Designing AI-powered translation education tools: a framework for parallel sentence generation using SauLTC and LLMs

Introduction

As the world becomes increasingly interconnected, often described as a “global village”, translation emerges as a pivotal practice facilitating communication and understanding across diverse languages. Translation bridges cultural differences and plays a vital role in fostering social and cultural development (Zhao, Li & Tian, 2020). The domain of translation education (TE) is central to preparing future translators and interpreters, equipping them with the necessary skills to proficiently translate written and spoken content across linguistic systems. This critical field shoulders the responsibility of enabling global communication. However, TE is inherently dynamic and continuously evolving to keep pace with advancements in translation practices and pedagogical approaches (Igualada & Echeverri, 2019).

Academic translation programs are shaped by two primary forces: the evolving demands of the translation market, driven by technological innovations, and progressive trends in higher education teaching methodologies (Sawyer, Austermühl & Raído, 2019). Recent technological advancements, such as artificial intelligence (AI), big data, and deep learning, have revolutionized the translation industry (Zhao, Li & Tian, 2020; Al-Batineh & Al Tenaijy, 2024). Concurrently, higher education has experienced transformative shifts, including the shift from teacher-centered to student-centered learning models in the late 1990s (Igualada & Echeverri, 2019) and the widespread adoption of competency-based training at the start of the 21st century (Albir, 2017).

To remain relevant, translation programs must adapt their curricula to integrate these technological advancements and align with modern pedagogical practices. This requires the incorporation of state-of-the-art translation tools and the adoption of learning approaches that reflect market needs (Sawyer, Austermühl & Raído, 2019). In response, scholars have underscored the importance of improving existing teaching tools and developing new resources to enhance TE (Zapata, 2016). Key recommendations include revising curricular designs and refining assessment methods to better reflect real-world translation practices (Albir, 2017; He, 2021).

The PACTE group believes that a pivotal aspect of this transformation is the emphasis on translation competence, grounded in cognitive-constructivist and socio-constructivist learning theories (Albir, 2017). The authors in He (2021) advocates for student-led teaching approaches that leverage technological resources. By encouraging students to independently engage with multiple platforms, revise translations, and conduct peer assessments, TE can better reflect the demands of the contemporary translation landscape. Considering emerging technologies, He (2021) emphasizes that translators’ primary role will increasingly involve post-editing machine-generated translations, highlighting the need to cultivate post-editing competencies within TE programs.

The advent of AI, particularly in the domains of natural language processing (NLP) and natural language understanding (NLU), has introduced sophisticated tools capable of supporting these objectives. AI-powered applications can personalize learning experiences, analyze student translations, and automate certain assessment processes, thereby enhancing efficiency and objectivity. Such innovations ensure that students remain competitive in the evolving translation market by acquiring the necessary skills to collaborate with AI tools rather than being replaced by them.

Despite these advancements, a significant challenge persists: the scarcity of high-quality, comprehensive datasets tailored to TE. A parallel corpus, in this context, is defined as a collection of texts in two languages that are direct translations of each other. The Saudi Learner Translation Corpus (SauLTC) represents a notable attempt to address this gap. Developed to facilitate the teaching of translation from English to Arabic (Al-Harthi & Al-Saif, 2019), SauLTC is a multi-version, English-Arabic parallel corpus featuring part-of-speech tagging. It provides two target versions of 366 source texts (STs), allowing for the analysis of translation and revision processes from initial to final drafts. This corpus supports error identification, the evaluation of teaching feedback, and the investigation of individual differences in manual verification. The SauLTC project further aims to develop AI-based translation technologies and eLearning resources to support digital TE.

