Mixtec–Spanish Parallel Text Dataset for Language Technology Development

Abstract

This article introduces a freely available Spanish–Mixtec parallel corpus designed to foster natural language processing (NLP) development for an indigenous language that remains digitally low-resourced. The dataset, comprising 14,587 sentence pairs, covers Mixtec variants from Guerrero (Tlacoachistlahuaca, Northern Guerrero, and Xochapa) and Oaxaca (Western Coast, Southern Lowland, Santa María Yosoyúa, Central, Lower Cañada, Western Central, San Antonio Huitepec, Upper Western, and Southwestern Central). Texts are classified into four main domains as follows: education, law, health, and religion. To compile these data, we conducted a two-phase collection process as follows: first, an online search of government portals, religious organizations, and Mixtec language blogs; and second, an on-site retrieval of physical texts from the library of the Autonomous University of Querétaro. Scanning and optical character recognition were then performed to digitize physical materials, followed by manual correction to fix character misreadings and remove duplicates or irrelevant segments. We conducted a preliminary evaluation of the collected data to validate its usability in automatic translation systems. From Spanish to Mixtec, a fine-tuned GPT-4o-mini model yielded a BLEU score of 0.22 and a TER score of 122.86, while two fine-tuned open source models mBART-50 and M2M-100 yielded BLEU scores of 4.2 and 2.63 and TER scores of 98.99 and 104.87, respectively. All code demonstrating data usage, along with the final corpus itself, is publicly accessible via GitHub and Figshare. We anticipate that this resource will enable further research into machine translation, speech recognition, and other NLP applications while contributing to the broader goal of preserving and revitalizing the Mixtec language.

Dataset:https://doi.org/10.6084/m9.figshare.28557182.v1.

Dataset License: CC-BY-4.0

1. Summary

This work presents the first publicly available preprocessed Spanish–Mixtec parallel corpus in Mexico, designed to facilitate the development of language technologies for this historically low-resource language [1,2]. The dataset, currently comprising 14,587 sentence pairs, was collected from digital sources such as official government websites, religious organization websites, Mixtec enthusiast blogs, and physical texts obtained by scanning printed books at the Autonomous University of Querétaro (UAQ). By including texts from diverse domains, such as education, law, health, and religion, and drawing on 13 Mixtec variants spread throughout Guerrero and Oaxaca, this corpus provides a rich and representative linguistic resource.

Mixtec is regarded as a low-resource language in a digital context, lacking large-scale documentation and linguistic tools [3,4,5]. According to the National Institute of Indigenous Languages (INALI), Mixtec belongs to the Oto-Manguean family, which includes Amuzgo, Triqui, Zapotec, and others [6]. It is one of the most spoken indigenous languages in Mexico, with 526,593 speakers [7]. This situation has motivated the development of technologies aimed at reducing language barriers across various domains. In education, projects such as Aprendamos Mixteco [8] have used gamified digital resources to teach Mixtec to migrant communities in the U.S., while mobile applications for teaching Nahuatl with voice recognition have proven effective in both urban and rural contexts [9]. In another context, automatic translators between Spanish and indigenous languages (such as Nahuatl and Mixtec) offer new possibilities to improve initial interpretation in legal proceedings where speakers do not understand Spanish [10]. Recent projects have also explored the use of multilingual models and synthetic data to translate medical instructions into Mixtec [11]. However, it faces significant challenges in computational processing due to fragmented documentation and regional variants.

In response to these challenges, this dataset underwent an extensive methodology that included the systematic search and collection of bilingual texts, OCR-based digitization, and manual and semi-automatic correction to normalize special Mixtec glyphs [12,13,14,15], along with data cleaning to remove duplicates, empty lines, and nonparallel fragments. The final corpus adheres to UTF-8 encoding, ensuring broad usability across multiple NLP pipelines.

