Four approaches to retrieving model English essays

Finding model English argumentative essays with AI: a comparison of four retrieval models

Guo Shiyi, Zhu Tianning, Bie Zini, Hu Jiawen  

Uppsala University (Sweden)  

Abstract:Based on data from English argumentative essays, this case study presents a system for retrieving examples of English argumentative essays that utilizes artificial intelligence and natural language processing techniques. The system integrates four mainstream models: BM25, Word2Vec, Sentence-BERT, and Elasticsearch, and a simple UI is provided to display the system’s functions. Experiments were conducted to compare the performance of these models in terms of relevance to sample texts and ranking accuracy. The results show that all four models are capable of performing the task of retrieving examples of English writing, but the Sentence-BERT model based on Transformers exhibits the highest accuracy in identifying semantic similarities and retrieving relevant texts. Traditional models such as BM25 and Elasticsearch also have practical applications when resources are limited or when keyword-based retrieval is required. The system allows users to choose the appropriate model according to teaching needs and available resources. This case study provides a reference for the development and application of intelligent tools for retrieving English teaching materials.

Keywords:Artificial intelligence; Information retrieval models; English writing instruction

In the process of teaching and learning English writing, model essays are a very important resource. Traditional methods of keyword retrieval and manual selection have significant limitations in terms of relevance and efficiency, and they cannot meet the needs of large-scale and personalized teaching. Currently, using these traditional methods for selecting model essays has certain problems: on one hand, keyword retrieval does not fully take into account the details that students are interested in in real-world situations; on the other hand, manual selection is slow, and when dealing with dozens or even hundreds of essays, the process becomes very complicated and inefficient. With the development of AI and NLP, using information retrieval models based on deep learning and semantic understanding can greatly improve the accuracy and intelligence of selecting model essays. How to use these new technologies to provide high-precision, intelligent recommendations for model essays has become a key area that needs further research. In this project, approximately 2,000 essays related to IELTS writing were collected and organized, with the aim of developing a system for recommending model essays that is efficient, intelligent, and relevant. A system integrating several mainstream retrieval models was created. The system includes BM25, Word2Vec, Sentence-BERT, and Elasticsearch. By comparing these models through experiments, their differences in accuracy and ranking capabilities were analyzed, and the advantages and disadvantages of each model in English writing teaching were also examined. The results show that the Sentence-BERT model based on Transformers is superior in terms of semantic similarity and retrieval accuracy, and it is well-suited for semantic understanding and personalized recommendations. Although BM25 and Elasticsearch are more traditional, they have advantages in keyword matching and retrieval speed, making them suitable for scenarios with limited resources or large-scale searches. The Word2Vec model has limited performance when the amount of data is limited, but it remains useful as a compromise solution that balances efficiency and basic semantic capabilities. In general, all four models can be used for retrieving model essays in different resource and demand conditions, and they can be selected based on the actual needs and the characteristics of each model.  

1. The practical needs and system innovations in the retrieval of English writing samples  

(1) Realistic problems and needs in the retrieval of model texts in English writing teaching  

In English writing instruction, the use of high-quality model texts can help improve students’ writing skills, especially for academic exam essays and task-based writing. However, since most teachers and students currently use keyword searches or manual filtering to find model texts, this method is very inefficient. Moreover, the results obtained are not very accurate, and it cannot meet the demands of increasing teaching loads as well as the individual needs of students. On one hand, traditional search methods are limited by the low degree of match between word meanings, making it difficult for students to access high-quality model texts. On the other hand, with the increasing variety of writing tasks, there is a pressing need for automated, intelligent, and highly relevant model text recommendations. With the development of artificial intelligence and natural language processing in recent years, new possibilities have emerged for searching English model texts. Information retrieval models based on deep semantic understanding can better provide personalized and precise recommendations for model texts, thereby contributing to the high-quality development of English writing instruction.  

