Language is a fundamental aspect of human identity, deeply connected to geographical origins, cultural heritage,

and social belonging. However, many indigenous languages across the world are gradually declining due to

providing an accessible platform for both native speakers and learners to understand, translate, and communicate

in their local dialects.

To evaluate the usability and effectiveness of the application, User Acceptance Testing (UAT) was conducted

among selected users. Data were collected through structured interviews, document analysis, and standardized

evaluation tools to ensure comprehensive assessment and validation. Experimental results showed that the

TextConvoNet model achieved the highest accuracy rate of 74.00 percent, surpassing the performance of both

LSTM and CNN-based models. This demonstrates the model’s efficiency in identifying and classifying Blaan

dialects, highlighting its potential in the field of Natural Language Processing (NLP).

Future research should focus on expanding the dataset by collecting transcriptions from diverse age groups,

locations, and communication contexts to improve model generalization and accuracy. Further refinement of the

Language is the foundation of human communication, enabling us to share thoughts, feelings, and more. Each

country around the world has its own unique native languages. However, due to technological evolution, many

languages are now under threat of extinction, which could result in the loss of cultural identity and traditional

knowledge. Approximately 6,000 indigenous languages are recognized, and half of these, specifically 1,500, are

at risk. One way to safeguard cultural heritages and indigenous languages is to integrate this knowledge into

international science-policy-society frameworks, where diversity is recognized and contributes to strengthening

Indigenous Peoples [1].

Currently, at least 182 languages are spoken in the Philippines [2], yet six languages are recognized as extinct

[2][3][4]: Agta Villaviciosa, Agta Dicamay, Katabaga, Ayta Tayabas, Inagta Isarog, and Agta Sorsogon.

Notably, all these languages were once used by various Negrito ethnolinguistic communities. Lobel (2015)

The study preserves indigenous languages and keeps pace with technological advancement by applying Machine

Learning Techniques. It identifies and translates texts from different Blaan dialects using a mobile application

developed by the researchers. Furthermore, the study provides a technological foundation for creating

educational resources, translation tools, and digital archives which fosters a greater appreciation and

understanding of regional linguistic diversity.

In the field of Natural Language Processing (NLP), language identification plays a crucial role in automatically

determining the language used in a given text or document. This process is essential for tasks involving text

analysis, as identifying the language is a prerequisite for further linguistic processing. A study examined the

Machine Learning is a diversified subject of computer science and control science. It is widely used in various

fields in which it learns patterns and relationships of complex data. Machine Learning has also the ability to

adapt and generalize data models enabling it to compute complex processes. Moreover, there are three categories

of Machine Learning, supervised learning, semi-supervised learning, and unsupervised learning [8].

Convolutional Neural Networks (CNNs) are deep learning models effective for recognizing patterns in both

visual and sequential data. Originally designed for image processing, they are now widely used in Natural

Language Processing (NLP) tasks like text classification, sentiment analysis, and Blaan dialect identification.

CNNs capture essential text features, enhancing accuracy in language tasks [9].

A recent advancement, TextConvoNet improves text classification by capturing both intra- and inter-sentence

analysis and question classification. By utilizing pre-trained word embeddings, his model could capture relevant

text representations and improve classification accuracy. Kim's methodology primarily relies on convolutional

layers for feature extraction, utilizing various filter sizes to learn patterns effectively direct from the input text

data. This approach laid the groundwork for later developments in CNN models, influencing subsequent

architectures like TextConvoNet.

Recurrent Neural Networks consist of interconnected nodes which are trained through the process of forward

and backward propagation utilizing “binary cross-entropy loss” and optimized using the “Adam optimizer” to

minimize the difference between predicted and actual labels. Under the variation of Recurrent Neural Network

LSTMs have proven highly effective in different activities, covering areas like understanding human language,

forecasting sequences over time, and recognizing spoken language. It is mainly designed for processing

sequential data of varying lengths [11].