This project provides the basis for future linguistic research and technology development in Mexico. The resource has immediate applications in machine translation (including multimodal translation [16]), optical character recognition [17,18], information retrieval [19], and voice recognition [20,21]. In addition, it can serve as an essential step towards the preservation of the Mixtec language and culture [22,23] by supporting educational materials, dictionaries, and other language revitalization tools. Preliminary experiments with GPT-based models, as well as open source frameworks such as mBART-50 and M2M-100, have shown promising but still limited translation quality, highlighting the need for further research and larger corpora.

By publicly releasing and describing the dataset, our aim is to empower researchers, developers, and community members to collaborate and advance language technologies for Mixtec, ultimately contributing to both linguistic innovation and cultural preservation. The usage instructions are available on the following GitHub repository: https://github.com/hermilocap/ParalellCorpusMixtec-Spanish (accessed on 18 June 2025).

2. Background

A language is considered digitally low-resourced when it lacks sufficient documentation, either in digital format or otherwise, to support the development of language technologies [3,4,5]. In many cases, these languages are also classified as endangered due to a declining number of speakers and limited resources (e.g., dictionaries, textbooks) available either online or in print. Consequently, both the scarcity of digital data and the small population of native speakers hamper the creation of natural language processing (NLP) tools such as tokenizers, morphological analyzers, and spell checkers.

Based on our preliminary investigations, the Mixtec language (which belongs to the Oto-Manguean family) can be classified as a low-resource digital language. Despite the existence of written materials in various dialects, these materials are scattered across physical libraries, such as those of local universities and social organizations, and they are not readily accessible in digital form. Unlike Nahuatl, which has 1,651,958 speakers [24] and historical corpora (38 Nahuatl–Spanish texts, multiple dictionaries, and extensive literature [25]), Mixtec lacks digital resources, and its 81 varieties [26] hinder the development of language technologies. While languages such as Quechua and Guarani have made significant advances due to larger corpora and a degree of orthographic standardization [27], Mixtec has required alternative strategies, including the generation of synthetic data, the manual construction of parallel corpora, and the use of neural network-based translation, which can achieve high accuracy even with limited corpora in controlled contexts [28]. This lack of digital documentation severely restricts the development of core NLP tasks and language applications.

At present, there is no automatic tokenizer or morphosyntactic parser available for Mixtec. The absence of essential NLP tools, compounded by the shortage of domain-diverse corpora, makes it challenging to build robust language technologies. Moreover, the fragmentation of data across multiple dialects, each with its orthographic conventions, adds further complexity to any standardization efforts.

In light of these challenges, our primary motivation is to contribute to the preservation and promotion of the Mixtec language by compiling a parallel corpus that encompasses various domains (e.g., health, legal, and educational) and dialectal variants. This resource can provide the foundation for critical NLP technologies, including text classifiers for language identification, speech recognizers, and multimodal translation systems. These tools not only serve researchers and developers, but they can also address urgent community needs in domains such as public health, legal services, and education.

This work represents a substantial endeavor to collect, digitize, and manually correct parallel texts in Mixtec. Notably, the process required scanning and converting physical materials into editable digital formats, followed by meticulous verification and the alignment of Mixtec–Spanish sentences. The result is a novel resource that can be used for academic research, educational initiatives, and the broader development of language technologies intended to improve the vitality of Mixtec.

3. Data Description

According to INALI, Mexico is home to 68 linguistic groups, 364 variants, and 11 linguistic families [6]. Among these families are Oto-Manguean, Yuto-Nahua, Seri, and Mixe-Zoque. Mixtec belongs to the Oto-Manguean family, sharing various linguistic similarities with languages such as Amuzgo, Triqui, Zapotec, and Tlapanec. The states of Guerrero, Oaxaca, and Puebla have the largest number of Mixtec speakers. Official estimates suggest that there are more than 81 distinct Mixtec variants in Mexico, although this number may be higher if small communities and municipalities with their own dialectal differences are considered.

Based on data from the National Institute of Statistics and Geography (INEGI), the Mixtec language is spoken by 526,593 individuals in Mexico, making it the fourth most spoken indigenous language in the country [7]. Despite the significant number of speakers, the resources necessary for computational analysis remain fragmented or limited to physical formats, thereby hindering the creation of robust language technologies.

