Structure dependent weighted loss function for triple extraction using BART model

Manu Shrivastava

doi:10.1007/s10044-025-01604-9

Return

Structure dependent weighted loss function for triple extraction using BART model

2026-01-29

PRE

Manu Shrivastava *

Kosei Shibata

Hiroaki Wagatsuma

DOI：10.1007/s10044-025-01604-9

Original

Original request for help

Save

Abstract

En 中文

A Knowledge-Based System (KBS) comprises two parts, a knowledge base where all domain-specific knowledge is stored, typically organized according to an ontology, and an inference engine that processes and interprets knowledge. In the manufacturing domain, KBS can be queried to understand the functioning of machines and their failure causes thereby aiding in maintenance and reducing the dependency on expert technicians. Automating the creation of KBS is a critical challenge, and neural network models (NNMs) have emerged as a solution to generate Resource Description Framework (RDF) triples from natural language input. These models take a sentence as an input and generate RDF triples, the complexity and structure of input sentences can significantly impact the performance of these models. In this work, we analyze the effect of sentence structure on the accuracy of the sequential model’s performance. We conducted statistical analysis on the embeddings learned by NNMs to examine how different sentence structures influence model learning behavior and how the distribution of learned embeddings depends on sentence structure. Our findings suggest that sentence structure has an effect similar to class imbalance in traditional machine learning tasks, leading to performance disparities across different sentence types. To address this issue, we propose a structure-dependent weighted loss function (SDWLF), which categorizes sentences based on their structure and applies varying weights to loss for optimizing model performance for each type. We evaluate the effectiveness of the SDWLF on sequential models like BART, demonstrating measurable improvements in the translation of sentences with varying structures.

Keywords:

Information retrieval

Sentence structure

Kolmogorov–Smirnov and Cramer Von Mises test

RDF Extraction

Class imbalance

Encoder decoder

Weighted loss function