Article Preview
TopIn the landscape of English oral proficiency assessment, a multitude of methodologies and models have been developed to gauge linguistic competence, reflecting the ever-evolving intersection of technology and language education. This section provides a comprehensive overview of the existing approaches employed in the realm of English oral assessment, ranging from traditional methods to contemporary technological advancements. By examining the current state of the field, the aim is to contextualize this present research within the broader landscape, and highlight the gaps that motivate the exploration into the integration of deep learning and swarm intelligence for a more personalized and effective English oral education experience.
Research on automatic scoring of spoken language was carried out relatively early. As early as 2000, Witt and Young (2000) proposed a classic algorithm for scoring reading questions: the global optimization (GOP) algorithm, and it has been widely used in scoring of spoken pronunciation. However, the limitation of this algorithm is that it is a text-dependent algorithm, so the GOP algorithm is only suitable for those question types where the answers are unique. The most typical application scenario of this algorithm is the pronunciation scoring of reading questions. de Wet, Van der Walt, and Niesler (2009) conducted feature extraction on the data set of reading questions, and extracted three features: speaking speed, pronunciation, and accuracy. The pronunciation feature, leveraging the GOP algorithm, alongside three other features, was utilized to evaluate students' spoken language proficiency, resulting in a record human-machine rating correlation of 0.72. Klaus et al. (Year) have done a lot of work on a variety of speaking question types in the automatic speaking scoring project in the “TOEFL” online test, and proposed that in actual deployment, a 0.17 difference in scoring correlation with human raters is enough. The automatic scoring system can be considered effective (Zechner, Higgins, Xi & Williamson, 2009) and Automated Speech Recognition, (ASR) technology was used in speaking scoring to score two open-ended speaking questions. However, the effect is ideal. The correlation between human and machine scores is only 57.2%, which is still far from actual deployment. Su et al. (Year) proposed a human-computer spoken language-scoring method in 2017. Yoon and Zechner (2017) used a screening system to screen out some speech sounds that were difficult to ensure accurate machine scoring, and then used manual scoring methods. This part of the speech was evaluated, with manual scoring acting as a support to enhance the reliability of the automated scoring system. This approach also introduces an additional concept: employing minimal human assistance to increase the precision of the automatic scoring system. This approach not only reduces the burden caused by manual scoring, but also allows automatic scoring. The system becomes more reliable. Tao, Ghaffarzadegan, Chen, and Zechner (2016) studied a method of using deep learning to improve speech recognition. The human-machine scoring correlation of the spoken language scoring system SpeechRater, based on speech recognition, has been effectively improved. It can be seen that the accuracy of speech recognition is crucial to the spoken language scoring system (Tao, Ghaffarzadegan, Chen & Zechner, 2016).
In studying the relevant literature on automatic oral correction in foreign countries, the researchers found that because foreign language training methods and examination methods are somewhat different from those in China, foreign research on automatic oral correction mainly focuses on reading questions and follow-up questions. The research help for this article was limited, but many classic scoring algorithms are worthy of learning and reference. Based on the study of these classic algorithms, this article will improve some of the algorithms and apply them to the multi-feature intelligent correction model.