以時間分析與多維度語句呈現為基礎之熱門話題萃取

Abstract

主題偵測(Topic Detection)是主題偵測與追蹤(Topic Detection and Tracking)裡其中一個研究領域，該領域試著從新聞媒體裡，進行搜尋、組織及建構文字形式的新聞資料。我們的研究為如何偵測「熱門」的主題(Hot Topic)。所謂熱門的主題，是在某一段時間之內，它會被很多人常常討論與報導。在之前的研究裡，可透過TF*PDF計算文字權重的式子，找到描述熱門主題的「熱門字」 (Hot Term)。不過它仍然會有一些問題存在：(一)只以字的出現頻率和文獻比例頻率為基礎的TF*PDF，萃取熱門字會導致不可靠的結果;(二)只用單一的句子向量並不足以表達句子的涵義。 因此我們提出了改良的熱門主題的萃取系統，來解決上述的兩個問題。首先，我們透過紀錄字在時間上的使用變化，來萃取熱門字;也就是說，追蹤任一個字的生命週期，可以幫助我們來分辨它是否為足以描述「熱門」主題的字。之後，我們使用多維度的句子向量，來描述句子的資訊。最後，我們對所有新聞報導裡的句子進行叢集(cluster)，而每一個叢集代表著一個新聞話題。透過以上兩個流程的改善，根據實驗結果顯示，不但增進了每一個叢集的品質，也能夠萃取出一段時間內所包含的熱門主題。

Topic detection is part of the Topic Detection and Tracking field, which seeks to develop technologies that search, organize, and structure news-oriented textual materials from various broadcast news media. We are interested in detecting “hot” topics that are frequently discussed by people in a given period of time. A prior work on hot topic extraction that designed an innovative term-weighting scheme called TF*PDF, which extracts “hot” terms that can describe hot topics. One of the problems that happens in the process of extracting hot topics using TF*PDF is the unreliability of results when the weight is determined solely on term frequency and document frequency. Another problem is that using one single vector misrepresents the meaning of a sentence. We propose a hot topic extraction system that aims to solve the two problems mentioned above. First, we extract the hot terms by capturing their variations of the time distribution within a timeline. In other words, tracking the life cycles of the terms can help us differentiate which term is a real hot term that describes a hot topic. Second, we use multi-dimensional sentence vectors to feature the information of a sentence. Finally we group the sentences of news report into clusters, which represent hot topics. Clustering the sentences by the multi-dimensional sentence vectors not only improves the quality of each cluster, but also extracts most of the actual hot topics over a period of time.