(II) The innovative integration of information retrieval techniques and language teaching  

With the further development of AI technology, information retrieval models are presenting new opportunities for integration in language teaching. In traditional English writing teaching methods, the retrieval of model texts is usually carried out based on keywords or manually, which results in low efficiency and poor semantic understanding. These issues can be improved or resolved through information retrieval technologies. BM25, as a classic keyword probability retrieval model, maintains good retrieval efficiency and stability when applied to large-scale teaching resources, thanks to its normalization based on word frequency and document length. As the system moves toward semantic-driven approaches, word vector models such as Word2Vec, FastText, and GloVe convert words into vector forms and capture semantic similarities within context, enabling semantic synonym replacement and matching in teaching systems. Currently, Sentence-BERT, a model based on the Transformer architecture, is used to achieve a deeper semantic understanding of sentences; by generating high-quality semantic vectors, it allows for more accurate matching and ranking of the structure, logic, and arguments in writing samples, thereby enhancing the relevance of retrieval results and the quality of teaching recommendations. In the field of educational technology, numerous studies have explored the potential of AI and information retrieval technologies in writing instruction. It has been noted that the text retrieval and model-text recommendation functions provided by AI writing assistance tools can offer adaptive feedback for second-language learners, thereby helping them improve their language organization and content creation skills. Based on extensive experimental data related to language models, Yang, Zhang, and Lin (2019) concluded that using the BERT semantic retrieval model is more effective than traditional keyword matching methods for sorting informal documents. It is feasible to implement semantic-driven text retrieval using educational technology. Alves Pereira, Komarlu, and Mobeirek (2023) conducted a comparative analysis of the usage patterns of various AI writing tools from the perspective of information retrieval. They found that an improvement in the quality of user searches not only increases users’ participation in writing activities, but also significantly enhances the diversity of ideas presented in the content created by users. These studies indicate that the use of AI to assist in the creation of teaching materials has evolved from the previous approach of matching keywords and manual selection to a more advanced technique of deep semantic understanding. Moreover, this application of information retrieval techniques in language teaching aligns with the development trends of Computer-Assisted Language Learning (CALL). It can be considered a form of “Data-Driven Learning” (DDL). This teaching approach was proposed by Johns (1991); it emphasizes allowing students to actively discover language patterns through corpus retrieval, thereby enabling them to better understand language knowledge. Additionally, this approach is consistent with the development direction of Intelligent Computer-Assisted Language Learning (ICALL). For example, Amaral and Meurers (2011) pointed out that ICALL systems, which utilize Natural Language Processing (NLP) technologies, can provide language exercises based on artificial intelligence. They can also combine students’ learning data to provide personalized feedback, thereby enhancing learners’ autonomy and overall learning experience.  

(3) Introduction and presentation of the search system  

In response to the issues such as low efficiency and poor relevance in the retrieval of model examples in English writing teaching, based on the aforementioned research, this case study utilizes the IELTS writing corpus. By employing four different retrieval models—BM25, Word2Vec, Sentence-BERT, and Elasticsearch—it covers a range of techniques, from word-frequency-based retrieval to word-vector semantic calculation, and finally to deep semantic understanding. An intelligent retrieval system that integrates these various retrieval models is developed, enabling modular and integrated operation. Additionally, users can freely choose different retrieval models on the same interface, directly compare the performance of various algorithms, and conduct systematic comparisons and evaluations of the models, thereby identifying suitable teaching applications. Furthermore, to enhance the user interaction experience, a lightweight graphical interface was created using PySide6. After the program starts, it calls different model operators based on the user’s input, and the results are then presented. The page also provides a full-text preview function, which is useful for teaching and research purposes, and serves as a foundation for open-source templates for various auxiliary functions in the future.  

II. Comparison of retrieval models in essay retrieval: experimental schemes and evaluation methods  

(1) Research issues  

This study aims to verify the feasibility and practical value of information retrieval methods applied to the retrieval of English writing models. It also attempts to examine the performance of various information retrieval methods such as BM25, Word2Vec, Sentence-BERT, and Elasticsearch in this context. Through experiments, the effectiveness of these traditional retrieval methods, neural network models, and well-established commercial search engines is compared. Finally, the educational implications and limitations of these methods are analyzed, along with recommendations for using different models in actual English writing instruction.  

(2) Description of the dataset  

The “IELTS-writing-task-2-evaluation” dataset used by this research comes from the HuggingFace open platform. It contains over 9,000 essays on IELTS Task 2 written by real test takers. Each essay is accompanied by a manual score and a specific prompt. The topics covered in these essays include education, technology, and the environment. Since the focus of this study is to examine how various information retrieval models perform in retrieving relevant information in English writing tasks, rather than assessing or evaluating the English writing skills of the candidates, the writing prompts and the articles themselves were used to construct the word extraction model and for subsequent comparative analysis. The scoring criteria were not utilized in this study.  

