- 學位論文
基於自然語言處理的股票消息生成
Stock News Generation with Natural Language Processing
摘要
現有與股票市場相關的機器學習研究，多是以股票資訊做股價預測，然而社群公司的受眾所想要閱讀的，是投資評析股票消息的『文字』，而非單純股價預測的『數值』。所以，社群公司需浪費許多時間人力成本把股價預測數值結果表達成股票消息的文字，以利受眾閱讀。故本研究提出了直接由股票資訊生成股票消息的創新構想，透過機器學習自然語言生成的方法來直接生成股票消息。我們使用公開的股票資訊資料集，以 teacher forcing方法訓練生成器，並採用對抗式的訓練方式來解決exposure bias問題，希望可以生成與人類所寫相仿的股票消息。在我們的實驗中，我們不僅使用通用的指標來評估生成的消息，也找了專家來對這些消息做評估。根據指標和專家評估的結果，生成的消息可以根據股票主題生成類似人類產生的股票消息，更進一步地分析，我們發現模型在這個任務中已經可以理解"多"、"空"、"漲"與"崩"等情緒，並能讀懂每個字詞的含義，如"長榮"與"長榮航"的不同、"2330"是"台積電"股票標號、"APPLE"與"蘋果"在股票中都代表同一家公司，幫助社群公司節省時間人力成本。
並列摘要
Nowadays, machine learning researches on the stock market mainly apply stock information for price prediction. However, their obtained results may not help the social technology company's target audience, who wants to read stock news more than the stock prices. The social technology company needs to take time and labor costs to generate stock news from the price prediction results. In this paper, we propose an innovative idea to generate stock news directly from stock information through natural language generation with machine learning. We train the generator using teacher forcing with an open-source dataset of stock topics and news. Adversarial training is applied to solve exposure bias for producing sequences indistinguishable from human-generated ones. We evaluate the generated sentences using general metrics and human evaluation. The generator can consider the stock topic and then generate human-generated-like news. We found that the generator has learned the sentimental conditions "多", "空", "漲", and "崩" to generate the proper stock news. It can understand the meaning of the stock topic. The generator can recognize the difference between "長榮" and "長榮航". The generator knows that "2330" is the stock symbol of "台積電". The generator understands synonyms like "APPLE" and "蘋果" as the same (company name) in the stock market. It can help the social technology company to save time and labor costs.