Please use this identifier to cite or link to this item:
Title: 使用深度學習方法於產品評論之建議探勘
Mining Suggestions from Product Reviews by Deep Learning Methods
Authors: 侯文娟
Hou, Wen-Juan
Keywords: 深度學習
Deep Learning
Product Reviews
Suggestion Mining
Word Vector
Word Frequency
Issue Date: 2019
Abstract: 隨著網路的普及,許多的資訊愈來愈容易被搜索。其中,產品評論相關的訊息也隨著社群網路的發展被公開在社群中。這類的社群包含,論壇、社群網站、和產品官方網站等,甚至有組織專門在蒐集這些評論後組成評論相關的網站,並且將評論分類,給受用戶查看。消費者可以在上述的網站中查看產品的使用狀況與心得和是否符合自身所需,再決定是否購買;產品提供者也可以透過評論持續收集使用者的使用狀況與心得,對產品進行迭代設計,對產品進行改良以符合大眾的需求。 本研究將評論分為建議句 (suggestion)和非建議句 (non-suggestion)。使用 Stanford Core NLP 斷詞系統將文本以詞為單位進行處理;詞的表達方式分為兩種,詞向量與詞頻率;模型採用深度學習類神經網路,分為全連結類神經網路 (Full Connected Neural Networks, FCNN)、卷積類神經網路 (Convolutional Neural Networks, CNN)與長短期記憶類神經網路 (Long Short-Term Memory, LSTM)。本研究以 Z-Score 方式標準化詞頻率表達式,並用全連結類神經網路訓練,此架構即可和利用詞向量在卷積類神經網路與長短期記憶類神經網路有差不多效果,但速度上快非常多。此外,本研究提供結合詞向量與詞頻率表達式在上述三種模型上訓練,進而對結果做比較分析。效能的評估方式以精準率 (Precision)、召回率 (Recall) 和 F1 分數 (F1-measure, F1)作比較。
With the popularity of the Internet, much of the information are searched easily. The information related to product reviews is also pushed to the social community, such as forums, social media, and official product websites. There are some organizations that collect the comments to form a review-based website and classify the comments to potential customers. Customers can see the usage and experience of the products to determine whether the products satisfy their needs to decide whether they purchase the products. The product providers can collect the user's usage and experience through comments, and iterate the product design to improve the product and satisfy the needs of customers. This study divided comments into two class, suggestion and non-suggestion. The Stanford Core NLP can segment corpus in word units. The expression of words were divided into two types, word vector and word frequency. The model uses deep learning neural network, which include Fully Connected Neural Networks (FCNN), Convolutional Neural Networks (CNN) and Long-Short Term Memory (LSTM) Neural Networks. This study normalized the word frequency in the Z-Score and trained it with a fully connected neural network. This architecture's performance can be similar to the architecture of word vectors in convolutional neural networks and long-short term memory neural networks, but the speed is very fast. In addition, this study provides a combination of word vector and word frequency. The method of performance evaluation is based on Precision, Recall, and F1-measure (F1).
Other Identifiers: G060647031S
Appears in Collections:學位論文

Files in This Item:
File SizeFormat 
060647031s01.pdf2.19 MBAdobe PDFView/Open

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.