In alignment with the evolving landscape of TE, this research seeks to build upon SauLTC by developing a parallel didactic corpus enriched with an innovative annotation scheme. A key component of this process is parallel sentence generation, which refers to the automated process of identifying or constructing sentence pairs that convey the same meaning across bilingual texts. Additionally, alignment is the process of matching sentences from two different languages, where each sentence is a translation of the other. This scheme aims to enhance students’ translation competence, foster self-directed learning, and promote autonomy (F Alshihri, M Alharthi, 2024, unpublished data). By focusing on linguistic equivalence between English and Arabic, the annotation highlights structural and lexical correspondences, encouraging learners to reflect on key aspects of the translation process, such as translation units, problem-solving strategies, and the impact of errors.

Moreover, the proposed tool incorporates error analysis and post-editing features, allowing students to identify and rectify inaccuracies systematically. This hands-on approach equips students with the practical skills needed to address real-world translation challenges, bridging the gap between theoretical knowledge and applied practice.

The first stage of this project outlines a framework for transforming SauLTC’s parallel texts into parallel sentences, laying the foundation for future annotations that stimulate critical thinking and active engagement with translation problems. By harnessing AI and large language models (LLMs), the framework facilitates the generation of high-quality parallel sentences, ensuring contextual appropriateness and linguistic precision. The integration of AI-driven tools enhances the pedagogical value of the corpus, offering students a more comprehensive learning experience.

Ultimately, the development of a SauLTC-based parallel sentence dataset promises to advance AI-powered TE tools, providing students with invaluable resources for practical analysis, comparative evaluation, and cultural competency. This initiative supports instructors in delivering targeted feedback, fostering error correction, and enhancing classroom engagement.

The remainder of this article is organized as follows: “Translation and Technology” reviews the role of technology in translation education, while “Related Work” explores recent initiatives in corpus development for TE. “Methodology and Proposed Parallel Sentence Generation Framework” describes the methodology and framework for parallel sentence generation. “Experiment and Results” presents the results and analysis, and “Discussion” concludes with future research directions.

Translation and technology

The rapid advancement of information technology, notably AI technology, is one of the realities affecting translation services and teaching/training (Garbovskiy & Kostikova, 2020; Kanglang & Afzaal, 2021; Wang, 2023; Mohamed et al., 2024; Mohsen, 2024), resulting in a process of constant evolution in the field (Sawyer, Austermühl & Raído, 2019; Kenny, 2019). The translation industry is now AI-driven which improved the levels of machine translation accuracy and efficiency (Zaghlool & Khasawneh, 2024). Machine translation tools and CAT tools are constantly updated and upgraded because of the advances in artificial intelligence technologies (Steigerwald et al., 2022; Mohamed et al., 2024), and this means that the translator, who is expected to translate a huge number of texts quickly in a little amount of time (Rodríguez De Céspedes, 2019), needs to be kept abreast of the latest translation technologies, and therefore, these technologies should be a core component in any TE (Kenny, 2019). In fact, translators will not have a place in today’s translation market if they do not know how to use translation technology (Bowker, 2023).

Technological innovations have led to huge transformations in the translation profession, education, and assessment (Kong, 2022). One of the most prominent transformations is the Computer-Assisted translation (CAT) change of the translators’ role into post-editors (Igualada & Echeverri, 2019). Therefore, translation students need to be trained to use the most recent AI-based translation technologies to enhance their learning experience and empower them to meet the rising demands of the market (Zaghlool & Khasawneh, 2024). Embracing AI technologies provides teachers with ample opportunities to create high-quality translation environments with favorable conditions to improve students’ foreign language skills, improve their translation skills and broaden their horizons, and support active learning (Kong, 2022; Liu, 2022; Wang, 2023).

Related work

Learner translation corpora (LTC) are comprised of translations produced by students, either into their first language (L1) or a second language (L2), aligned with their STs. This subset represents a specialized area within corpus-based translation studies (CBTS), which typically focuses on professional or expert translation corpora (Lefer, 2020). The emergence of LTC began in the early 2000s, approximately a decade following the establishment of CBTS. Foundational initiatives such as PELCRA (Uzar & Waliński, 2001) and MeLLANGE (Castagnoli et al., 2011) were instrumental in shaping and advancing LTC. Espunya (2014) highlights the primary objectives of LTC, which include analyzing the development of translation skills, assessing the efficacy of pedagogical approaches, and developing resources for translator training. Similarly, Andrey & Kunilovskaya (2014) present a comprehensive research plan based on their work with the Russian Learner Translator Corpus (RusLTC), focusing on translation variability, decision-making processes, and the identification of challenging linguistic areas.

