面向方面级情感分析的QLoRA轻量微调与提示视图一致性探索

QLoRA Lightweight Fine-Tuning and Prompt-View Consistency Exploration for Aspect-Based Sentiment Analysis

摘要: 方面级情感分析需要同时抽取评论中的方面词并判定对应情感，传统任务专用模型在资源受限场景中往往需要针对结构和标签重新设计。本文将SemEval-2014 Restaurant的方面–情感抽取统一表述为严格JSON数组生成任务，基于本地Qwen2.5-7B-Instruct构建Prompt-only、4-bit QLoRA及双提示视图一致性约束Co-LoRA实验。所有结果均在Restaurant测试集full800上以aspect-sentiment二元组micro-F1、JSON有效率和精确匹配率评价。实验显示，schema Prompt-only的micro-F1为0.4815；QLoRA-500、QLoRA-1000与QLoRA-full分别达到0.6126、0.6529和0.7334，QLoRA-full相对最优Prompt-only基线提升0.2519，精确匹配率从0.3038提升到0.6913。QLoRA-full单轮训练耗时831.9 s，峰值显存7.91 GB。进一步的Co-LoRA-1000探索获得较高的多提示预测一致性，但schema micro-F1为0.5657，未超过常规QLoRA。结果表明，QLoRA能够以较低资源有效适配结构化ABSA，而提示一致性约束仍需在更充分的对照与多域实验中验证。

Abstract: Aspect-based sentiment analysis requires jointly extracting aspect terms from reviews and determining their corresponding polarities. Traditional task-specific models often need to be redesigned for structures and labels in resource-constrained settings. We formulate aspect-sentiment extraction on the SemEval-2014 Restaurant dataset as strict JSON-array generation and conduct Prompt-only, 4-bit QLoRA, and dual-prompt-view consistency-constrained Co-LoRA experiments based on local Qwen2.5-7B-Instruct. All results are evaluated on the full 800-item Restaurant test set using aspect-sentiment tuple micro-F1, JSON validity, and Exact Match. The experiment shows that the schema Prompt-only micro-F1 is 0.4815, while QLoRA-500, QLoRA-1000, and QLoRA-full achieve 0.6126, 0.6529, and 0.7334, respectively. QLoRA-full improves the best Prompt-only baseline by 0.2519 and raises Exact Match from 0.3038 to 0.6913. It completes one training epoch in 831.9 s with a peak GPU memory usage of 7.91 GB. The exploratory Co-LoRA-1000 setting attains relatively high multi-prompt prediction consistency, but its schema micro-F1 of 0.5657 does not exceed conventional QLoRA. These results show that QLoRA can effectively adapt structured ABSA under modest resource budgets, whereas prompt-consistency constraints require validation with more comprehensive controls and multi-domain experiments.