A Novel Summarization-based Approach for Feature Reduction Enhancing Text Classification Accuracy

Abstract

Automatic summarization is the process of shortening one (in single document summarization) or multiple documents (in multi-document summarization). In this paper, a new feature selection method for the nearest neighbor classifier by summarizing the original training documents based on sentence importance measure is proposed. Our approach for single document summarization uses two measures for sentence similarity: the frequency of the terms in one sentence and the similarity of that sentence to other sentences. All sentences were ranked accordingly and the sentences with top ranks (with a threshold constraint) were selected for summarization. The summary of every document in the corpus is taken into a new document used for the summarization evaluation process.

Técnicas de sumarização de textos jurídicos para suporte à classificação de documentos de decisões judiciais

“…Entretanto, pouco ainda foi explorado acerca da classificac ¸ão de textos com o auxílio de técnicas de sumarizac ¸ão. Em [Rahamat Basha et al 2019] é proposto um novo método de selec ¸ão de características para o classificador KNN (K-nearest neighbor) resumindo os documentos de treinamento originais com base na medida de importância da sentenc ¸a. A abordagem para sumarizac ¸ão de documento único usa duas medidas para similaridade de sentenc ¸as: a frequência dos termos em uma sentenc ¸a e a similaridade dessa sentenc ¸a com outras sentenc ¸as.…”

Proceedings of the Sixth Workshop on Financial Technology and Natural Language Processing

“…Polyanskaya and Brillet (2023) demonstrates the superiority of GPT-3.5 Turbo over BERT for En-glish dataset which was enhanced by the translated French dataset. (Winatmoko and Septiandri, 2023) explores ST5 and SBERT for generating embeddings and in Mishra (2023), fine-tuning Llama2 on the English dataset with prompts detailing the classification criteria gives the best result, on both the English and French datasets. The introduces an ensemble learning method with mBERT, FlauBERT, ALBERT, and MLP models, incorporating feature representations (LSA and TF-IDF), demonstrating superior performance with early fusion ensemble across all four languages.…”

“…LIPI and FinNLU also joined the Japanese subtask with the proposed methods, and SPEvFT (Mishra, 2023) applies prompt engineering to Japanese articles.…”

Identifying ESG Impact with Key Information

“…Techniques utilizing summarization for classification also fall within this category (e.g. Basha et al, 2019).…”