Drawing from the MISTiC corpus, Castagnoli (2020) observes that full lexical consistency is most evident in the translation of concrete nouns, functional words, and numerical data, while abstract nouns and metaphorical expressions frequently exhibit greater variability. A more recent contribution to the field of language and translation education is the Multilingual Student Translation Corpus (MUST) which is a multilingual LTC project that provides extensive metadata for translation research (Granger & Lefer, 2020a, 2020b).

Moreover, as part of the MUST initiative, the translation-oriented annotation system (TAS) was developed to enhance translator training and facilitate research on translation quality across various language pairs (Granger & Lefer, 2021).

The multifunctionality of LTC is evident in its diverse applications. These corpora are integral to translator education, offering students exposure to parallel sentence structures that deepen their understanding of linguistic patterns and cultural nuances. They also serve as valuable resources for research on translation quality by facilitating comparative studies of various translation strategies. Furthermore, LTC supports the development of machine translation (MT) systems and computational linguistic studies, underscoring its relevance across both academic and professional domains.

Annotation systems employed in LTC play a critical role in enhancing the analytical process of student translations. Commonly used annotation methods include part of speech (POS) tagging, which labels each word with its grammatical category, allowing for the examination of syntactic patterns and learner language use. Parsing, another key method, analyzes sentence structures to identify syntactic dependencies and relational hierarchies. Error annotation, a cornerstone of LTC research, involves marking errors within learner translations to pinpoint areas of improvement, such as grammatical inconsistencies or lexical inaccuracies.

Several taxonomies within LTC vary widely, ranging from simple frameworks comprising five categories to more elaborate schemes encompassing over 50 classifications. While projects such as RusLTC and KOPTE have developed comprehensive taxonomies for English, French, and German pairs, some annotation systems lack sufficient documentation, posing challenges for achieving consistency in error categorization. Nevertheless, error annotation yields valuable insights into translation competence development. For instance, Wurm (2020) leverages KOPTE error data to investigate the empirical dimensions of translation proficiency, analyzing factors such as time spent abroad, media exposure, and temporal changes in student performance. Her findings indicate that intensive training significantly reduces errors, enhances problem-solving efficiency, and accelerates translation speeds, whereas external variables such as international experience exhibit limited influence on the acquisition of translation competence.

An emerging trend within CBTS is the annotation of corpora focusing on student post editing. Researchers, including Kübler, Mestivier & Pecman (2022), have focused on identifying errors in editing, particularly those involving complex noun phrases in specialized discourse. Their taxonomy categorizes errors into three groups; overconfidence in MT where incorrect MT output remains unchanged; underconfidence in MT, where correct MT output is unnecessarily altered; and failure to correct MT, where errors in MT are acknowledged but not adequately corrected. In their corpus, the error type concerning noun phrases is failure to correct MT.

Further advancing this domain, Lefer, Piette & Bodart (2022) introduced the Machine Translation Post-Editing Annotation System (MTPEAS), It consists of seven categories; value adding edits, successful edits, unnecessary edits, incomplete edits, error introducing edits, unsuccessful edits and missing edits. They have effectively integrated this system with the TAS developed by Granger & Lefer (2021). This classification system helps in structuring annotations for student editing datasets, allowing for the identification and clarification of incorrect segments that persist in the final edited texts. This approach enhances the investigation and understanding of errors encountered during student editing tasks.

These initiatives collectively enrich the scope of LTC research, fostering advancements in translation education and contributing to the diversification of corpus-based translation studies. By accommodating multiple languages, text genres, and proficiency levels, LTC continues to serve as a vital resource for advancing translation pedagogy and practice.