(3) Model introduction and similarity calculation methods  

The four information retrieval models used in this paper are as follows: the traditional BM25 model based on word frequency, the Word2Vec word vector model that has been widely used in recent years, the Transformer neural network model Sentence-BERT, which is used for in-depth analysis of sentence semantics, and Elasticsearch, an industrial-grade search engine that is widely applied today.  

BM25 (Best Matching 25) is a classic term matching retrieval model. It was first proposed by Robertson and Walker. Its simplified version is widely used in various information retrieval systems, such as Elasticsearch and Lucene. BM25 utilizes non-linear term frequency functions and document length normalization to address the issues caused by overly frequent keywords or overly long documents, which can affect the relevance score. However, it remains a shallow matching method based on term overlap; it cannot extract more information from a semantic or contextual perspective. In this experiment, we used the BM25Okapi implementation from the rank_bm25 library to implement the BM25 model and call its scoring function. As mentioned above, BM25 takes into account term frequency, document length, and the distribution of terms in the corpus, in order to evaluate the relevance between queries and documents.  

Word2Vec is a word embedding model developed by Google in 2013. This model draws upon J.R. Firth’s idea that “the meaning of a word depends on its context.” It maps words into a vector space, where words with similar meanings are located closer in this space. This method is then used to calculate the semantic relationships between words. In this study, the Gensim library was used for implementation, and articles were sorted and retrieved in a descending order based on cosine similarity. Although Word2Vec can capture superficial semantic relationships, since it relies only on local context-related word vectors, it cannot reflect multi-word meanings or deeper semantic aspects. First, we used NumPy to obtain the average word vectors for the queries and each essay, and these vectors were used as the text vectors. Then, the articles were sorted in descending order based on cosine similarity, thereby achieving semantic-based retrieval results.  

Sentence-BERT (SBERT) is an improved version of the classic BERT model. By integrating Siamese and Triplet network architectures, it can efficiently generate sentence-level semantic representations and calculate the similarity between sentences. SBERT supports various pooling strategies, including CLS token pooling, average pooling, and maximum pooling. In this study, we used the pre-trained model paraphrase-MiniLM-L6-v2 provided by HuggingFace, and the sentence-transformers library was utilized to extract sentence vectors. The vector dimensions of this model are 384; thus, it can be directly used for semantic retrieval tasks without further fine-tuning. In terms of scoring mechanisms, similar to Word2Vec, SBERT uses cosine similarity to measure the semantic similarity between sentences or texts. A higher score indicates a stronger semantic relevance.  

Elasticsearch is a distributed, simple, and flexible search and analysis engine developed based on Apache Lucene. It allows for the easy configuration of multi-node clusters, as well as rapid response to various types of queries. In this paper, Elasticsearch v8.7 was used for the design and development. The system adopts a hybrid architecture that combines inverted indexes with JSON document storage. Automated methods are used to set up nodes and manage their operation, ensuring the accuracy of data. Additionally, the original text, its segments, and processed tokens are stored together in the index structure. The `copy_to` field is used to improve search relevance. The BM25 parameters (k_1=1.5, b=0.8) are applied, which is suitable for articles of different lengths. Finally, in this paper, the document relevance scoring mechanism of Elasticsearch is considered to be a modified version of the BM25 model. The relevance scoring and sorting are performed using the BM25 algorithm; only the implementation details such as the idf calculation method are followed according to the official specifications.  

(4) Experimental evaluation methods and indicators  

In information retrieval systems, “K” is often used to refer to the number of search results that are displayed to users after each search attempt, up to K results (Top-K). This setting determines the scope of user browsing, and it is an important parameter in various retrieval system evaluation metrics such as Precision@K, Recall@K, and NDCG@K. To better reflect the actual performance of the retrieval system, this study introduces two evaluation methods: static K and dynamic K. For static K, a fixed value is used (e.g., K=10 or K=20), which simulates the situation where users only browse the first two pages of search results. This approach helps to assess whether the model can capture more highly relevant documents under a limited amount of exposure; this parameter reflects the actual browsing habits of users. From user experiments, it can be seen that most users only read the 10th to 20th results during the search process. In such cases, it is most appropriate to evaluate the Top-K retrieval model. For dynamic K, a method is adopted where the value of K varies depending on the actual number of relevant documents for each search task. K is set to 0.8, 1.0, 1.2, 1.4, or 1.6 times the actual number of relevant documents (denoted as K_ratio). This approach avoids the problem of unbalanced evaluations caused by values that are too high or too low. Based on this, comparisons are made between models for different sizes or difficulty levels of search tasks, along with the dynamic changes in the precision and recall rates.  

