KMITL

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

Browse

Settings

Subject: search.filters.subject.large language models

Search Results

Now showing 1 - 3 of 3
  • Loading...
    Thumbnail Image
    Some of the metrics are blocked by your 
    consent settings
    Item type:Publication,
    A Comparative Study of Deep Reinforcement Learning Agents for Gold Trading with Technical Indicators and LLM-Filtered News Sentiment
    (2026-06-16)
    Thanasarn, Thanapong
    ;
    Anuntachai, Anuntapat
    ;
    Netisopakul, Ponrudee
    Integrating macroeconomic news into deep reinforcement learning (DRL) for daily gold (XAU/USD) trading remains challenging. This study implements a natural language processing pipeline using majority voting from three large language models (Llama-3.1-8B-Instruct, Qwen3-8B, and Gemma-3-12B-it) to filter articles from The New York Times from 2010 to 2024, yielding 592 relevant articles that are subsequently assigned sentiment scores using FinBERT. We evaluate A2C, PPO, and SAC agents in a FinRL and Stable-Baselines3 framework using a four-fold walk-forward expanding-window protocol and Optuna hyperparameter tuning. The models are compared under two feature settings: (1) six technical indicators only and (2) technical indicators combined with sentiment features. Results show that incorporating LLM-filtered sentiment can modestly improve trading performance for some agents. PPO with sentiment achieves the best average cumulative return (10.05%) and Sharpe ratio (0.74), compared with its indicator-only version (9.80%, 0.72), and slightly outperforms the Buy-and-Hold (B&H) baseline (9.58%, 0.71). A2C also improves with sentiment (9.50% to 9.90%), while SAC shows no improvement in this setting. These findings suggest that LLM-filtered sentiment provides a modest benefit for some DRL agents in daily gold trading in our experiments.
      2
  • Loading...
    Thumbnail Image
    Some of the metrics are blocked by your 
    consent settings
    Item type:Publication,
    Color-Aware Structured Parsing and Self-Consistent Reasoning for Chart Question Answering
    (2025-01-01)
    Vorathammathorn, Supasate
    ;
    Sintarasirikulchai, Wassana
    ;
    Sakdejayont, Theerat
    ;
    Panichpapiboon, Sooksan
    Colors are critical elements in charts. They allow us to visualize data in different series or categories effectively. Currently, generative artificial intelligence such as vision language models (VLMs) and large language models (LLMs) can be combined together to form a pipeline for extracting information and answering questions regarding the data presented on a given chart. However, most chart question answering pipelines still treat every chart as if it were rendered in grayscale, ignoring the valuable chromatic attributes. This paper introduces ColorAware Structured Parsing (CASP) with Self-Consistent Reasoning, a two-stage pipeline that extracts both numerical and chromatic information to improve chart understanding. On a 2,500 question Chart Question Answering (ChartQA) benchmark, CASP attains 91.12% exact-match accuracy, a significant gain over the strongest structured baseline. The answers produced by CASP are also auditable. They include explicit pointers to the table cells used for reasoning. By exploiting both numerical and chromatic information on the chart and enforcing agreement across reasoning paths, CASP turns colors into usable evidence, delivering accurate and transparent answers regarding the information embedded in the chart.
  • Loading...
    Thumbnail Image
    Some of the metrics are blocked by your 
    consent settings
    Item type:Publication,
    DataDecon: Data Cleansing Tools for Large Language Model with Efficient Decontamination Techniques
    (2024-01-01)
    Yuenyong, Sumeth
    ;
    Buppodom, Norapat
    ;
    Sangkaew, Koravich
    ;
    Boonmeeprakob, Konthee
    ;
    Boonkwan, Prachya
    Large language models (LLMs) play an important role in modern NLP technology as they are versatile for a wide array of NLP tasks. However, constructing an LLM is challenging due to concealed construction pipelines, the lack of cleansed datasets, and hyperparameter settings, making it almost irreproducible. This paper presents an efficient pipeline for constructing an LLM tailored to a low-to-medium-sourced language with a high level of data contamination and tools to cleanse the dataset. Following our pipeline, we constructed OpenThaiGPT, an LLM for Thai, with only open-sourced datasets such as CC100, OSCAR, and mC4, and achieved the state-of-the-art accuracies on our downstream tasks. Here, we disclosed the data statistics and all hyperparameter settings for reproducibility.