SauLTC represents a unidirectional, multi-version parallel LTC, comprising student translators’ graduation projects, some of which have undergone professional editing and subsequent publication. SauLTC offers two distinct translation versions: a sub-corpus of unedited draft translations reflecting the initial outputs of student translators, and a final sub-corpus containing professionally edited versions. The SauLTC software facilitates both independent searches within each sub-corpus and parallel searches across two or three sub-corpora. The availability of draft and final submission sub-corpora allows for the differentiation between linguistic features attributable to the initial translation process and those arising from post-feedback revisions. By analyzing both draft and final versions, researchers can gain deeper insight into the translation process, offering valuable contributions to the understanding of translation development and competence (Kruger, 2012; Bisiada, 2017). This dual-layer structure enables the identification of mediations occurring during the transition from source to target text and through subsequent editorial interventions. These mediations may manifest as “translationese” (unusual features characteristic of translated texts) or as editorial modifications aimed at enhancing lexical diversity, simplification, or explicitation. The dual availability of draft and final versions positions SauLTC as an essential resource for process-oriented, corpus-based translation studies. The authors in Alasmri & Kruger (2018) highlighted the scarcity of corpus-based translation Arabic research. Thus, to address this scarcity, particularly in learner corpus research and corpus-based translation studies, this article works towards a more accurate sentence-alignment system for SauLTC.

Sentence alignment methods fall into three main categories: length-based, lexical-based, and neural approaches. Length-based models (Church et al., 1993) align sentences based on length correlation, which has been effective for some European languages but struggle with morphologically rich languages and those with different character structures. Lexical-based methods (Varga et al., 2007) align sentences using bilingual dictionaries and word-overlap heuristics, making them effective for direct translation equivalents. However, they struggle with structurally divergent translations and paraphrased sentences, where direct word matches are insufficient for accurate alignment. Authors in Grégoire & Langlais (2018) introduced a neural-based parallel sentence extraction system that maps sentences into a shared vector space, improving alignment accuracy over traditional methods. Their fully neural approach eliminates the need for predefined feature engineering but relies on large-scale parallel corpora, making it less applicable to low-resource languages. Additionally, their model does not handle out-of-vocabulary (OOV) words, which may lead to alignment errors when dealing with rare or unseen terms. The Parallel Hierarchical Attention Network (PHAN) (Zhu, Yang & Xu, 2020) further improves alignment by assigning different weights to key words, yet it also depends on large parallel datasets for training, posing challenges for low-resource language pairs.

Methodology and proposed parallel sentence generation framework

To address the problem at hand, this section details the methodological framework employed in this study to derive a didactic corpus of parallel sentences from the SauLTC. The chosen approach ensures a robust and systematic investigation of recent LLMs in achieving parallel sentence generation from document pairs. The following sub-sections dive into the specific methods utilized to achieve the research objectives. Figure 1 shows the end-to-end methodological framework adopted in our study. It encompasses the following main phases:

Data preparation

The data preparation phase is critical in developing LLMs, especially when dealing with the complexities of English-Arabic texts. This phase involves several key steps to ensure the dataset is optimized for subsequent analyses:

• Initial data assessment and correction: We began with a comprehensive review of the dataset to identify and document any discrepancies or mismatches. This included checking for instances where English STs did not have corresponding Arabic translations and vice versa. We also addressed and corrected any data entry errors, particularly where texts are mislabeled or placed in incorrect database columns. Ensuring that each text is correctly categorized is essential for accurate data handling and subsequent processes.

• Data cleaning: We meticulously cleaned the dataset by identifying and eliminating non-essential columns such as ID, group number, version, and year. This step simplifies the dataset and focuses attention on the core aspects of the translation corpus.

• Data filtering: We filtered out all entries containing preliminary draft texts, opting to retain only those entries that included the final revised versions of the Arabic translations corresponding to their English STs. This selective inclusion ensures that our analysis is based on the most accurate and refined data available.

