Recognizing Textual Entailment in Indonesian Using Individual Biplet Head-Dependent and Multi-Head Attention Mechanism

Recognizing Textual Entailment (RTE) has become essential to determine inferential relationships between sentences in text-understanding systems. Traditionally, RTE models have addressed textual inferences at both syntactic and semantic levels. However, the development of RTE models for the Indonesian language has mainly been limited to lexical sentence-level analysis, overlooking syntactic dependencies among words in the Premise (P) and Hypothesis (H) sentences. This limitation often leads to the capture of irrelevant information within Premise sentences. In this study, we proposed Indo-Biplet Entailment Model (Indo-BiEnt), a deep learning architecture for RTE capable of learning sequential information from sentence pairs transformed into individual Head-Dependent word pairs (Biplet (h-d)). These Biplets are derived from a word-pair-dependency process to capture syntactic relationships between words. Each Biplet pair undergoes direct comparison within the pair segments. Additionally, we employed a BiLSTM network augmented with multi-head attention to emphasize critical words more effectively. Through experiments carried out using the SNLI Indo dataset, our approach achieved a test accuracy of 79%, demonstrating its effectiveness in exploiting syntactic dependencies to enhance the performance of RTE in Indonesian.