KMITL

Permanent URI for this communityhttps://dspace.kmitl.ac.th/handle/kmitl/1

Browse

Settings

Subject: search.filters.subject.Thai NLP

Search Results

Now showing 1 - 3 of 3
  • Loading...
    Thumbnail Image
    Some of the metrics are blocked by your 
    consent settings
    Item type:Publication,
    Virtue-Based Thai Folktale Recommendation with Ensemble LLMs
    (2026-06-16)
    Daeng-Am, Wassana
    ;
    Anuntachai, Anuntapat
    ;
    Netisopakul, Ponrudee
    Moral learning plays an important role in early childhood education, yet teachers often spend considerable time identifying moral lessons in stories and deciding which virtues they represent. This study investigates whether a multi-model ensemble strategy can produce more teacher-aligned moral extractions from Thai folktales than a single-LLM baseline. We propose an LLM-based ensemble framework that extracts concise moral statements and classifies them into eight core virtues promoted by the Thai Ministry of Education: diligence, frugality, honesty, discipline, politeness, cleanliness, unity, and kindness. The framework combines outputs from Google Gemini 2.0 Flash, OpenAI GPT-4o-mini, and Anthropic Claude 3.5 Sonnet through a semantic consensus mechanism using BGE-M3 embeddings and majority voting. The system is evaluated on a corpus of 200 Thai folktales annotated by three experienced early childhood educators using multi-label metrics including Hamming Loss, Jaccard Similarity, and Exact Match. The results show that the ensemble approach achieves a Hamming Loss of 0.208, Jaccard Similarity of 0.564, and Exact Match of 0.175, consistently outperforming all single-model baselines. These findings suggest that consensus-driven ensemble inference provides a more robust and teacher-aligned foundation for automated moral education tools in Thai NLP.
      1
  • Loading...
    Thumbnail Image
    Some of the metrics are blocked by your 
    consent settings
    Item type:Publication,
    MANJU: A Multi-Agent Framework for Natural Just-in-Time Understanding in Thai
    (2026-01-01)
    Koseeyaumporn, Sawit
    ;
    Saiprom, Siratee
    ;
    Promfiy, Watin
    ;
    Tarnpradab, Sansiri
    This paper presents MANJU (Multi-Agent AI for Natural Just-in-Time Understanding), an automated platform for Thai-language customer service delivered through just-in-time voice interaction. MANJU transcribes a customer's spoken query, applies multi-agent reasoning with retrieval-augmented generation (RAG) to determine the optimal response strategy, and synthesises a reply in natural Thai, delivering the first audio segment within approximately 1.8 s via sentence-level streaming. A visual drag-and-drop work flow editor enables non-technical operators to compose and modify conversation logic without programming.Evaluated on 50 standard Thai customer-service queries and 23 adversarial edge cases, MANJU achieves 100% intent routing accuracy, 62% RAG-grounded responses (GR), 78% edge-case acceptability with 100% safety, and Mean Opinion Score (MOS) = 4.33 for VoiceDesign TTS. End-to-end median latency is 6.07 s; the primary bottleneck is Qwen3-TTS at Real-Time Factor (RTF) = 1.54, partially offset by sentence-level streaming.
      1
  • Loading...
    Thumbnail Image
    Some of the metrics are blocked by your 
    consent settings
    Item type:Publication,
    ThaiScamBench: Toward a Benchmark Dataset for Scam and Phishing Detection in the Thai Language
    (2025-01-01)
    Chalaemwongwan, Nutthakorn
    Recent reports from Thai financial regulators reveal a sharp increase in online scams, resulting in major financial damage each year. While English-language research has made considerable progress, Thai-language scam detection remains underexplored and lacks standardized benchmarks. The study establishes *ThaiScamBench (pilot)*, a curated corpus of 1,750 Thai messages labeled as scam or legitimate across seven categories. Reproducible baselines (Logistic Regression, Linear SVM) and a rigorous evaluation protocol are provided to address class imbalance, Thai-English code-switching, and adversarial obfuscation. All scam texts contained URLs; none appeared in legitimate messages, revealing dataset bias, with all scam samples containing a URL and none among legitimate texts, artificially inflating headline metrics. We addressed this issue by introducing URL-masked and domain hold-out evaluations. We outline our roadmap toward version 1.0 (around 50k messages), focusing on dataset scaling and robust benchmarking for Thai scam detection, adversarial stress tests, and PDPA-compliant release artifacts. *ThaiScamBench∗ establishes the first standardized evaluation for Thai scam detection, enabling transparent comparison and artifact-conscious evaluation. This benchmark provides a reproducible foundation for robust Thai-language scam detection.