Searching Sahih Hadiths Based on Queries using Neural Models and FastText

Abstract

Hadith is the second source of Islamic law after the Qur’an, and the availability of accurate and easily accessible information about hadith is crucial, as it directly affects a person’s belief (aqidah). This highlights the importance of having hadith collections as essential guidance in everyday life. Today, digital versions of hadiths are available in various applications, e-books, and websites. However, users often complain that these sources are incomplete and do not contain the entire collection of the Prophet's hadiths from al-Kutub as-Sittah. Additionally, the complex presentation of these digital resources makes it difficult to find relevant hadiths efficiently. This study aims to improve access to accurate and relevant hadith information, focusing specifically on al-Kutub as-Sittah, using Information Retrieval systems that search for hadiths based on keywords. IR is employed because it has proven effective in retrieving precise documents according to the search terms. A Neural Network is used to match user queries with the document collection, while FastText word embedding is implemented for text representation. FastText is particularly useful for detecting similar meanings across different words, which is essential when interpreting Indonesian-translated hadiths that require nuanced understanding. The dataset used in this study consists of 31,275 Indonesian-translated hadiths from al-Kutub as-Sittah. In this study, it was found that many hadith translations have ancient language so that query reformulation is needed to get the right hadith because users often enter commands with currently trending words. In this study, it was also found that word2vec has less performance than FastText in weighting words in hadith translations. The results indicate that the neural network performs well in retrieving relevant hadith content according to the user’s commands or keywords. With a training data proportion of 70% and a testing data proportion of 30%, the Recall value was 0.7721 and the Precision value was 0.75112.