Persistent Homology-Based Data Topology in Relation to the Accuracy of Translation Results Based on Transformer Models Euis Asriani a), Intan Muchtadi-Alamsyah b) c), Ayu Purwarianti c) d)
a) Doctoral Program of Mathematics, Faculty of Mathematics and Natural Sciences, Institut Teknologi Bandung, Bandung 40132, Indonesia
b) Algebra Research Group, Faculty of Mathematics and Natural Sciences, Institut Teknologi Bandung, Bandung 40132, Indonesia
c) University Center of Excellence Artificial Intelligence on Vision, Natural Language Processing and Big Data Analytics (U-CoE AI-VLB), Institut Teknologi Bandung, Bandung 40132, Indonesia
d) Informatics Research Group, School of Electrical Engineering and Informatics, Institut Teknologi Bandung, Bandung 40132, Indonesia
Abstract
Topological data analysis plays an important role in data analysis,
particularly in exploring data shapes. In the realm of transformer-based
translation models, data preparation is a crucial phase that can affect the
quality of translation results. This research aims to investigate the correlation
between the quality of translation results and the structure or shape of the
dataset based on topological data analysis. Experiments were conducted on a modified
transformer-based translation machine using a block circulant weight matrix and
the DCT-DST matrix-vector multiplication algorithm. The quality of the translation
results is based on BLEU scores on two pairs of datasets: Portuguese-English and
Indonesian-English. The structure of both language pair datasets was explored using persistent homology and represented by persistence diagrams and barcodes. The Wasserstein distance was used
to measure the similarity between two persistence diagrams. The experimental results
show that the dataset structure reflects the BLEU scores achieved in translating
both language pairs.