To identify a Blaan dialect, a dataset of text samples from the General Santos City, Polomolok, and Sarangani

Blaan dialects was collected. The data was obtained from transcribed literature and interviews, ensuring a diverse

representation of dialectal variations.

Table I. Blaan Texts by Category

dependencies in text data [9]. While conventional CNN models rely on 1D convolutional filters to extract intra-

sentence features, TextConvoNet extends this by leveraging 2D multi-scale convolutional operations, enabling

feature extraction across sentence boundaries. The model represents textual input as a paragraph-level

embedding matrix, allowing for a richer representation of contextual relationships. Its architecture consists of

multiple convolutional layers with varying kernel sizes, followed by ReLU activations, feature concatenation,

and a classification layer.

To adapt TextConvoNet for multi-class classification, the researchers replaced the binary cross-entropy loss with

sparse categorical cross-entropy and modified the activation function from sigmoid to softmax. This adjustment

allows the model to output probability distributions over three classes rather than a binary decision. The

restructuring ensures effective distinction among the three dialects while leveraging a more efficient loss

In machine learning, class imbalance occurs when certain categories in a dataset have significantly more samples

than others, leading to biased models that favor the majority class [12]. This imbalance can hinder a model’s

ability to accurately classify underrepresented categories, as it may learn to prioritize the dominant class while

overlooking the minority ones. A common approach to address this issue is oversampling, which involves

increasing the number of instances in the minority class by duplicating existing samples or generating synthetic

ones. By balancing the dataset, oversampling improves model performance and ensures that all classes are

adequately represented during training [13].

negatively impact the model’s ability to classify underrepresented dialects accurately. To address this issue and

develop a more effective classification model, they applied oversampling techniques to balance the dataset [12].

Regularization is a technique in machine learning used to prevent overfitting, a common issue where models

perform well on training data but fail to generalize to new, unseen data [14]. Overfitting often occurs when

models become too complex, capturing noise and random fluctuations rather than meaningful patterns [15].

The researchers applied kernel regularization to enhance the model’s ability to generalize across different Blaan

dialects. This approach adds a penalty term to the loss function based on the magnitude of the kernel weights,

discouraging excessively large weights and reducing the risk of overfitting. By incorporating kernel

regularization, they aimed to develop a more robust and reliable classification model [16].

findings were carried out objectively and independently, solely for academic and research purposes.

the imbalance in performance across dialects, while the weighted average of 0.74 suggests that, overall, the

Figure 1. Text Convo Net Confusion Matrix

The results indicate that TextConvoNet effectively classifies Blaan Sarangani texts but struggles with Blaan

Figure 3. Dialect Identification Feature of the Mobile Application

Figure 3 illustrates the dialect identification feature, where users input a Blaan text, and the application predicts

the most likely dialect along with confidence scores. The results are displayed in both text and graphical formats,

providing a clear breakdown of the classification probabilities for Gensan, Polomolok, and Sarangani dialects.

Figure 4. Translation Feature of the Mobile Application

Figure 4 showcases the translation feature, where users can input either a Blaan text or an English sentence. The

application allows users to select the appropriate Blaan dialect before translating. Upon clicking the Translate

button, the system generates the corresponding translation from Blaan to English or from English to Blaan,

displaying the output below the button. This feature enables users to understand Blaan phrases and also construct

The translation feature provides reliable translations by using predefined translations from collected data,

Some challenges, such as limited data for certain dialects and variations in spelling, affected accuracy. However,

The study highlights the importance of using technology to support the preservation and accessibility of the

Blaan language. The app serves as a valuable tool for Blaan speakers, students, and researchers by providing an

easy way to classify dialects and access translations.

Table V. Detailed UAT Result of the Program Design

design and usability. Users find the interface visually appealing with a score of 4.30, easy to navigate with 4.83,

and consistently designed with 4.80. The layout is intuitive, scoring 4.70, while the app performs smoothly with

minimal delays, receiving 4.57. These results confirm that the application is highly functional, user-friendly, and

well-optimized.

Table VI. Detailed UAT Result of the Acceptability