1 Introduction

In recent years, the application of information technology in the field of education has promoted the development of foreign language teaching models. Automatic Writing Scoring (AES) technology provides learners with dynamic automatic scoring and personalized skill guidance. The reliability and validity of automatic scoring on the online platform of domestic sentence correction website have been studied, and its feedback effectiveness has been verified, which helps to improve the quality of students' compositions and their ability to write independently. With the popularization of artificial intelligence in foreign language education, the transparency, interpretability, and prediction rate of automatic scoring algorithm frameworks are constantly being validated and iterated. The automatic writing scoring model integrates technologies such as natural language processing and machine learning to predict scores from surface language features to deep semantic content. Although the scoring algorithm of neural network paradigm has improved in model performance, there are still problems such as low interpretability and few feature dimensions. The future scoring model should be based on multi-dimensional text feature selection to construct a high-quality writing scoring model.

2 Literature Review

The research on automatic essay scoring models can be mainly divided into three categories: traditional paradigm, deep neural network paradigm, and pre training paradigm. The traditional paradigm relies on machine learning models such as multiple linear regression and manually selected text feature indicators, while the deep neural network paradigm automatically learns essay features through models such as LSTM or CNN, achieving end-to-end scoring. The pre training paradigm utilizes pre trained language models for fine-tuning, but its advantages are not significant in automatic writing scoring tasks. The multidimensional text complexity quantification features include language complexity, discourse complexity, and proposition complexity, involving aspects such as grammar spelling errors, phonological complexity, morpheme complexity, lexical complexity, syntactic complexity, and discourse coherence. This study selected 327 text feature indicators and used tools such as GAMET, TAMMI, TAALED, L2SCA, and TAASSC to calculate language errors, morpheme complexity, lexical complexity, syntactic complexity, and discourse coherence indicators to construct a multidimensional English writing evaluation model.

3 Research Design

This chapter introduces the research and design of an automatic writing scoring model based on machine learning. The research used the ASAP dataset of argumentative essays, consisting of 1783 articles, including training, testing, and validation sets. The scoring range for essays is 2-12 points, with an average word count of 150-650 words. Each essay can be reviewed by up to 3 people. The study used a multiple linear regression model and 327 text feature indicators for training and model construction, aiming to answer questions about the model's goodness of fit, prediction accuracy, and text features that affect essay scoring. The research steps include downloading the dataset, extracting essay text, calculating 327 language complexity metric features, constructing a multiple linear regression model, selecting important features, and generating an automatic scoring model algorithm. Finally, the QWK algorithm is used to calculate the consistency between machine scoring and manual scoring.

4 Data Analysis and Results

This chapter analyzes the English writing evaluation model based on machine learning. The multiple linear regression model showed successful model construction after 22 iterations, with high goodness of fit, explaining 74.3% of the variation in manual scoring scores. Compared with previous studies, the prediction rate of our research model ranks second, with a QWK value of 0.835, close to the highest value of 0.836, indicating that the model has high self consistency. In the analysis of text complexity indicators, 22 indicators were included as model parameters, covering multiple levels such as language errors, morphemes, vocabulary, syntax, and discourse coherence, indicating that composition scores are influenced by multidimensional text complexity. Fine grained syntactic indicators and phrase complexity indicators are more effective in predicting writing quality. This study also provides a matrix formula for the automatic scoring model, which provides theoretical reference for future research on constructing automatic scoring models.

5 Conclusion

The automatic scoring model in this study has good fit, strong interpretability, and high accuracy in predicting grades; The important text complexity indicator for model calculation has multidimensionality. The intelligent evaluation of foreign languages should consider multidimensional text features, explore deep multimodal model architectures, focus on model interpretability, identify writing elements and techniques, consider writing purposes and contextual influences, strengthen format and structure evaluation, pay attention to fairness issues, and integrate with teaching to promote writing ability development.