In this section, we describe the data generation process, dataset statistics, and final data organization.

3.1. Dataset Generation Process

The following steps were followed to generate the parallel corpus.

• We visited university libraries and official government websites and had direct contact with Mixtec advocates who provided support on data sources.
• We selected texts that had translations in Mixtec and Spanish and were in the public domain, discarding those that did not meet these criteria.
• We collected texts from various categories whenever possible, including educational, health, legal, and religious texts.
• Each Mixtec–Spanish pair was rigorously reviewed to ensure the correct alignment.
• We corrected spelling errors and removed empty and duplicate sentences.

In generating the dataset, we identified that the parallel corpus was translated by native Mixtec speakers from Mixtec communities who were invited by educational institutions, government agencies, and private companies to collaborate in the translation of the texts. The texts generated by these institutions are intended to communicate and inform the Mixtec community, provide healthcare recommendations, and offer educational resources in the Mixtec language.

3.2. Mixtec Variants and Domains

By analyzing the collected texts, we identified the following 13 Mixtec variants: Western Coast, Southern Lowland, Santa María Yosoyúa, Central, Lower Cañada, Western Central, San Antonio Huitepec, Upper Western, Southwestern Central, Tlacoachistlahuaca, Northern Guerrero, and Xochapa. Figure 1 shows the total number of documents per variant. In particular, the Santa María Yosoyúa variant from the state of Oaxaca is represented by 12 documents, while two documents represent Northern Guerrero.

Depending on the type of document, we organize the corpus into the following four main domains: education, law, health, and religion. Figure 2 illustrates the categories and their proportion from the total documents. The largest category is educational texts (80%), followed by legal texts (14%), health-related texts (3%), and religious texts (3%). Health materials primarily include patient care guidelines and patient rights documentation.

Although this distribution is unavoidably skewed, reflecting the real-world availability of written Mixtec, machine translation models can readily compensate for such bias through lightweight strategies such as temperature-based sampling or dynamic data weighting during fine-tuning [29].

3.3. Corpus Statistics

Table 1. Number of sentences and words in each corpus category. Religious texts largely derive from the Bible, while legal texts include the Mexican Constitution and official documents pertaining to indigenous language rights.

Category	Number of Mixtec–Spanish Sentences	Total Words (Mixtec–Spanish)	Unique Words (Mixtec–Spanish)
Educational	9220	42,654	6971
Laws	3498	180,236	4634
Health	22	278	70
Religious	16,434	461,234	5035

provides an overview of the corpus at the sentence and word level by category. The religious category (primarily consisting of Bible passages) contains the highest number of sentences, 16,434, while health-related texts comprise only 22 sentences, limiting the representativeness of the parallel corpus in this domain.

3.4. Organization of the Repository

All parallel texts are available through the project repository. The main folder, data, contains two subfolders as follows: train and test. The train folder includes mixtec-train.txt and spanish-train.txt, while the test folder contains mixtec-test.txt and spanish-test.txt. Each file corresponds line by line to its parallel counterpart as follows:

• mixtec-train.txt: Used for training; each line in Mixtec aligns with the corresponding line in spanish-train.txt.
• spanish-train.txt: Contains the Spanish version of each sentence, matching each line from mixtec-train.txt.
• mixtec-test.txt: Used for testing; each Mixtec sentence aligns with the corresponding line in Spanish.
• spanish-test.txt: Contains the Spanish version of each sentence in mixtec-test.txt.

Formally, if a Mixtec file x contains sentences $x_i$ for $i=1\dots n$, the corresponding Spanish file y contains sentences $y_i$ aligned line by line with $x_i$.

4. Methods

This section details the methodology used to create the dataset. We followed a three-phase approach for building the Spanish–Mixtec parallel corpus. This same approach can be extended to other digitally low-resourced languages to create robust and relevant parallel corpora that can support various language technologies. Figure 3 illustrates the methodology, which includes the following phases: (1) systematic search for parallel texts, (2) scanning of physical materials and semi-automatic correction, and (3) preprocessing including deduplication, removing empty fragments, and manual classification of texts [12,13,14,15].

