利用卷積神經網路串連新創產業故事鏈

Abstract

發布在網路上的文章數量成長速度非常地快，造成使用者擷取資訊的困難，若能將大量的文章串聯成一個個的故事鏈，則可以幫助使用者快速的閱覽大量文章中的重要資訊。因此，我們提出一個方法串聯新創產業相關的故事鏈。我們的架構包含五個階段，第一階段，我們去除掉對於串連故事鏈影響力小的字詞；第二階段，我們建立一個News-CNN模型從文章中擷取語意的特徵；第三階段，我們將每篇文章的專有名詞提取出來作為專有名詞特徵；第四階段，我們利用兩個擷取出的特徵，計算文章與文章間的相似度，然後將相關的文章串連起來，構成文章層級的故事鏈，接著，我們設定一些故事鏈的規則，以去除不適合的故事鏈；最後，我們將故事鏈提取摘要，建構句子層級的故事鏈。實驗結果顯示，我們提出的方法比SteinerTree更能串連出好的故事鏈，且我們所串連出的故事鏈能夠提供使用者豐富的產業資訊，讓使用者快速掌握相關產業的資訊，協助其擬定相關商業策略。

Nowadays, the amount of articles on the Internet has grown very rapidly. Mining storylines from such a large amount of articles may give us a picture what is going on or the whole picture of related events. Therefore, we propose a framework to build storylines from news articles of entrepreneurial industry. The proposed framework contains five steps. First, we preprocess the articles to remove stop words and low-frequency words. Second, based on Convolutional Neural Network (CNN), we build a model, called News-CNN, to extract semantic features from news articles and transform each news article into a feature vector. Third, we convert the name entities in each news article into a feature vector. Fourth, we use the feature vectors derived by News-CNN and the feature vectors derived from name entities to compute the similarity between news articles and then group similar news articles into a document-based storyline. Also, we set some rules to remove redundant storylines. Finally, based on NMF, we propose a method to summarize each document-based storyline into a sentence-based storyline. The experiment results show our proposed method can generate better storylines than SteinerTree. Also, it can provide valuable business insights for users to implement their business strategies in the related industry.