| 摘要: | 社群媒體網站提供了豐富且多樣性的資訊，使輿情分析與網路聲量偵測，變成調查與理解市場的方法之一。而上述的方法能夠為網路服務平台提供平台資訊以外的訊息，有助於解決客群侷限於特定平台的問題，使平台能夠獲得較為全面性的數據來進行分析與決策。然而，多半的社群媒體都是以文字敘述為主，該如何將輿情分析的結果與文字訊息有效的轉化成數值資料，利於後續數據分析中是本研究關注的主題。
本研究主軸設定在「歌手網路聲量」以及「意見目標的情緒分類」兩項，並應用於「熱門歌曲預測」。在「歌手網路聲量」的項目中，我們鎖定在華語區的歌手及歌曲，針對來自於「批踢踢 PTT 實業坊」的社群媒體，來關注歌手、歌曲的網路聲量。由於一個句子談論內容可能會針對不同的意見目標(Target)來進行評論，造成模型情感標記的困難。
因此本研究進一步採用「意見目標的情感分類(Aspect-based Sentiment Classification, ASC)」，來對與歌手、歌曲相關的評論進行意見目標上情感預測，在ASC的任務中使用運用不同的嵌入層方法，如pre-train 好的 Word Embedding、Character embedding，以及 BERT[5] 來將字詞轉換成向量，並在這些不同嵌入層方法上，運用 Huang 等人[11]的意見目標情感分類方法，以及運用Parikh 等人[17]的自然語言推論(Natural Language Inference)方法，來對意見目標的進行情緒分析。接著運用訓練好的Aspect情感模型來對社群媒體評論進行輿情分析，並將分析結果加入平台的點播量資訊中，來預測歌曲的未來播放量。
並在「熱門歌曲預測」的任務中運用類神經網路來設計預測模型，解決高維度的特徵資料；我們希望透過類神經網路來學習特徵之間的關係，藉此省去以往需要仰賴人工經驗、繁瑣的統計方法來進行特徵處理。實驗結果顯示，加入「歌手網路聲量」以及「意見目標的情緒分析」有助於後續歌曲點播量提升的預測。
;Social media networks provide rich and diverse information, making opinion analysis and network volume analysis have become one of the methods to investigate and understand the market. The above method can provide information other than platform. It can helps to solve the problem that the customer group is restricted to a specific platform, Enable the platform to obtain more comprehensive data for analysis and decision-making. However, most social media are based on text narratives. How to effectively convert the results of public opinion analysis and text messages into numerical data, which will facilitate subsequent data analysis, is the subject of our research.
Our main research is to apply "singer & song network mentions" and "Aspect-based sentiment analysis" to "Hit Song prediction". In "singer & song network mentions" tasks, we focus on Chinese district singers and songs, and follow the social media "PTT" to pay attention to follow the network mentions of singers and songs. Since the content of a sentence may be commented on different targets, it is difficult to label the emotions of the model.
Therefore, In our Research, Further use the "Aspect-based Sentiment Classification" to predict sentiment polarity on discussions related to singers and songs. We use different embedding layer methods in ASC tasks, such as pre-trained Word Embedding, Character embedding, and BERT[5] to transfer words and character into vectors, and on these embedding layer methods, Huang et al.[11] Aspect-based Sentiment Classification method and Parikh et al.[17] natural language inference method are used for the Aspect-based Sentiment Classification tasks. Then use the pre-train Aspect sentiment model to analyze the social media comments, and add the analysis results to the platform’s on-demand information to predict the future play volume of songs.
In Hit Song prediction task, use neural network to design prediction model, Solve the high dimensional features data and learning the relationship between features through neural networks, This eliminates the need to rely on artificial experience and cumbersome statistical methods for feature processing in the past. The experimental results show that the addition of "network mention information" and "sentiment analysis" can improve hit song prediction performance. |