In the remainder of this section, we detail each phase.

4.1. Phase 1: Systematic Search for Parallel Texts

The first phase involves searching for parallel materials (Spanish–Mixtec), carried out in two main steps as follows: an online exploration and in-person visits to the libraries of the Autonomous University of Querétaro (UAQ). The goal is to discover and collect bilingual texts suitable for constructing the corpus. We apply inclusion and exclusion criteria to ensure that only relevant and high-quality documents are retained.

In Step 1, we began by identifying potential online portals that could host parallel texts, focusing on government resources such as the government website of Mexico, the INALI website, and the National Institute of Indigenous Peoples (INPI), as well as Mixtec language blogs and websites maintained by the Summer Institute of Linguistics (ILV). After visiting each platform, we downloaded texts that explicitly contained both Mixtec and Spanish content. This step yielded 22 digital books that contained parallel segments.

In Step 2, we then conducted a targeted search of the UAQ’s virtual library catalog to identify additional bilingual materials. Following this digital exploration, we visited the libraries of the Faculty of Philosophy and the Faculty of Social Sciences in person. Through these visits, we obtained six additional physical books that contained parallel texts.

4.1.1. Inclusion Criteria

During the selection of relevant resources, the following inclusion criteria were used:

• The title of the book or document explicitly indicates that it includes both Mixtec and Spanish.
• There is a clear correspondence between a segment in Mixtec and its Spanish translation, at least at the paragraph or sentence level.
• For texts with long paragraphs, we confirmed that extensive sections in Mixtec are accurately matched with their Spanish equivalents.

4.1.2. Exclusion Criteria

The following factors led to the exclusion of certain materials as follows:

• Texts in which the Mixtec segment was disproportionately short compared to a very long Spanish passage (or vice versa). Long texts were on the order of 7000 words.
• Multilingual materials containing Mixtec, Spanish, and English, where the English component was not essential or only partially parallel.
• Poetry, songs, or lyrical verses, even if they contained some parallel segments, were omitted to maintain consistency in prose-based alignments.

4.2. Phase 2: Scanning of Physical Parallel Corpora

In this phase, we focus on digitizing the physical materials and ensuring their accuracy through a semi-automatic correction process.

• Digitization. Physical books from the UAQ libraries were manually scanned. Each page was scanned using a flatbed scanner, resulting in a digital file in either PDF or JPEG format. A total of eight books were digitized. This process is especially significant given the scarcity of digital resources for Mixtec, as it opens these texts to broader usage by researchers and computational linguists.
• Semi-automatic correction. To convert the scanned pages into machine-readable text, we used standard optical character recognition (OCR) software (e.g., ABBYY FineReader Engine 11). Since Mixtec is not natively supported by most commercial OCR tools, we used a Spanish-based recognition model and then customized it with a new Mixtec character set. Although ABBYY FineReader supports various high-resource languages, our review did not find direct support for Mixtec. Therefore, we constructed a special dictionary that contains relevant Mixtec glyphs. We ran the recognition process, reviewed the output, and corrected misread characters caused by faded printing and typed text. This iterative procedure was crucial in generating reliable digital text. Mixtec glyphs are those characters that OCR has difficulty recognizing [28]: $\underset{¯}{\mathrm{A}\underset{¯}{,}}$ $\underset{¯}{\mathrm{a}\underset{¯}{,}}$ $\underset{¯}{\mathrm{E}\underset{¯}{,}}$ $\underset{¯}{\mathrm{e}\underset{¯}{,}}$ $\underset{¯}{\mathrm{I}\underset{¯}{,}}$ $\underset{¯}{\mathrm{i}\underset{¯}{,}}$ $\underset{¯}{\mathrm{O}\underset{¯}{,}}$ $\underset{¯}{\mathrm{o}\underset{¯}{,}}$ $\underset{¯}{\mathrm{U}\underset{¯}{,}}$ $\underset{¯}{\mathrm{u}\underset{¯}{,}}$ Ā, ā, Ē, ē, Ī, ī, Ō, ō, Ū, ū. The incorrect words appear like x§én?, TahvT=tl, and caj o, while the correct forms are xēēn, Táhvi-ti, and caj-o. The blurred printouts that the OCR was unable to correct automatically were resolved by manually completing the characters.
  To validate the linguistic quality of the corpus, the texts were manually reviewed by a native Mixtec expert. After correction with ABBYY FineReader, the texts were reviewed to identify and correct spelling errors. A manual review of Mixtec grammar was also performed throughout the corpus.

