Siamese Networks for Inference in Malayalam Language Texts

Proceedings of Recent Advances in Natural Language Processing, pages 1167–1173 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_131 1167 Siamese Networks for Inference in Malayalam Language Texts Sara Renjit Department of Computer Science CUSAT, Kochi, India sararenjit.g@gmail.com Sumam Mary Idicula Department of Computer Science CUSAT, Kochi, India sumam@cusat.ac.in Abstract Natural language inference is a method of find- ing inferences in language texts. Understand- ing the meaning of a sentence and its infer- ence is essential in many language process- ing applications. In this context, we consider the inference problem for a Dravidian lan- guage, Malayalam. Siamese networks train the text hypothesis pairs with word embeddings and language agnostic embeddings, and the re- sults are evaluated against classification met- rics for binary classification into entailment and contradiction classes. XLM-R embed- dings based Siamese architecture using gated recurrent units and bidirectional long short term memory networks provide promising re- sults for this classification problem. 1 Introduction Textual entailment is a uni-directional relationship between two text pairs. It is the method of identi- fying the meaning from two sentences. Sentence pairs are labeled as entailment pairs if a sentence is inferred from the context of the other sentence in the pair. It is defined in different ways as classical def- inition, applied definition, and mathematical def- inition (Ghuge and Bhattacharya, 2014). In the classical definition, a text entails a hypothesis if the hypothesis is valid in all circumstances where the text is true. In applied definition, text entails hypothesis if the hypothesis is mostly true when a human reads it. The mathematical definition (Glick- man et al., 2005) is a text that entails a hypothesis if the probability of the hypothesis is true given the text is greater than the likelihood of the hypothesis being true. The sentence from which we derive the informa- tion is called text, and the sentence which identifies its information as derived from text is called a hy- pothesis. Sentence pairs are named as entailed if the hypothesis has its meaning derived from the text. Contradictory pairs indicate that the informa- tion from the hypothesis is contradictory concern- ing the information conveyed in the text. Nowadays, textual entailment has become part of the primary natural language processing tasks. It is also called natural language inference. The terms text and premise are used interchangeably. The performance of the latest transformer-based approaches is evaluated in NLP tasks like entail- ment recognition, semantic textual similarity, and paraphrase detection. Hence, textual entailment recognition has also become an evaluation criterion for many NLP tasks. It is also a necessary sub-task in applications like multi-document summarization, information retrieval, information extraction, and question answering systems. Many text entailment or natural language infer- ence related works in English use different sized datasets but not in the Malayalam language. Malay- alam is a Dravidian language used in the southern part of India. It is a language that has various dialects and has many inflections. Malayalam lan- guage computing is developing, with few resources, namely stemmer, POS tagger, sandhi splitter, and few …

Content truncated to 500 words