• Text preprocessing: The text preprocessing phase involved carefully planned steps to preserve the integrity and meaning of the text while addressing formatting irregularities. Regular expressions were used to handle issues such as extraneous spaces, newline characters (\n), line breaks (\\n), and non-breaking spaces (\xa0), ensuring a clean and standardized text format. This approach maintained the original structure and semantics, making the data suitable for alignment and analysis. Figure 2 shows a general example of text preprocessing process.

Parallel sentences generation

Following the successful pairing of text documents, the next pivotal step in our methodology involves the generation of parallel sentences pairs in English and Arabic. For this task, we employed the Generative Pre-trained Transformer (GPT) model, developed by OpenAI. GPT leverages transformer-based architecture and serves as a prime example of a LLM. Its versions GPT-3 and GPT-4 are some of the most well-known and powerful LLMs currently available. This model is trained on a diverse corpus of text data and excels in recognizing and generating sentence boundaries, not merely based on punctuation but through a deep contextual understanding of language syntax and structure. GPT’s sophisticated capabilities allow it to handle nuanced expressions and complex grammatical constructions effectively, surpassing traditional methods in segmenting and aligning sentences in complex texts (Wu et al., 2023).

To facilitate sentence alignment using this LLM through GPT, we implemented a multi-step methodology to address some limitations of the GPT model. Initially, we utilized a Sentence Embedding model to segment the texts into smaller units based on semantic similarity and a predefined token count limit for English.

The segmentation process began by splitting the English text into smaller units, each limited to approximately 1,000 tokens. This process employed sentence-level tokenization to segment the text into coherent, semantically meaningful units. Sentence boundaries were identified using basic delimiters, such as punctuation and whitespace, ensuring sentence structure was preserved while adhering to the token limit.

For Arabic, corresponding segments were identified dynamically using multilingual embeddings to ensure semantic equivalence with the English segments. Specifically, the cosine similarity of sentence embeddings was calculated to match the last sentence in each English segment with the most semantically similar Arabic sentences. This dynamic alignment determined the Arabic segment boundaries, ensuring contextual and semantic alignment with the English segments.

Although the Arabic segments were not explicitly constrained by a fixed token limit, the semantic alignment process ensured they were approximately equivalent in length to the English segments in terms of tokens. The underlying Sentence Embedding model internally employs subword-level tokenization, enabling accurate representation of linguistic nuances in both English and Arabic, including morphological variations and structural differences.

By integrating sentence-level segmentation for English and semantic similarity-based alignment for Arabic, this approach ensures that the English and Arabic segments are balanced in terms of tokens and content, making them suitable for subsequent parallel sentence generation.

For sentence alignments, each segment pair is processed by GPT to split the segments into aligned sentences. The following prompt was used to instruct GPT:

• Prompt: “Segment each English text and Arabic text, then respond by aligning each sentence in the English text with its corresponding Arabic text. Format each pair as ‘English: (English Text) Arabic: (Arabic Text)’. Separate each sentence pair with a newline.”

The responses from GPT are long texts where each English text segment is introduced with the keyword “English:” followed by its corresponding Arabic translation, which starts with the keyword “Arabic:”.

Despite GPT’s advanced capabilities, it may occasionally encounter technical and formatting issues, necessitating a thorough review and correction process to ensure accuracy. In this study, we assess the quality of GPT’s responses and make necessary adjustments to align with our quality criteria. These criteria include ensuring there are no null values, and that all data adheres to the correct format. Responses that meet these standards are archived for future use, while discrepancies are carefully documented for in-depth analysis.

The detailed procedure for generating parallel sentences is outlined in the pseudocode illustrated in Fig. 3.

Experiment and results

In this section, we detail the experimental setup and procedures used to test and validate the parallel sentence generation framework previously described. This experiment was specifically designed to evaluate the effectiveness of our methodology in generating high-quality English-Arabic sentence pairs with a particular focus on the health text genre within the SauLTC.