4.3. Phase 3: Preprocessing of the Parallel Corpus

Preprocessing is indispensable in any NLP pipeline, with the goal of producing a clean, well-structured corpus for subsequent tasks, such as machine translation or language modeling. We followed the six steps described below.

• Removing duplicate sentences. We identified and deleted duplicate segments that appeared multiple times within our corpus. These duplicates often arise from repeated passages in source documents or OCR artifacts. As duplicates can bias statistical models or neural networks, we removed them using a regular expression-based script that searched for repeated lines and deleted them in both the Mixtec and Spanish versions simultaneously. Sentences such as “Te tu cuni ñuhün-un cahmu yunu xíi”, “Furthermore, with your spirit”, “—Vaha —cáchí da”, “T. Amen” were removed. Figure 4 shows our duplicate removal workflow.
• Removing empty fragments. We found fragments containing only numbers or whitespace, plus non-textual metadata such as editorial information and printing addresses. We removed these lines from both language files, since they do not contribute meaningful linguistic content. Words such as “nan”, “- 21 -”, “- 3k -”; section titles from documents such as “Section 1: who you are”, “Section 2: I LIVE IN THIS TOWN”; and section descriptions such as “CURRENT PRESENT. IDENTIFYING PHRASE. RELATIONSHIP” and “WAY OF EXPRESSING THE SEARCH” were removed. Algorithm 1 outlines the process of detecting and deleting empty fragments.

  Algorithm 1 Removing empty fragments or non-linguistic lines from the Mixtec–Spanish parallel corpus.

  Require: MixtecFile: plain-text file in Mixtec

  SpanishFile: plain-text file in Spanish

  Ensure: Cleaned text files (Mixtec and Spanish) 1: Identify every empty fragment in both files {blank lines or lines with only whitespace} 2: Remove lines that contain no valid characters or only digits 3: Write the cleaned data into new text files

• Removing non-parallel fragments. We also discarded segments where the Mixtec or Spanish part was disproportionately longer, indicating a lack of true parallel alignment.
  Disproportionately long fragments are those where the Mixtec text and its Spanish translation were present, but there was no division into alignment pairs. Sometimes the Mixtec text is shorter than the Spanish text, and there are no correspondences between the Mixtec and Spanish sentences.
• Removing incorrect parallel segments. Some lines included Spanish text in the Mixtec file and vice versa. We manually aligned and corrected them to ensure each line in Mixtec corresponds to the correct line in Spanish.
• Adding missing segments. We conducted a final review to identify incomplete lines, often caused by outdated translations or editorial omissions. In such cases, we referred back to the original source. This step was especially important for the Mexican Constitution’s translations, which had changed over time in Spanish but not always in Mixtec.
• Categorizing texts. We completed the process by classifying the cleaned corpus into the following four major domains: education, law, and religious health. Figure 2 shows the number of documents per category, and Table 1 reports the total number of sentences and words in each category.

