Practical Research on BERT-based Domain Knowledge Question Answering System

• 1、Guangdong Haiqixing Marine Technology Co., Ltd.

Abstract：Domain knowledge question answering systems face practical constraints in offline operation, data localization, and long-term invocation costs. Current mainstream solutions tend to adopt large-parameter generative models. However, for domain-specific tasks, this paper argues that most parameters in large models are used to encode general linguistic knowledge, while the proportion of parameters truly serving a specific domain is extremely low, resulting in significant model redundancy. Taking marine domain knowledge services as the background, this paper returns to the discriminative architecture based on experiments with edge-side Large Language Models (LLM), and adopts a "text classification + preset answers" approach, using bert-base-chinese with only about 102 million (102M) parameters to build an offline-deployable question answering system. Experiments first compared the performance of Ollama CPU inference and RKLLM NPU inference on domestic ARM development boards, revealing limitations of edge-side LLMs in latency, hallucination, and resource occupation, leading to the shift toward the BERT classification scheme. This paper emphasizes "domain means specialty": the training data originate from national standards, industry norms, and enterprise knowledge to ensure authority. A reproducible engineering pipeline covering knowledge extraction, training-data construction, model training, and deployment verification is established. A confidence-analysis-based training-data deduplication method is proposed for the semantic-consistency problem where the same concept is expressed differently across multiple authoritative sources; taking the short question "sea temperature (海温)" as an example, the confidence increased from 49.85% to 94.16%, and the validation accuracy reached 96.79%. Comparative experiments show that the proposed scheme requires no GPU and runs smoothly on x86 and ARM CPUs (approximately 87ms on x86 and 307ms on ARM), serving as a useful exploration of localized AI deployment with practical value in offline, confidential, and standard-question-answering scenarios, validating the design philosophy that "a suitable model matters more than a larger one."

Keywords： AI BERT knowledge question answering text classification training data cleaning edge deployment