In order to more accurately assess the performance of various retrieval models in the task of retrieving English writing samples, we have used a variety of commonly used information retrieval evaluation metrics. Precision measures the proportion of relevant documents among the first K results returned by the system, indicating the efficiency with which users can obtain high-quality documents. Recall represents the proportion of relevant documents that are successfully retrieved by the system, reflecting the system’s ability to find all relevant documents. The F1 score is the harmonic mean of precision and recall, used to evaluate the overall balance between the “true positive rate” and “true positive rate” of the retrieval system. Mean Average Precision (MAP@K) measures the overall ranking quality of the model for all queries: for a single query, the precision of each relevant document is calculated first, then the average is taken, and finally the average is computed for all queries. Normalized Discounted Cumulative Gain (NDCG@K) assesses whether the most relevant documents are ranked at the top, providing a more user-friendly experience.  

(5) Experimental procedure  

In order to comprehensively evaluate the applicability and effectiveness of various information retrieval models in the task of retrieving English writing samples, the following experimental procedure was designed in this study:  

1. Experimental data preprocessing: The IELTS Essay Task 2 dataset was downloaded from HuggingFace. The first 2,000 records were selected, and appropriate text cleaning was performed. Subsequently, standardized preprocessing steps such as text cleaning, tokenization, removal of stop words, and punctuation were applied.  

2. Experimental model integration: The BM25 retrieval model from the rank-bm25 library, the Word2Vec-Cosine retrieval model based on Gensim and sklearn, the Sentence-BERT retrieval model that uses pre-trained models from the sentence-transformers library and cosine similarity, and the Elasticsearch retrieval model implemented using the elasticsearch-py interface and a back-end inverted index are all integrated together. Due to the differences in the retrieval mechanisms of these four different models, a RandomIR model was additionally developed to facilitate a comparison of their performance. In this model, a list of K randomly selected files is returned each time, without considering any correlation between the query and the files. All models are written in Python, and they implement class-based encapsulation for functions such as multi-model retrieval and returning results in a sorted order to the user. This allows for easy invocation and quick switching between different retrieval methods.  

3. Setup and preparation of retrieval tasks: The experimental design involves using writing prompts as inputs for the retrieval process, thereby simulating the scenario of actual user retrieval attempts. To minimize the impact of bias caused by a single amount of ground truth, five different query values were used; the number of articles corresponding to each ground truth value was 200, 112, 63, 22, and 397 respectively. These values follow a stepped distribution, and they are randomly divided into two parts—80% for training data and 20% for test data. Finally, a performance comparison analysis is conducted between retrieval tasks with different scales of ground truth (large and small) to ensure the credibility of the experimental results, thereby enabling a better assessment of how data size affects the performance of the retrieval tasks.  

4. Comparison of multiple models and output of results: All models were tested on a common platform and dataset. Detailed retrieval performance under various evaluation metrics and K values was reported, and relevant charts were created for easy horizontal comparison and comprehensive analysis.  

III. Experimental Results, Analysis, and Discussion  

(1) Analysis and interpretation of dynamic K experiment results  

Under the dynamic K setting with k_ratio=1 (i.e., K equals the number of true labels per query), the average MAP (Mean Average Precision) of the five models across five sets of queries showed significant differences. As a random baseline, RandomIR had an MAP value of only 0.01, indicating that its retrieval results had little to no relation to the ordering of relevant documents. Word2VecIR had an average MAP of 0.64, which was significantly inferior to the other mainstream models. Both BM25IR and Elasticsearch achieved an average MAP of 0.86, indicating reliability and robustness based on term matching and industrial-grade search engines. Sentence-BERT performed best, with an average MAP of 0.91, which was significantly better than the other methods; this indicates that it has a clear advantage in comprehensive semantic understanding and relevance in document ordering. Overall, the deep semantic model (Sentence-BERT) has stronger capabilities in ordering relevant documents in English writing sample retrieval tasks, while BM25 and Elasticsearch provide reliable retrieval performance. Word2Vec shows certain limitations under small datasets.  