Following the completion of the data collection, comprehension, selection, preparation, and document pair generation phases, as elaborated in the preceding section, a total of 85 document pairs were successfully created and prepared for use in the parallel sentence generation phase. These document pairs were subsequently segmented into smaller units to facilitate processing by GPT. The segmentation process involved splitting the English text into smaller units, each limited to approximately 1,000 tokens, using sentence-level tokenization. Arabic segments were dynamically aligned with their corresponding English segments by leveraging semantic similarity scores calculated using the paraphrase-multilingual-mpnet-base-v2 embedding model. This ensured contextual and semantic equivalence between the segments, even though Arabic segments were not constrained by a fixed token limit. The technical details of this process are outlined in Table 4.

The segmented pairs were processed by GPT, configured as described in Table 4, to generate aligned English-Arabic sentence pairs. The experiment utilized the GPT-3.5 Turbo model, chosen for its balance between accessibility, cost-effectiveness, and suitability for the task. The process used a structured prompt to align segments, producing responses with clear formatting. GPT outputs were designed to pair each English segment with its corresponding Arabic translation, marked with the keywords “English:” and “Arabic:” for clarity.

Upon generating the sentence pairs, each response underwent a thorough evaluation to verify its adherence to the quality and formatting standards established for this study. Responses with mismatched English and Arabic counts, improper formatting, or alignment errors were flagged for reprocessing. These flagged outputs were either corrected manually or excluded if they failed to meet the quality criteria upon re-evaluation. During the evaluation process, several issues were identified in the alignment of English-Arabic parallel sentences; examples are presented in Table 5.

These examples illustrate the challenges in ensuring accurate alignment of parallel sentences and demonstrate the importance of our review and correction processes.

Of the 469 initial responses generated by GPT model, 443 were accepted as meeting the high standards set for inclusion in the study. The remaining 26 responses did not meet the necessary criteria and were therefore discarded. This process of validation ensures that our data is both reliable and robust, suitable for further analysis and application in translation studies. Ultimately, the results of the sentence alignment processes are organized into a structured table, formatted with two columns—one for English and one for Arabic, as depicted in Fig. 4.

As a result from the sentence generation phase, we generate a total of 15,845 English-Arabic sentence pairs. The aligned sentence pairs were then evaluated through a dual approach—quantitative analysis and expert human review. Quantitative assessment involved the use of four multilingual sentence embedding models to convert the sentences into vectors for semantic alignment analysis. The models selected for this experiment included:

• “distiluse-base-multilingual-cased-v2”: based on the Universal Sentence Encoder (mUSE).

• “paraphrase-multilingual-MiniLM-L12-v2”: a smaller and more efficient version of the MiniLM architecture.

• “paraphrase-multilingual-mpnet-base-v2”: based on the MPNet.

• “LaBSE”: The Language-agnostic BERT Sentence Embedding (LaBSE).

Following the transformation of sentences into vector embeddings using the selected multilingual models, the semantic alignment between English and Arabic sentence pairs was assessed using cosine similarity, a metric that quantifies semantic similarity based on vector relationships. The average cosine similarity scores and SD for each model are presented in Table 6, reflecting their alignment performance and variability.

Among the evaluated models, Mpnet achieved the highest average similarity score 0.863, indicating effective semantic alignment. LaBSE followed closely with a score of 0.852 and demonstrated greater consistency, as evidenced by its lower standard deviation (SD = 0.09). These results suggest that while Mpnet provides strong overall alignment, LaBSE offers more consistent performance across diverse sentence pairs.

A cosine similarity threshold of 0.7 was used to categorize sentence pairs:

• Aligned pairs: Scores ≥ 0.7 were considered semantically aligned.

• Misaligned pairs: Scores < 0.7 indicated misalignments.

Figure 5 illustrates the distribution of cosine similarity scores, with the threshold prominently marked.

