Soyutlayıcı Özetlemek, Benzerlik, Gereklilik, ve Kabul Edilebilirliği Kullanan Kapsamlı Değerlendirme Metriği
Loading...
Date
2023
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Open Access Color
OpenAIRE Downloads
OpenAIRE Views
Abstract
Uzun metinlerden otomatik olarak anlamlı özetler üretmek, birçok alanda büyük önem taşımaktadır. Transformer modeli gibi yeni sinir ağı mimarilerinin ortaya çıkması, kaliteli özetler üretebilen çok sayıda büyük dil modellerinin gelişmesine neden olmuştur. Fakat, özetleme modellerinin ürettiği özetler, önemli bir sorunu beraberinde getirmektedir. Özetleme modellerinin kalitesini ölçen, ROUGE gibi, standart otomatik değerlendirme metrikleri, kapsamlı bir değerlendirme yapmakta eksik kalmaktadır. Bu çalışmada, modeller tarafından üretilen ve insanlar tarafından yazılan örnek özetleri kullanan, SEAScore adlı yeni bir model tabanlı metrik sunuyoruz. Bu metrik, semantik benzerlik, doğal dil çıkarımı ve dilsel kabul edilebilirlik gibi çeşitli Doğal Dil İşleme yöntemlerini kullanır. Geliştirdiğimiz SEAScore metriği, daha önce eğitilmiş dil modelleri tarafından çıkarılan özellikleri kullanarak, özetleme modellerinin kalitelerini ölçen bir puan üretir. Bu tezde, üç tane özetleme modeli kullanarak yeni metriğimizin kalitesini ölçen deneyler yaptık. Deneysel sonuçlara göre, geliştirdiğimiz SEAScore metriği, bilinen standart metriklerine göre, insan tarafından üretilen değerlendirme puanları ile daha yüksek korelasyon sergileyerek başarılı sonuçlar sunmuştur.
Producing meaningful automatic summaries from long textual documents is essential in various fields. The emergence of novel neural network architectures, such as the Transformer model, has led to the development of large pre-trained language models that can produce quality summaries. However, model-generated summaries suffer from many issues. Thus, standard automatic evaluation metrics, such as the ROUGE metric, fail to effectively evaluate the quality of summarization models. In this study, we introduce SEAScore, a new model-based automatic evaluation metric that can evaluate model-generated summaries against their counterpart reference summaries by utilizing multiple Natural Language Processing tasks such as Semantic Similarity, Natural Language Inference, and Linguistic Acceptability. SEAScore takes features extracted by pre-trained language models and produces an evaluation score to measure the quality of summarization models. In this thesis, we develop our new evaluation metric SEAScore and train three summarization models to assess our new metric. Experimental results show that SEAScore correlates better with human judgment than some standard metrics.
Producing meaningful automatic summaries from long textual documents is essential in various fields. The emergence of novel neural network architectures, such as the Transformer model, has led to the development of large pre-trained language models that can produce quality summaries. However, model-generated summaries suffer from many issues. Thus, standard automatic evaluation metrics, such as the ROUGE metric, fail to effectively evaluate the quality of summarization models. In this study, we introduce SEAScore, a new model-based automatic evaluation metric that can evaluate model-generated summaries against their counterpart reference summaries by utilizing multiple Natural Language Processing tasks such as Semantic Similarity, Natural Language Inference, and Linguistic Acceptability. SEAScore takes features extracted by pre-trained language models and produces an evaluation score to measure the quality of summarization models. In this thesis, we develop our new evaluation metric SEAScore and train three summarization models to assess our new metric. Experimental results show that SEAScore correlates better with human judgment than some standard metrics.
Description
Keywords
Bilgisayar Mühendisliği Bilimleri-Bilgisayar ve Kontrol, Computer Engineering and Computer Science and Control, Derin öğrenme, Deep learning
Turkish CoHE Thesis Center URL
Fields of Science
Citation
WoS Q
Scopus Q
Source
Volume
Issue
Start Page
0
End Page
123