面向情緒分析及文件產生之探討：以遊戲評論為例

Abstract

在機器學習的廣泛運用之前，最常在網路上看見使用文件產生的機制通常為隨機更改一些關鍵字，並產生出一段文字，例如時常在年輕人之間流行的輸入名字就會產生各式各樣結果的測驗。在以機器學習進行文件產生的研究中，隨著人們在網路上進行聊天、提問、評論、投稿、甚至出版，這些都成為文件產生很好的訓練資料來源。也有許多投入產生更好的文件產生模型的研究，讓產生出來的文件不是只有一段文字，而是有意義的、能讓人讀懂並且邏輯通順的文章。本文藉由面向情緒和GPT-2模型對遊戲評論進行文件產生的深入探討。不只用評分產生文件，還利用關鍵字抽取面向特徵，再以面向特徵的不同表示方法進行實驗，加以觀察哪一種表示方法，產生出來的文件最能維持遊戲評論的情緒跟面向的情緒。效能的評估方式以正規化方均根差(Normalized root mean square error)作比較。

Before the widely use of machine learning, the most common text generation on the Internet is randomly changing some keywords and generating a simple text. For example, the game of inputting your name and then generating many kinds of text is often popular among young people.In text generation with machine learning, people chatting, asking questions, commenting, submitting, even publishing on the Internet, have become good resources of training data. There is also a lot of research to produce better text generation models, so that the text generated are not just a paragraph, but a meaningful, understandable, and logical article.This thesis uses aspect-based sentiment and a GPT-2 model to study text generation of game reviews. Not just using user scores to generate text, but also using keywords to extract aspects, and then experiment with different ways to represent aspects. We want to observe which method can generate game reviews with most similar sentiment and aspect-based sentiment. The performance evaluation method is Normalized root mean square error.