由於網路的便利，顧客可以在網路上留下對於店家的評論，並成為他人在選擇店家時的參考，而店家所累積的評論往往數量龐大，若沒有經過整理則無法讓人馬上了解其優缺點。透過人工整理評論顯然非常繁瑣且缺乏效率，所以利用自然語言處理的技術來進行自動分析成為了一個重要且普遍的解決方法。本篇論文的目標為自動提取顧客評論的關鍵字和產生摘要，然而以往進行文字摘要(Text Summarization)的模型如T5[1]、BART[2]，這些模型訓練的資料為新聞、期刊等文章，其大多數有特定的段落格式，且句子之間具有上下文關係，而顧客評論之間則沒有，這些原因使先前提到的這些預訓練好的模型不適合直接應用在顧客評論上。因此我們建立了一個系統和流程來自動提取顧客評論的關鍵字和摘要，此系統目前應用的範圍為旅館評論，往後希望能將其應用在其他不同主題的評論上(如餐廳、電影評論等等)。 在顧客評論自動關鍵字提取與摘要系統的流程中，我們使用預訓練好的T5模型過濾文法錯誤的評論，避免影響系統後續的分析；使用BERT[3]模型提取評論的關鍵字，在訓練階段時使用少量人工定義好的關鍵詞自動標記評論中的字詞，並將標記好的評論作為訓練資料，以此讓模型自動學習關鍵詞在評論中的特徵；透過POS-Tagging和Dependency Parsing來提取關鍵字的形容詞，顯示出明確的優缺點；最後利用預訓練好的SBERT[22]模型產生的評論特徵計算句子相似度來進行評論的分群與摘要。 ;Due to the convenience of the Internet, customers can leave store reviews on the Internet and become a reference for others when choosing a store. The goal of this paper is to automatically extract keywords from customer reviews and generate summarization. However, in the past, text summarization models such as T5[1] and BART[2] were trained for news and journal articles, and these articles have specific paragraph formats and contextual relationships between sentences, but customer reviews don’t have these features. For these reasons, the pre-trained models mentioned earlier are not suitable for direct application on customer reviews. Therefore, we have established a system and process to automatically extract keywords from customer reviews and generate summarization. The current application scope of this system is hotel reviews, and we hope to apply it to other reviews on different topics (such as restaurants, movie reviews, etc.). In the process of the automatic keyword extraction and summarization system for customer reviews, we use the pre-trained T5 model to filter grammatically wrong reviews to avoid affecting the subsequent analysis of the system. Use the BERT[3] model to extract keywords for reviews, and during the training phase, a small number of manually defined keywords are used to automatically label the words in the reviews, and the labeled reviews are used as training data to make the model automatically learns the features of the keywords in the reviews. Extract keyword adjectives through POS-Tagging and Dependency Parsing to show clear advantages and disadvantages. Finally, use the review features generated by the pre-trained SBERT[22] model to calculate sentence similarity to cluster and summarize reviews.