Except for RandomIR, the average nDCG values of the other models are relatively high and stable under various k_ratio values. This indicates that mainstream retrieval models can still place most relevant documents at the top of the search results, achieving an ideal forward ranking effect, even when the number of candidate results changes. In particular, when k_ratio=0.8, the nDCG values of BM25IR, Elasticsearch, and Sentence-BERT are all close to 1, indicating that both ranking accuracy and user experience are well ensured. Further analysis shows that Sentence-BERT consistently ranks first, with minimal variation as k_ratio changes, demonstrating the strong robustness of deep semantic models in document retrieval. BM25IR and Elasticsearch also maintain values above 0.95, indicating stable and reliable results. Word2VecIR, on the whole, is lower than the above three models, but it shows a significant improvement compared to the random method (RandomIR, with nDCG values of only 0.05–0.1), which clearly demonstrates the advantage of mainstream models in ranking relevant documents.  

The F1 scores of different retrieval models vary slightly. However, when k_ratio is set to 1 for all models, both accuracy and recall are optimal, and the F1 scores reach their peak levels. It can be seen that the Sentence-BERT model has the highest F1 score, almost reaching 0.9, which is several times higher than that of other models. Therefore, this model exhibits good overall performance in document retrieval. As k_ratio increases, the F1 scores of most models gradually decrease. In particular, BM25IR and Elasticsearch drop from 0.86 to around 0.7, while Word2VecIR remains at a relatively low level, never exceeding 0.7. RandomIR always has a very low F1 score; it is used merely as a baseline for comparison. It is evident that a more appropriate value for k_ratio (such as k_ratio=1) allows for a more accurate representation of the retrieval model’s performance, without causing a decrease in precision as a result of an increase in the overall accuracy.  

As k_ratio increases, the precision of each model generally decreases, while the recall rate gradually increases. This reflects the inherent trade-off between these two metrics in the retrieval task. It is worth noting that Sentence-BERT remains at the forefront across all values of k_ratio: at k_ratio=0.8, its precision reaches 0.98, and at k_ratio=1.6, its recall rate reaches 0.93. This trend is consistent with the performance of the F1 score earlier; it indicates that this model has a significant advantage in balancing precision and recall.  

In addition to comparing the final ranking results of each model, we also present the PR curves and AUC values for all models. The PR curve shows the proportion of relevant documents obtained by each model under different recall rates, i.e., their values at different retrieval thresholds. Additionally, the AUC of each model was calculated to provide a comprehensive measure of the model’s overall retrieval ability. A higher AUC indicates that the model can always find highly relevant documents at some stage and maintain a high precision rate; conversely, if no highly relevant documents can be found at a certain stage, then the retrieval performance at that stage is poor. The best model is one that meets this requirement at every stage; this also indicates that the model is able to retrieve highly relevant documents within a certain range and maintains a high precision rate. As can be seen from the results, Sentence-BERT has the highest AUC value, at 0.14. It demonstrates excellent retrieval performance under all different threshold conditions, while the AUC of RandomIR is only 0.004, which is much lower than that of other mainstream models.  

(II) Analysis of Static K Experiment Results  

The static K evaluation involved using K=10 and K=20 to assess the performance of various models in simulating the retrieval process of real user text. By comparing the performance of different models in simulating real user text retrieval scenarios, it was found that the three mainstream models—BM25IR, Sentence-BERT, and Elasticsearch—achieved a result of 1.00 in both Precision@10 and nDCG@10. This indicates that they have good ability to retrieve all relevant texts at the 10th hit; Word2VecIR also achieved good results (Precision@10 was 0.88, nDCG@10 was 0.89); while RandomIR achieved the lowest values for all metrics.  

