Dataset Construction and Opinion Holder Detection Using Pre-trained Models

With the growing prevalence of Internet, increasingly more people and entities express opinions on online platforms, such as Facebook, Twitter, Amazon. As it is becoming impossible to detect opinion trends manually, an automatic approach holders essential a means identify specific concerns regarding particular topic, product, or problem. Opinion holder detection comprises two steps: presence in text identification holders. The present study examines both steps. Initially, we this task binary classification problem: INSIDE OUTSIDE. Then, consider sequence labeling prepare appropriate English-language dataset. Subsequently, employ three pre-trained models for task: BERT, DistilBERT, contextual string embedding (CSE). For task, logistic regression model top layers BERT DistilBERT models. We compare models’ performance terms F1 score accuracy. Experimental results show that obtained superior performance, with 0.901 accuracy 0.924. utilize feature- fine-tuning-based architectures. Furthermore, combined CSE conditional random field (CRF) DistilBERT. feature-based architecture, five models: CSE+CRF, BERT+CRF, (BERT&CSE)+CRF, DistilBERT+CRF, (DistilBERT&CSE)+CRF. six All language are evaluated processing time. experimental indicate (DistilBERT&CSE)+CRF jointly yielded optimal 0.9453. However, CSE+CRF incurred lowest time 49 s while yielding comparable by optimal-performing