4.4. Exploratory Domain Analysis

To verify that the four domains are reflected in the lexical-semantic space of the corpus, we performed a lightweight, unsupervised analysis as follows:

• All Spanish sentences and their Mixtec counterparts were embedded with LaBSE [30], yielding 768-dimensional vectors that are language-agnostic yet semantically aligned.
• We projected the vectors to two dimensions using UMAP (cosine metric, $n_{\mathrm{neighbors}}=15$, $\min \_\mathrm{dist}=0.1$) to visualize the global structure while preserving local neighborhoods.
• We plotted every sentence as a point, colored by its label, as shown in Figure 5 (Spanish) and Figure 6 (Mixtec). Dense, well-separated regions emerge for the Religious and Law domains, whereas Education and Health partially overlap, consistent with the class imbalance in Table 1. A Davies–Bouldin Index calculated in the original 768-D space yields $5.5$ for Spanish and $5.3$ for Mixtec. These high values (lower is better, with values less than one considered well separated) suggest appreciable overlap among the four topical clusters, consistent with the broad, partly interrelated nature of the domains.

This exploratory analysis reveals that lexical cues imprint an identifiable but diffuse domain structure, providing the corpus with a realistic and moderately challenging profile for downstream text classification tasks.

5. Translation Experiments

We used BLEU [31] and TER [32] as metrics in our experiments. BLEU evaluates translation quality from 0 to 1, using modified n-gram accuracy and a brevity penalty to align candidate translations with reference lengths in terms of word choice and order. This penalty applies corpus-wide for sentence flexibility. The reference length r is matched to the candidate sentence lengths, decreasing exponentially $r/c$. c indicates the total translation length of the candidate.

The BLEU metric is calculated by multiplying the geometric mean of the precision of modified n-grams by the brevity penalty, using formulas in Equations (1) and (2). The precision is derived for n-grams of length N with weights $w_n$. Here, c indicates the candidate translation length and r the reference corpus length.

$$BLEU=BP·exp(\sum _{n=1}^N{w}_{n}log{P}_n)$$

(1)

$$BP=\left\{\begin{array}{cc}1& \mathrm{if}\mathrm{c}>\mathrm{r}\\ e^{(1-r/c)}& \mathrm{if}\mathrm{c}\le \mathrm{r}\end{array}\right.$$

(2)

TER, on the other hand, measures the error rate in the prediction and the percentage of necessary corrections in the translation relative to the reference translation. The formula for obtaining the TER metric is shown in Equation (3).

$$TER={\displaystyle \frac{numberofedits}{averagenumberofreferenceword}}$$

(3)

We used 14,587 Mixtec–Spanish sentence pairs divided into 80% training and 20% test sets. The experiments validated the usability of the data for the Spanish–Mixtec and Mixtec–Spanish translation using fine-tuned GPT, mBART-50, and M2M-100 models. The test set served as a reference for evaluation, with predictions reviewed by a native Mixtec expert and assessed using BLEU and TER metrics.

Table 2. Evaluation of the Spanish–Mixtec dataset. Higher BLEU scores in the range of 0–100 indicate closer matches with the reference translation, whereas lower TER scores indicate fewer edits needed to match the reference.

Model	BLEU	TER
GPT-4o (zero-shot)	0.10	143.35
GPT-4o-mini (zero-shot)	0.07	131.32
Fine-tuned GPT-4o-mini	0.22	122.86
Fine-tuned mBART-50	4.2	98.99
Fine-tuned M2M-100	2.63	104.87

and

Table 3. Evaluation of the Mixtec–Spanish dataset. The fine-tuned gpt-4o-mini-2024-07-18 model yields a higher BLEU score than M2M-100, although mBART-50 achieves better overall performance in terms of BLEU.

Model	BLEU	TER
GPT-4o (zero-shot)	0.46	182.28
GPT-4o-mini (zero-shot)	0.14	453.48
Fine-tuned GPT-4o-mini	0.23	121.01
Fine-tuned mBART-50	2.87	111.26
Fine-tuned M2M-100	0.05	112.57

summarize the evaluation in terms of the BLEU and TER metrics.

The results show that mBART-50 achieved the best results with a BLEU score of 4.2 and requires fewer edits (TER = 98.99).