At the 0.7 threshold, LaBSE aligned the highest percentage of sentence pairs (95%), followed by Mpnet (94%), Distiluse (91%), and MiniLM (88%). While these percentages provide a broad overview, examining specific examples shows more about each model’s performance.

For instance, the English sentence “Instead, your body will try to keep you alive by slowing down its metabolism” and its Arabic equivalent “بدلاً من ذلك، سيحاول جسمك الحفاظ على حياتك من خلال إبطاء عملية الأيض” were scored differently across models. LaBSE achieved the highest score (0.919), closely followed by Mpnet (0.841), demonstrating their ability to align nuanced sentence pairs effectively. In contrast, Distiluse (0.656) and MiniLM (0.646) struggled to capture the full semantic equivalence, scoring below the 0.7 threshold.

This example aligns with the overall percentages, indicating that LaBSE and Mpnet are more suitable for tasks requiring precise semantic alignment.

After calculating the cosine similarity scores, a comprehensive human evaluation was conducted to further validate the quality of the aligned sentence pairs. A subset of 100 sentence pairs was evaluated by human experts based on predefined criteria outlined in the Methodology section. The human evaluators assigned an average MQM quality score of 85%, indicating a high level of satisfaction with the results. To ensure the consistency and reliability of human judgments, Gwet’s AC1 Coefficient was computed as a measure of inter-rater reliability, yielding a score of 0.98, which indicates almost perfect agreement.

A key observation was that sentence pairs with high cosine similarity scores (>0.70) generally aligned well with human evaluations, confirming the reliability of cosine similarity as an initial filtering metric. However, cases involving idiomatic expressions or culturally adapted translations sometimes received lower similarity scores due to differences in surface-level wording. For example, the English phrase “We want what we are told we can’t have” was translated as “فكل ممنوع مرغوبا”, receiving a similarity score of less than 0.5 across all models. Despite the low score, human evaluators recognized it as a valid translation, illustrating the limitations of cosine similarity in capturing deeper semantic relationships.

Overall, our study confirms that cosine similarity serves as a useful complementary tool in translation evaluation, particularly for identifying lexical alignment. However, human judgment remains essential for assessing meaning in cases involving idiomatic expressions or significant paraphrasing.

Discussion

This study introduces a novel approach to parallel sentence alignment by integrating GPT, advanced sentence embedding models and human expert reviews into a single, adaptive framework.

Traditional length-based and lexical-based methods rely on fixed heuristics or bilingual dictionaries, which struggle with paraphrased translations and structurally divergent sentences. Similarly, neural models improve alignment by mapping sentences into shared vector spaces, but they typically require large parallel datasets for training, making them less suitable for low-resource languages. Many of these models also struggle with out-of-vocabulary words, affecting alignment accuracy when dealing with rare or unseen terms.

Unlike previous models, our approach reduces the need for predefined segmentation rules, as GPT utilizes contextual understanding to assist in identifying sentence boundaries. Our method effectively handles OOV words, paraphrased sentences, and structural variations without requiring extensive corpus-specific tuning. Additionally, our framework remains adaptable and scalable across different languages and text genres, overcoming the dataset limitations of earlier models.

Our approach combines automated alignment with expert validation, ensuring both precision and adaptability, particularly LTC. By addressing key challenges in previous methods, it offers a more flexible and efficient solution for parallel sentence generation.

We identified several challenges that impacted the accuracy of the alignment process. One significant issue was inconsistency in sentence length and structure between English and Arabic. This inconsistency is primarily due to some translations being shortened, generally translated, or not matching the source text. For example, the English word “Yes” is translated into Arabic as “نعم، بالتأكيد,” which is longer because the Arabic version adds emphasis and certainty with “بالتأكيد” (certainly). This problem is more commonly observed in translating English words that are not lexicalized in Arabic. For example, the English sentence “Metabolism is a word frequently used and seldom understood by dieters” is translated into Arabic as “الأيض هي كلمة كثيرًا ما تستخدم، ونادرًا ما يفهمها الأشخاص الذين يتبعون النظام الغذائي.” The Arabic translation explicitly expands “dieters” to “الأشخاص الذين يتبعون النظام الغذائي” (people who follow a diet), because Arabic does not have one lexical item to express the meaning of the English term “dieters”. In general, these differences arise due to variations in grammar, cultural preferences, and contextual expression.

