Hey there fellow BSIT student! Super stoked to hear about your capstone project—Cebuano-Tagalog translation is such a practical, impactful idea for local users. Let’s break down how to tackle dataset training step by step, no overwhelming jargon, just actionable steps:

第一步：搞定高质量的平行语料（双语对照数据） #

Since Cebuano and Tagalog are regional languages, you won’t find massive datasets like English-Spanish, but there are solid ways to build your own:

• Leverage local open resources: Check out public datasets from Philippine academic institutions (like UP’s language research departments) which often release parallel texts from news articles, folk tales, or government documents. Open-source NLP communities also have small shared corpora for Philippine languages—keep an eye out for those.
• Build your own parallel pairs: This is where you can make your dataset unique and tailored to real-world use:
  • Grab bilingual news: Local news outlets often publish the same story in both Cebuano and Tagalog—copy-paste the corresponding paragraphs to create pairs.
  • Crowdsource simple sentences: Ask classmates, family members (who are native speakers of either language) to translate daily phrases (e.g., "Where is the nearest canteen?" → Cebuano: "Unsa ang pinakaduol nga kantina?" / Tagalog: "Saan ang pinakamalapit na kantina?") focusing on campus, market, or community scenarios.
  • Validate with native speakers: Always double-check pairs you build to avoid translation errors—bad data leads to bad model performance.
• Filter low-quality data: Ditch duplicate sentences, empty lines, or pairs where the translation doesn’t match the source (e.g., a Cebuano sentence about weather paired with a Tagalog sentence about food).

第二步：预处理数据（让模型能读懂） #

Raw data is messy—clean it up before training:

• Standardize text:
  • Remove special characters (like emojis, random symbols) except basic punctuation (.,!? ).
  • Unify spelling variants (e.g., some Cebuano words have alternate spellings—pick one standard form and stick to it).
  • Convert all text to lowercase to avoid case sensitivity issues.
• Format the data: Save your pairs in a tab-separated values (TSV) or CSV file, with one column for Cebuano and one for Tagalog. For example:

  Unsa imong plano karong weekend?	Anong plano mo this weekend?
  Gusto ko magluto sinigang.	Gusto kong magluto ng sinigang.
  

• Filter sentence lengths: Cut out sentences that are too short (1-2 words) or too long (over 50 words)—these don’t add much value and slow down training. Aim for 5-40 word pairs.

第三步：模型训练（start small, iterate fast） #

You don’t need a supercomputer to train a basic translation model—start with lightweight tools and scale up if you can:

• Choose a beginner-friendly framework: TensorFlow/Keras or PyTorch are great. For translation, start with a simple Seq2Seq (Sequence-to-Sequence) model—there are tons of beginner-friendly tutorials that walk you through building one for small languages.
• Split your data: Divide your dataset into three parts:
  • 80% for training (the model learns from this)
  • 10% for validation (check if the model is overfitting during training)
  • 10% for testing (final evaluation of how well it works)
• Training tips to avoid headaches:
  • Use early stopping: If your validation loss stops improving after 3-5 epochs, stop training—this prevents overfitting (when the model memorizes training data instead of learning to translate).
  • Adjust batch size: Start with a small batch size (16 or 32) if you’re using a laptop—this uses less memory.
  • Fine-tune a pre-trained model (if possible): Some pre-trained models are built for Southeast Asian languages. You can take one of these and fine-tune it on your Cebuano-Tagalog pairs—this gives you a better starting point than training from scratch.
• Evaluate your model: Use the BLEU score (a standard metric for machine translation) to measure how accurate your translations are. Most NLP libraries have built-in functions to calculate this.

第四步：毕设 extra credit tips #

• Focus on niche use cases: Instead of trying to translate everything, optimize for scenarios that matter to Filipinos—like campus announcements, market transactions, or medical advice phrases. This makes your app more useful and easier to showcase.
• Build a simple demo UI: Use tools like Streamlit or Flask to make a basic web interface where users can type a Cebuano sentence and get a Tagalog translation. It’s way more impressive to show a working demo than just code.
• Document everything: Keep track of how you built your dataset, what preprocessing steps you took, and how you adjusted your model. This will make your defense way smoother—professors love seeing your thought process.

Good luck with your capstone! This project is going to be really cool—can’t wait to see how it turns out. If you hit a snag with code or data cleaning, feel free to follow up with specific questions!