For Mixtec–Spanish translation, mBART-50 also performs best, achieving a BLEU score of 2.87 and TER of 111.26. These experiments were conducted using five epochs and a learning rate of $3\times {10}^{-5}$ on Google Colab with an A100 GPU, monitored using the Weights & Biases (Wandb) platform.

6. Discussion

In this section, we compare the advantages of our dataset with those of related works focused on parallel data collection for digitally low-resource languages.

The primary advantage of our dataset compared to the work of Tonja et al. [1] is its completeness. Every Spanish sentence in our corpus has a corresponding Mixtec sentence, eliminating the need for researchers to review and fill in missing data manually.

Another advantage compared to previous efforts [2,33] is that our corpus is thoroughly preprocessed, publicly available, and clean. We conducted rigorous preprocessing, including the removal of duplicate sentences, and we made the dataset openly accessible through FigShare and the code on GitHub.

Table 4. Comparison of corpus size with related works. A check-mark ✓ indicates that the authors released a dataset that is already pre-processed or manually corrected, whereas a cross × indicates that such resource is not available.

Parallel Corpus	Number of Sentences	Languages	Preprocessed Datasets	Manual Corrected Datasets
Gaustad et al. [34]	110,367	English, Siswati	✓	×
Chiruzzo et al. [35]	14,500	Guarani, Spanish	✓	×
Oliver et al. [36]	173,530	South Tyrolean German, Italian	✓	×
Zhu et al. [37]	20,106	Chinese, Lao	×	✓
Ours	14,587	Mixtec, Spanish	✓	✓

compares the size and processing properties of our corpus with related datasets.

A significant advantage of our dataset compared to other studies is its compilation of a diverse corpus sourced from multiple domains. This includes educational texts that can serve in the development of language technologies applicable to the educational sector, such as educational data mining [38] for identifying patterns that facilitate teaching and learning, Mixtec dialect classification [39], and correction of spelling errors [40].

Furthermore, several studies have shown that techniques such as translation [41], transfer learning [42], and the use of multilingual models [43] enable the development of functional tools even with limited available data. In the field of education, for example, the Lesan system has enabled translation between low-resource languages, facilitating access to educational content in local languages [44]. Mobile applications have also been developed to teach indigenous languages, which have proven useful in communities with limited connectivity [24]. In the legal domain, tools have been designed to help people access legal information through the training of models that summarize court rulings in various languages [43]. Therefore, adapting NLP models not only improves access to educational and legal domains but also extends to the healthcare sector [45].

7. Conclusions

This study outlines a method and dataset for creating a Spanish–Mixtec parallel corpus, addressing the challenges of digitizing bilingual materials for the low-resource Mixtec language. Using systematically online repositories and libraries, we collected texts from educational, legal, health, and religious fields. We digitized physical documents using manual scanning and semi-automatic OCR corrections, customized for Mixtec glyphs not supported by regular software.

The processed corpus, free of duplicates and non-parallel segments, serves as a reliable resource. Automatic translation experiments with the GPT, mBART-50, and M2M-100 models showed their potential to advance NLP technology.

Our work has some limitations. We did not address automatic text alignment between Spanish and Mixtec, and there is a disparity in sentence numbers across categories, with limited sentences in the health category. This affects the corpus’s representativeness for technology development in this domain. We suggest expanding the corpus and working with Mixtec communities for translations. Finally, grammatical analysis of Mixtec was outside the scope of our project, and our corpus focuses only on Mixtec variants from Guerrero and Oaxaca.

This Spanish–Mixtec parallel dataset constitutes an essential resource for future research, facilitating the creation of tokenizers, morphological analyzers, machine translation systems, speech recognition, and other essential language technology tools that can significantly benefit the Mixtec community. We anticipate that publicly releasing this dataset will promote collaborative research efforts and practical applications, contributing meaningfully to the digital preservation and revitalization of the Mixtec language. In the future, we plan to expand the corpus by incorporating additional linguistic variants and various application domains while refining the automatic processing methods specifically adapted to the distinct typological characteristics of Mixtec.