| 摘要: | 隨著YouTube成為全球第二大的熱門網站，不斷增加的流量、用戶量與營收，新興的職業與商業模型順應而生，背後的商機非常可觀，在流量變現的時代中，不僅僅是影片上傳者想要精準行銷，使用者們在茫茫的影片海中也想找到適合自己的影片，為了讓使用者可以快速搜尋到想要的影片，管理與分類這些巨量的影音內容成為主要的任務。
而YouTube上除了上傳影片資料外，還包含其他的使用者產生的內容(User-generated content)，如影片標題、標籤、描述、影片評論等等，大部分都是由創作者在上傳影片時自己輸入的，而評論則是由大量的使用者共同創造的，因此我們希望透過結合創作者與使用者共同產生的資訊，能夠提供更加客觀的分類。以往的YouTube影片分類方法多用機器學習方法分析文本，本篇論文利用深度學習法將網路影片分類到指定的情緒類別中。
本文利用text-CNN提取影片標題、標籤、描述、評論四組文字的局部特徵，再利用Bi-LSTM分析較長的評論特徵，將YouTube影片更有效分類到適合的情緒中，最高達到92.19%。
;With YouTube has been becoming the world’s second most popular website, YouTube has been increasing traffic, user volume and revenue. New careers and business models have been emerging, and the business opportunities behind it are very impressive. content providers want to precision marketing, and users also want to find videos that suit them in the vast of videos. In order for users to quickly search for the videos they want, managing and categorizing these huge amounts of videos become the main task.
In addition to uploading video, YouTube also contains other user-generated content, such as video descriptions, keywords, titles, comments, etc., most of which are uploaded by the content provide. While the comments are created by a large number of users, we hope to provide a more objective classification by combining the information generated by content provide and users. In the past, YouTube video classification methods mostly used machine learning methods to analyze text. This paper uses deep learning methods to classify Internet videos into designated emotional categories.
This paper uses text-CNN to extract the local features of the four groups of texts: title, keywords, description, and comments, and then uses Bi-LSTM to analyze longer comment features to more effectively classify YouTube videos into suitable emotions, up to 92.19%. |
