Multilingual sentiment analysis of summarized texts: a cross-language study of text shortening effects
Open Access
- 20 February 2026
- journal article
- Published by PeerJ in PeerJ Computer Science
- Vol. 12, e3406
Abstract
Sentiment analysis in multilingual contexts is significantly influenced by text summarization, with the effects varying across languages of different morphological complexities. This study examines the impact of extractive and abstractive summarization techniques on sentiment classification in eight typologically diverse languages: English (analytic), German (fusional), French, Spanish, Italian (moderately synthetic), Finnish and Hungarian (agglutinative), and Arabic (root-based inflectional). Multilingual transformer models (Multilingual Bidirectional Encoder Representations from Transformers (mBERT), Cross-lingual Language Model-Robustly optimized BERT approach (XLM-RoBERTa), T5, Bidirectional and Auto-Regressive Transformer (BART)) and language-specific adaptations (FinBERT, AraBERT) were employed to evaluate performance. It was found that extractive summarization better preserves sentiment accuracy, with modest declines of 2.6% in English (from 92.1% to 88.5%) and 5.2% in German compared to baseline results. In contrast, morphologically complex languages such as Finnish and Arabic exhibited more substantial drops of 8–12%. Abstractive summarization, while enhancing readability (ROUGE-1: 87.1% for English vs 89.3% for extractive methods), introduced significant sentiment distortion in agglutinative and inflectional languages due to semantic drift, including the omission of critical markers like Finnish case endings. A hybrid approach, combining extractive sentiment preservation through Term Frequency-Inverse Document Frequency (TF-IDF) keyphrase retention with abstractive fluency, was proposed and demonstrated superior performance. This method achieved a 4.2% higher F1-score compared to pure abstractive summarization in Arabic. The findings underscore the necessity of language-specific adaptations, particularly for applications such as social media monitoring and multilingual natural language processing (NLP) systems, where morphological complexity directly affects sentiment preservation accuracy.Keywords
All Related Versions
Funding Information
- Mexican Government, CONAHCYT, Mexico (A1-S-47854)
- Secretaria de Investigación y Posgrado of the Instituto Politecnico Nacional, Mexico (20241816, 20241819, 20240951)
This publication has 21 references indexed in Scilit:
- Exploring the Limits of Transfer Learning with a Unified Text-to-Text TransformerarXiv, 2023
- Attention Is All You NeedarXiv, 2023
- Generalization in Cooperative Multi-Agent SystemsarXiv, 2022
- Sentiment Lossless SummarizationKnowledge-Based Systems, 2021
- AraBERT: Transformer-based Model for Arabic Language UnderstandingarXiv, 2021
- Language Models are Few-Shot LearnersarXiv, 2020
- Unsupervised Cross-lingual Representation Learning at ScalearXiv, 2020
- Generalization of partitioned Runge--Kutta methods for adjoint systemsarXiv, 2020
- Multilingual is not enough: BERT for FinnisharXiv, 2019
- BERT: Pre-training of Deep Bidirectional Transformers for Language UnderstandingarXiv, 2019