Another specific challenge arose from instances where students added acknowledgments or extra information at the beginning of the Arabic texts, which were not present in the English versions. This inclusion created additional content in the Arabic translations that had no corresponding segments in the English texts, complicating the creation of accurate translation pairs.

Furthermore, biases inherent in GPT influenced the alignment process and introduced several challenges that required manual interventions to maintain alignment quality and consistency. These challenges included issues such as incomplete sentence alignments, duplication, language switching within sentences, repeated Arabic text, and formatting errors.

Collectively, these issues highlighted the need for optimizing alignment algorithms and incorporating manual interventions to ensure the accuracy and reliability of the generated sentence pairs. To address these challenges, manual interventions were employed to resolve cases where automated methods struggled, such as identifying, correcting, or removing misaligned pairs caused by cultural or linguistic variations. Additionally, flagged outputs with formatting errors or alignment issues were meticulously reviewed and corrected to ensure consistency and quality. These combined efforts not only improved the consistency and accuracy of alignments but also ensured that the generated pairs adhered more closely to contextual and semantic equivalence. The success of these interventions was validated by the high MQM quality scores obtained during human evaluations.

To strengthen the justification for using GPT, we conducted a comparative analysis of the embedding models employed in this study—Mpnet, LaBSE, Distiluse, and MiniLM—against existing sentence alignment methodologies. In comparison, Chimoto & Bassett (2022) used fine-tuned LaBSE embeddings to align sentences in their dataset, achieving 53.3% alignment accuracy and over 85% precision at the same similarity threshold. While our study shows that pre-trained models like Mpnet and LaBSE perform well without fine-tuning, it is important to acknowledge the differences in datasets. Their study focused on Luhya-English, which introduces unique alignment challenges, whereas our dataset of English-Arabic pairs benefits from broader representation in pre-trained models. This comparison highlights the adaptability of embedding models in different contexts and demonstrates the potential of combining them with GPT for generating accurate parallel sentences in resource-rich domains.

Conclusions

This study presents a comprehensive framework for leveraging the SauLTC and LLMs to generate high-quality parallel English-Arabic sentence datasets, addressing a critical gap in TE. By integrating advanced AI tools, such as GPT, and embedding models, the proposed methodology enhances the efficiency and accuracy of parallel sentence generation. The results highlight the effectiveness of this approach, with LaBSE achieving the highest alignment consistency. These findings suggest that AI can play a valuable role in supporting translation pedagogy and practice.

The results demonstrated the framework’s ability to generate 15,845 sentence pairs, with LaBSE exhibiting the highest consistency and aligning 95% of the sentence pairs at the 0.7 threshold. Human evaluation further validated the dataset’s quality with an MQM score of 98%, emphasizing its contextual accuracy and reliability in addressing alignment challenges.

Future directions will focus on expanding the framework to encompass the entire SauLTC, incorporating additional annotations and metadata to enhance the dataset’s applicability across diverse domains. Furthermore, exploring advanced alignment techniques such as retrieval-augmented generation (RAG) can reinforce the contextual relevance and semantic accuracy of parallel sentence generation. Since automatic alignments and human evaluations present some challenges due to variations in linguistic interpretation and contextual nuances, future improvements should focus on fine-tuning embedding models to better capture language-specific variations. Additionally, developing context-aware scoring systems that integrate machine learning classifiers with human annotations will enhance AI evaluation metrics, ensuring more accurate and reliable assessments. By continuing to refine and expand this work, translation educators and researchers can equip students with the necessary skills to navigate the evolving landscape of AI-driven translation services, ultimately enriching the global exchange of knowledge and cultural understanding.

Supplemental Information