Revisiting Transformer-based Models for Long Document Classification
Publikation: Bidrag til bog/antologi/rapport › Konferencebidrag i proceedings › Forskning › fagfællebedømt
Standard
Revisiting Transformer-based Models for Long Document Classification. / Dai, Xiang ; Chalkidis, Ilias; Darkner, Sune; Elliott, Desmond.
Findings of the Association for Computational Linguistics: EMNLP 2022. Association for Computational Linguistics, 2022. s. 7212–7230.Publikation: Bidrag til bog/antologi/rapport › Konferencebidrag i proceedings › Forskning › fagfællebedømt
Harvard
APA
Vancouver
Author
Bibtex
}
RIS
TY - GEN
T1 - Revisiting Transformer-based Models for Long Document Classification
AU - Dai, Xiang
AU - Chalkidis, Ilias
AU - Darkner, Sune
AU - Elliott, Desmond
PY - 2022
Y1 - 2022
N2 - The recent literature in text classification is biased towards short text sequences (e.g., sentences or paragraphs). In real-world applications, multi-page multi-paragraph documents are common and they cannot be efficiently encoded by vanilla Transformer-based models. We compare different Transformer-based Long Document Classification (TrLDC) approaches that aim to mitigate the computational overhead of vanilla transformers to encode much longer text, namely sparse attention and hierarchical encoding methods.We examine several aspects of sparse attention (e.g., size of local attention window, use of global attention) and hierarchical (e.g., document splitting strategy) transformers on four document classification datasets covering different domains. We observe a clear benefit from being able to process longer text, and, based on our results, we derive practical advice of applying Transformer-based models on long document classification tasks.
AB - The recent literature in text classification is biased towards short text sequences (e.g., sentences or paragraphs). In real-world applications, multi-page multi-paragraph documents are common and they cannot be efficiently encoded by vanilla Transformer-based models. We compare different Transformer-based Long Document Classification (TrLDC) approaches that aim to mitigate the computational overhead of vanilla transformers to encode much longer text, namely sparse attention and hierarchical encoding methods.We examine several aspects of sparse attention (e.g., size of local attention window, use of global attention) and hierarchical (e.g., document splitting strategy) transformers on four document classification datasets covering different domains. We observe a clear benefit from being able to process longer text, and, based on our results, we derive practical advice of applying Transformer-based models on long document classification tasks.
M3 - Article in proceedings
SP - 7212
EP - 7230
BT - Findings of the Association for Computational Linguistics: EMNLP 2022
PB - Association for Computational Linguistics
ER -
ID: 339145904