A text classification method combining in-domain pre-training and prompt learning for the steel e-commerce industry

• Date: 09 December 2024
• Pages: 96-119
• DOI: https://doi.org/10.1108/IJWIS-09-2024-0277
• Published date: 09 December 2024
• Subject Matter: Information & knowledge management,Information & communications technology,Information systems,Library & information science,Information behaviour & retrieval,Metadata,Internet
• Author: Qiaojuan Peng,Xiong Luo,Yuqi Yuan,Fengbo Gu,Hailun Shen,Ziyang Huang

A text classif‌ication method

combining in-domain pre-training

and prompt learning for the steel

e-commerce industry

Qiaojuan Peng,Xiong Luo and Yuqi Yuan

School of Computer and Communication Engineering, University of Science and

Technology Beijing, Beijing, China; Shunde Innovation School, University of

Science and Technology Beijing, Foshan, China and Beijing Key Laboratory of

Knowledge Engineering for Materials Science, Beijing, China, and

Fengbo Gu,Hailun Shen and Ziyang Huang

Ouyeel Co., Ltd, Shanghai, China

Abstract

Purpose –With the development of Webinformation systems, steel e-commerce platforms have accumulated a

large number of quality objection texts. These texts ref‌lect consumer dissatisfaction with the dimensions,

appearance and performance of steel products, providing valuable insights for product improvement and

consumer decision-making. Currently,mainstream solutions rely on pre-trained models, but their performance on

domain-specif‌ic data sets and few-shot data sets is not satisfactory.This paper aims to address these challenges by

proposing more effective methodsfor improving model performance on these specialized data sets.

Design/methodology/approach –This paper presents a method on the basis of in-domain pre-training,

bidirectional encoder representation from Transformers(BERT) and prompt learning. Specif‌ically,a domain-

specif‌ic unsupervised data set is introduced into the BERT model for in-domain pre-training, enabling the

model to better understandspecif‌ic language patterns in the steel e-commerce industry,enhancing the model’s

generalization capability; the incorporation of prompt learning into the BERT model enhances attention to

sentencecontext, improving classif‌ication performanceon few-shot data sets.

Findings –Through experimentalevaluation, this method demonstrates superior performanceon the quality

objection data set, achievinga Macro-F1 score of 93.32%.Additionally, ablation experiments further validate

the signif‌icantadvantages of in-domain pre-training and prompt learningin enhancing model performance.

Originality/value –This study clearly demonstrates the value of the new method in improving the

classif‌ication of qualityobjection texts for steel products. The f‌indings of this study offer practical insightsfor

product improvementin the steel industry and provide new directions for future research onfew-shot learning

and domain-specif‌icmodels, with potential applications in other f‌ields.

Keywords Data analysis, Text classif‌ication, In-domain pre-training, Prompt learning

Paper type Research paper

Funding: This work was supported in part by the Beijing Natural Science Foundation under Grant

L211020 and in part by the National Natural Science Foundation of China under Grants U1836106,

62271045 and 62202044.

Statements and declarations: The authors declare that there are no conf‌licts of interest regarding the

publication of this paper. No human or animal subjects are involved in the study described in this

paper. All authors have agreed to the submission of this paper.

Data availability: The data underlying this article were provided by Ouyeel Co., Ltd under license.

Data will be shared on request to the corresponding author with permission of Ouyeel Co., Ltd.

IJWIS

21,1

96

Received27 September 2024

Revised3 November2024

Accepted 12 November 2024

InternationalJournal of Web

InformationSystems

Vol.21 No. 1, 2025

pp. 96-119

© Emerald Publishing Limited

1744-0084

DOI 10.1108/IJWIS-09-2024-0277

The current issue and full text archive of this journal is available on Emerald Insight at:

https://www.emerald.com/insight/1744-0084.htm

1. Introduction

With the widespread adoption of Web information systems, including e-commerce platforms,

people are increasingly inclined to shop and trade online. In the steel e-commerce industry, the

rapid development of these Web information systems has shifted steel transactions from

traditional off‌line markets to online platforms. Consumers can easily purchase various

specif‌ications and varieties of steel products through e-commerce platforms.As the scale of steel

transaction continues to expand, consumers’attention to the quality and performance of steel is

also increasing, and expressing dissatisfaction with the quality of steel has become more and

more frequent. These quality objection texts usually express consumers’dissatisfaction with the

dimensions, specif‌ications, appearance, performance and other aspects of steel products. By

accurately classifying quality objection texts, steel manufacturers can gain valuable in sights into

specif‌ic issues or defects that impact their products. For example, they can identify common

problems such as surface irregularities, dimensional inconsistencies or material defects. With this

information, manufacturers can make informed decisions to improve their manufacturing

processes, optimize product design and address quality issues promptly. Additionally, precise

classif‌ication of quality objection texts directly inf‌luences consumer decision-making. Steel

products are widely used in various industries, including construction, automotive and

infrastructure. When consumers consider purchasing steel products, they rely on information

about the quality and performance of different steel options. Accurat e classif‌ication of objection

texts provides consumers with reliable data and insights, enabling them to make informed

choices based on their specif‌ic requirements. In this context, how to acc urately identify and

classify quality objection texts has become a pressing issue that requires immediate attention.

As a data analysis technique, text classif‌ication aims to automatically analyze and process text

to classify it into pre-def‌ined categories (Li et al.,2024;Qian et al.,2024). Text classif‌ication has

wide applications in many f‌ields such as information retrieval (Kumar and Sharaff, 2024),

sentiment analysis (Khan et al.,2024) and news classif‌ication (Matemilola and Aliyu, 2024). It can

help extract useful information from massive amounts of text, quickly locate the required text and

speed up information processing. Therefore, we can use text classif‌ication technology to

automatically classify quality objection texts in the steel e-commerce industry.

Currently, existing text classif‌ication methods can be broadly categorized into rule-based

methods, machine learning-based methods and deep learning-based methods. Rule-based

methods rely on manually designed rules to classify texts using prior knowledge and experience.

This method has the advantage of being fast and easy to adjust, but it requires domain experts to

design rules, making it diff‌icult to adapt to changes in dataand new situations. Machine learning-

based methods use statistical learning to train a classif‌ier from the training data and then use it to

classify new data. Common machine learning algorithms include support vector machine (SVM)

(Hearst et al.,1998), random forest (Breiman, 2001), decision tree (Harrag et al., 2009) and others.

This method is advantageous in terms of simple algorithm implementation, ease of understanding

and adjustment and good performance on small-scale data sets. Deep learning-based methods use

the deep neural network to model and learn from text data, allowing for more complex and

sophisticated representations of text. Typical deep learning models include recurrent neural

network (RNN) (Liu et al., 2016), convolutional neural network (CNN) (Song et al.,2024),

Transformer (Vas wa n i et al., 2017) and bidirectional encoder representations from Transformers

(BERT) (Devlin et al., 2019). The main benef‌it of this method is that it can use large-scale data

sets for training and automatically learn effective feature representations, resulting in good

performance on complex text classif‌ication tasks. As a consequence, deep learning-basedmethods

have gained prominence as the primary approach for text classif‌ication tasks, particularly withthe

emergence of pre-trained models like BERT, which has become a popular solution in the natural

language processing (NLP) f‌ield.

International

Journal of Web

Information

Systems

97