When the threshold is raised to K=20, BM25IR, Sentence-BERT, and Elasticsearch still maintain very high retrieval performance (Precision@20 is 0.97–1.00, and nDCG@20 is 0.98–1.00); only slightly decreased compared to K=10. The MAP@20 values of all models increase (BM25IR: 0.30, Sentence-BERT: 0.31, Elasticsearch: 0.29), but overall they are still significantly lower than Precision and nDCG. This is mainly because, in some queries, the number of relevant examples far exceeds the K value, resulting in a lower MAP score, as the denominator is the total number of relevant documents. Word2VecIR shows further declines in all metrics at K=20 (Precision@20: 0.86, nDCG@20: 0.87, MAP@20: 0.16), while RandomIR always performs very poorly. It should be noted that even though the mainstream models achieve near-perfect performance in terms of Precision and nDCG, the MAP value remains significantly lower than the other two due to the static K setting and the correlation distribution. This indicates that when the number of relevant examples far exceeds the retrieval threshold K, only when most of the relevant documents are ranked at the top does the MAP become close to 1, similar to Precision and nDCG. Overall, BM25IR, Sentence-BERT, and Elasticsearch demonstrate high robustness and ranking ability in topic-based example retrieval, while Word2Vec becomes more unstable in terms of ranking-sensitive metrics as the K value increases. This difference further confirms the effectiveness of complex retrieval strategies in retrieving example texts in argumentative essays.  

(3) Discussion of experimental results  

This study employed both dynamic K and static K evaluation methods to compare the performance of various retrieval models in the task of retrieving English writing samples. The experimental results show that the Sentence-BERT model outperforms other models significantly across most metrics—whether it is MAP, nDCG, or F1—by consistently ranking highly relevant documents at the top, and it also exhibits high robustness under different threshold settings. BM25 and Elasticsearch have certain advantages in traditional keyword matching and retrieval efficiency; however, when computational resources are limited, they still maintain high retrieval accuracy according to the experimental results, making them suitable for applications where resources are constrained or efficiency is a priority. Word2Vec, on the other hand, suffers from limitations related to the size of training data and its ability to handle polysemous words, resulting in relatively poor performance under the current experimental conditions. In the dynamic K evaluation, it was found that adjusting the K value allows for a comprehensive assessment of the balance between precision and recall among different models. The results also suggest that setting the dynamic K value appropriately is helpful for systematic evaluation. In the static K experiments, to better reflect real user retrieval scenarios, K values were set at 10 and 20, and the performance of various models was observed. BM25IR, Sentence-BERT, and Elasticsearch all achieved values close to 1.0 in Precision@K and nDCG@K metrics, allowing them to consistently rank relevant documents at the top. As the K value increases, ranking-related metrics such as MAP improve significantly, but since the number of relevant documents for certain queries far exceeds the K value, the MAP score is generally lower than that of Precision and nDCG. Word2Vec’s performance declines further in this evaluation, indicating its sensitivity to training data and semantic ambiguity issues. Overall, these results indicate that in everyday teaching or scenarios where the requirement for retrieval accuracy is not very high, mainstream retrieval models can achieve satisfactory results. In practical applications, different models can be selected based on specific needs and system resources.  

Sentence-BERT demonstrates excellent capabilities in understanding deep semantic meanings; it can accurately capture the meaning in complex texts and enable efficient retrieval. Reimers and Gurevych (2019) pointed out that, while generating high-quality sentence vectors, it significantly reduces the response time of large-scale semantic retrieval. This performance makes it highly suitable for educational tasks that require strong semantic understanding, such as personalized recommendations, intelligent writing assistance, and the provision of advanced model examples. However, it also requires more computing resources, so it is suitable for deployment in environments with better hardware conditions. Traditional retrieval models like BM25 and Elasticsearch still have wide-ranging applications in educational systems due to their stable keyword matching capabilities and low resource consumption. In the BEIR benchmark, BM25 serves as a baseline for zero-sample retrieval; it performs better than most sparse and dense retrieval models in multi-domain document retrieval tasks, demonstrating its practicality and stability. Therefore, in scenarios where efficiency and system stability are important, such as indexing course names, locating keywords in textbooks, or rapid retrieval of teaching resources, BM25/Elasticsearch may be a more appropriate choice. Word2Vec can effectively measure the semantic similarity between words; theoretically, it has the potential to expand expression diversity and is suitable for auxiliary tasks like automatic rewriting. However, studies show that when the training data size is limited or the context is complex, Word2Vec tends to produce low-quality word vectors, leading to inaccurate semantics and false similarities (for example, “bank” having ambiguous meanings in different contexts, or “cat” and “chair” being misidentified). Thus, Word2Vec is more suitable as a supplementary tool in writing education systems and should be used carefully, in conjunction with large-scale high-quality data and specific task requirements. In summary, each of these models has its own strengths: Sentence-BERT is ideal for deep semantic understanding and intelligent recommendations; BM25/Elasticsearch meet the needs of basic keyword retrieval and efficiency; Word2Vec can serve as a supplementary tool for tasks involving expression diversity.  

