A Study of the State of the Art Approaches and Datasets for Multilingual Natural Language Inference

Natural language inference is critical in Natural Language Processing where semantics is involved. Also known as textual entailment recognition, it defines a directional relationship between pair of sentences, namely the text and the hypothesis. Identifying entailment, contradiction, and neutrality in text pairs is necessary for various language processing applications to reduce text redundancy. An overview of the critical works in this aspect for all languages, including the Indian language perspective, is detailed here. There is a high volume of textual entailment and related attempts in English and foreign languages. In contrast, there are only a few attempts for low resource languages. This article presents the progress in textual entailment for various languages and the development of different datasets for textual entailment. Over these years, the datasets developed differ in size and kind. Most of the datasets are raw or generated, and a few of the latest are translated datasets. The article also points to observation notes on the progress that has happened throughout these years.