languages (such as English, which has only 550,000 samples in the SNLI (Bowman et al., 2015) dataset). Researchers have tried to solve the data scarcity problem by using cross-language …

Content truncated to 500 words