IV. Summary of Cases  

(1) Research conclusions  

In contrast to the current majority of related studies that only select one technical approach to compare its performance, this paper examines various mainstream information retrieval models such as BM25, Word2Vec, Sentence-BERT, and Elasticsearch. On the task of retrieving model essays in English writing, different technical approaches are comprehensively compared and analyzed. The study finds that Sentence-BERT, a neural network semantic model based on the Transformer architecture, performs best across all metrics, accurately reflecting the deep semantic relationships between documents. It achieves good stability for various types of queries, including topic-specific and full-text queries. Although traditional methods based on keywords or inverted indexes such as BM25 and Elasticsearch offer high performance and stability when using limited computational resources or focusing on keyword-based matching, they cannot support semantic search and are difficult to use for basic writing tasks when there is insufficient training data. Additionally, the typical word vector neural network model Word2Vec requires a large amount of training data; in this experiment, due to the limited scale of the experimental data, Word2Vec fails to distinguish and match polysemous words and complex semantics as well as Sentence-BERT, highlighting the need for large-scale high-quality data in real-world applications. From an educational application perspective, all four mainstream models perform well in model essay intelligent retrieval tasks: Sentence-BERT is suitable for scenarios where high-quality model essays are required and performance is important; BM25 and Elasticsearch have advantages in simplicity of operation, rapid response, and the need for minimal hardware resources, making them ideal for large-scale teaching retrieval, as well as daily searches or basic writing tasks. Word2Vec can provide a certain level of semantic capability and serves as a balanced solution under limited resource constraints. By choosing the appropriate model configuration based on platform computing resources, application requirements, and development timelines, the optimal cost-performance ratio can be achieved, satisfying both application effectiveness and speed. This study demonstrates that neural network models are superior to other models in the field of intelligent retrieval of English writing, and provides valuable insights for model selection and optimization of writing assistance systems in different educational contexts. Future efforts will focus on multi-model integration and system integration to promote the informatization and intelligence of English writing education.  

(2) Limitations  

Although the research system has compared the performance of various information retrieval models in retrieving sample essays on English writing, there are still certain limitations. First, the IELTS datasets used currently have an uneven distribution of topics, with a strong tail effect. In practical applications, the predictions made by these models tend to focus on topics that are easier for the models to handle. However, for less popular or interdisciplinary topics, the accuracy of retrieving sample essays is relatively lower. Second, the current method of determining relevance relies solely on label consistency, without taking into account the authenticity, topic relevance, and semantic diversity of the essays. There is no authoritative standard yet, let alone a way to select samples based on consistency labels. In modeling, relevance MAP and NDCG are mainly used, but these methods cannot fully quantify the performance of these models when there are a small number of relevant samples and varying article lengths. In terms of algorithm optimization, both improving algorithm efficiency and model performance, as well as enhancing writing quality and teaching applications, are areas that require further research. The sample size used for parameter sensitivity analysis is limited.  

(3) Future prospects  

This study has demonstrated the effectiveness of mainstream information retrieval models in retrieving model essays for English writing. However, further research can be conducted in the following directions in the future: First, increase the amount of data and expand the sources of data used; Second, on the basis of the selected models, combine other models to explore methods of multi-model integration, thereby improving the model’s ability to retrieve complex semantics and diverse expressions; Third, systematically examine the impact of text preprocessing techniques such as word form restoration, stem extraction, and stop word removal on model performance; Finally, consider user habits and preferences, so that the system can better integrate into real teaching scenarios, allow for feedback optimization, and utilize its data resources for further functionality enhancement. With the development of technologies such as artificial intelligence and natural language processing, the model essay retrieval system will gradually become more useful in English writing instruction.