運用深度學習從線上評論萃取顧客痛點

Abstract

在網際網路時代，顧客能夠在社交媒體以及線上評論中分享他們的感受。了解這些顧客需求，尤其是了解顧客不滿意的地方，對於服務改進與創新相當重要，因為這些不滿意的顧客指出了服務無法滿足他們期望的具體之處。顧客不滿意可以分為兩種類型：抱怨和痛點。痛點是帶有具體問題或是不滿意的抱怨，提供可以付諸行為的見解。進行痛點分析有助於公司作出明智的決策。 在過去的研究中，痛點被萃取為關鍵詞或是整個句子，可能導致語意上的模糊或是包含不相關的訊息。此外，只有少數研究包含了痛點分類，提供痛點在類別層面上的評估。 在本研究中，我們提出了一個兩階段的模型來預測顧客評論中的痛點，並將獲得的痛點分類至事先定義的類別之中。我們進一步在不同的領域測試了痛點萃取模型的預測能力。另外，我們採用特殊標記來表示整個評論以進行痛點分類。實驗結果顯示了我們所提出的痛點分析框架的有效性。

In the age of the Internet, customers can share their feelings on social media or through online reviews. Understanding these customer needs, especially customer dissatisfaction, is important for service improvement and innovation since unsatisfied customers highlight specific areas where services do not meet their expectations. There are two types of customer dissatisfaction: complaints and pain points. Pain points are complaints with specific problems or dissatisfactions, which provide actionable insights. Conducting pain point analysis assists companies in making informed decisions. Pain points were extracted as keywords or entire sentences in previous studies, potentially leading to semantic ambiguity or the inclusion of irrelevant information. Additionally, only a few prior studies include pain points categorization, which enables evaluation of pain points at category level. In this study, we propose a two-phase model to predict the pain point expressions in customer reviews and classify the obtained pain points into predefined categories. We further test the pain point extraction model across different domains. Besides, we adopt special tokens to represent entire reviews for pain point categorization. Experimental results show the effectiveness of our proposed framework for pain point analysis.