TY - GEN
T1 - Recognition Textual Entailment on Bahasa Using Biplet Individual Comparison and BiLSTM
AU - Suwija Putra, I. Made
AU - Siahaan, Daniel
AU - Saikhu, Ahmad
N1 - Publisher Copyright:
© 2023 IEEE.
PY - 2023
Y1 - 2023
N2 - Recognizing Textual Entailment (RTE) is one of the important tasks in Natural Language Processing (NLP). Various approaches have been taken, starting from a simple statistical framework to the neural network (NN) that is currently the mainstay, including RTE in Bahasa Indonesia. Currently, RTE in Bahasa Indonesia has started using the neural network approach, but the value of the resulting accuracy is still less than 77%. This is because the new NN architecture has just accommodated the lexical elements and not yet the syntactical elements of sentences. Syntactical elements in sentences are important components in obtaining the local information contained therein. In RTE, local information is useful for determining how closely related text fragments are. This study proposes a new approach to NN-based Bahasa Indonesia RTE using the Biplet (head-dependency) individual comparison technique. Biplet is generated from the process of word pair dependency. The concept of word pair dependency is used to improve alignment and inference assessment that is optimized by adjusting the weight of the phrase using an attention mechanism. From experiments conducted using the SNLI dataset that has been translated into Bahasa Indonesia (SNLI Indo), it was obtained that the highest training accuracy value is 83.56% with the validation accuracy value is 64.61% for the number of pairs of sentences of 100k.
AB - Recognizing Textual Entailment (RTE) is one of the important tasks in Natural Language Processing (NLP). Various approaches have been taken, starting from a simple statistical framework to the neural network (NN) that is currently the mainstay, including RTE in Bahasa Indonesia. Currently, RTE in Bahasa Indonesia has started using the neural network approach, but the value of the resulting accuracy is still less than 77%. This is because the new NN architecture has just accommodated the lexical elements and not yet the syntactical elements of sentences. Syntactical elements in sentences are important components in obtaining the local information contained therein. In RTE, local information is useful for determining how closely related text fragments are. This study proposes a new approach to NN-based Bahasa Indonesia RTE using the Biplet (head-dependency) individual comparison technique. Biplet is generated from the process of word pair dependency. The concept of word pair dependency is used to improve alignment and inference assessment that is optimized by adjusting the weight of the phrase using an attention mechanism. From experiments conducted using the SNLI dataset that has been translated into Bahasa Indonesia (SNLI Indo), it was obtained that the highest training accuracy value is 83.56% with the validation accuracy value is 64.61% for the number of pairs of sentences of 100k.
KW - Biplet individual comparison
KW - attention mechanism
KW - recognizing textual entailment
KW - syntactic relation
KW - word pair dependency
UR - http://www.scopus.com/inward/record.url?scp=85175439759&partnerID=8YFLogxK
U2 - 10.1109/ICSECS58457.2023.10256326
DO - 10.1109/ICSECS58457.2023.10256326
M3 - Conference contribution
AN - SCOPUS:85175439759
T3 - 8th International Conference on Software Engineering and Computer Systems, ICSECS 2023
SP - 41
EP - 46
BT - 8th International Conference on Software Engineering and Computer Systems, ICSECS 2023
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 8th IEEE International Conference on Software Engineering and Computer Systems, ICSECS 2023
Y2 - 25 August 2023 through 27 August 2023
ER -