سيدمحمدعرفان موسوي منزه

With the increasing use of artificial intelligence, many existing models still act as black boxes. The lack of transparency in the decision-making process of these models has created a serious challenge in their application in sensitive fields such as medicine, social sciences, and education. To deal with this problem, methods called explainable artificial intelligence have been developed. However, these methods are often complex and require specialized knowledge in the relevant field to analyze and use them effectively. One promising solution to overcome this challenge is the use of explainable artificial intelligence with natural language. Natural language, as a universal and understandable tool for humans, can provide a better understanding of the inner workings of artificial intelligence models. In this regard, natural language inference is proposed as one of the key areas in natural language processing. In this research, different architectures of explainable inference are reviewed and the strengths and weaknesses of each are compared. For the first time, two architectures based on transformers are implemented: “explanation then prediction” and “simultaneous explanation and prediction”. The results show that the “explanation then predict” architecture outperforms the other architectures by at least 3%. In addition, the present study shows that existing dataset have poor explanations. Using new models with reasoning, a new dataset containing more than 500,000 rows with high-quality explanations is generated and the quality of this data is eva‎luated. Also, for the first time, this dataset along with the explanations is machine translated into Persian and the translation quality is also examined. Next, models from different families are trained on three datasets and their results are compared. The performance of a set of large language models from pre-trained and reasoning models is also eva‎luated on these datasets. This dataset serves as a benchmark for measuring the quality of explanations produced by AI models. This research is an important step towards increasing the transparency and reliability of AI models in sensitive domains and could help expand the use of this technology in critical applications.

هوش مصنوعي توضيح پذير , پردازش زبان طبيعي , يادگيري عميق , مدل هاي زباني بزرگ

Explainable AI , Natural Language Processing , Deep Learning , Large Language Models

Seyyed Mohammad Erfan Moosavi Monazzah

Sayed Sauleh Eetemadi