隨著網路的快速發展，對網路評論的研究也日益增加，探討評論中是否含有建議內容已成為一個重要的研究方向。然而，隨著評論數量持續攀升，要瀏覽所有評論並篩選出所需的建議資訊變得愈加困難，因此如何從評論文本中有效萃取建議資訊成為一個值得深入探討的問題。 由於針對不同領域標記資料集需要大量成本，本研究採用領域適應架構進行跨領域訓練。在現有應用領域適應技術於跨領域建議探勘的研究中，較少有關於利用領域特定句法特徵的探討。因此本研究以領域適應架構為基礎，引入詞性標籤序列來表示句法，並使用Zhu et al. (2023)提出的注意力網路來萃取領域特定句法特徵，藉由提升模型對領域特定句法的關注程度，以改善跨領域建議探勘的準確性和效能。實驗結果顯示，與僅使用文本輸入的領域適應架構模型相比，額外引入詞性標籤序列來萃取領域特定句法特徵可以提升跨領域建議探勘的 F1 分數。此外，透過對含有領域特定句法的文本進行分析，本研究提出的方法能有效過濾掉更多不含建議內容的評論，進而減少使用者在大量評論中尋找建議資訊所需的時間。;With the rapid development of the internet, research on online reviews has also increased significantly. Investigating whether reviews contain suggestion content has become an important research direction. However, as the volume of reviews continues to rise, browsing all reviews to filter out the necessary suggestion information is becoming increasingly challenging. Thus, effectively extracting suggestion information from review texts has become a topic worthy of in-depth exploration. Due to the high cost of labeling datasets for different domains, this study adopts a domain adaptation architecture for cross-domain training. In existing studies that apply domain adaptation techniques to cross-domain suggestion mining, there is relatively little discussion about utilizing domain-specific syntactic features. Therefore, this study introduces part-of-speech (POS) tag sequences to represent syntax within a domain adaptation framework and employs the attention network proposed by Zhu et al. (2023) to extract domain-specific syntactic features. By enhancing the model′s focus on domain-specific syntax, the goal is to improve the accuracy and performance of cross-domain suggestion mining. Experimental results show that, compared to domain adaptation models that use only textual input, the additional incorporation of POS tag sequences to extract domain-specific syntactic features can improve the F1 score in cross-domain suggestion mining. Furthermore, by analyzing texts containing domain-specific syntax, the proposed method effectively filters out more reviews without suggestion content, thereby reducing the time users spend searching for suggestion information among a large number of reviews.
