運用文字探勘技術在社群行為上之人格預測

Abstract

現今網路無遠弗界，人與人溝通或社交行為，已由早前的書面作業或面對面交談，漸漸成為線上作業。因時代的變遷，人們已較常在社群網路上發表文章及紓發自已的情緒，這些個人行為會透過文字的表達呈現於文章中。本研究資料來源透過Facebook社群行為中文語料的部份進行人格分析，採用最被廣為接受Costa & McCrae(1985)的五大人格特質構面 (Five Factor Model,the Big Five) 。此五大人格特質，分為神經質型、外向型、開放型、 隨和型、嚴謹型五大類。 研究方法採用二種方法進行研究及比較，方法一為關鍵詞彙預測法，透過中研究的BOW - WordNet 擴充詞彙；方法二為機器學習預測法，採用自行開發的程式利用變型的貝氏理論加以研究及實作。研究結果顯示，針對顯著人格加以分析比較，方法一有61%的準確率，方法二有80%的準確率。方法二的實驗結果所預測的準確率高於方法一的預測結果，二項方法當詞彙數夠多時，更可增進研究分析的準確性。另研究透過方法二學習訓練後的詞彙，用來自動擴充方法一的詞彙。在使用測試集的資料加以驗證後，結果顯示有效的增進方法一的預測結果，由66.67%增至73.33%。故本研究證實透過在Facebook的中文貼文可有效的分析個人在社群網路上的人格特質，未來可供後續研究者參考，以及企業應徵人員的參考依據等效益。

Nowadays Internet is used for communication widely. People prefer communicating via Internet Services over talking face-to-face or writing letters. They are more often writing blogs or posting messages on social networks and the personality will be presented by habitual vocabularies they used. This research is analyzing Chinese vocabularies to predict personality from posted contents by Facebook users. The personality classification is based on Five Factor Model (Costa & McCrae, 1985). The five categories are Neuroticism, Extraversion, Openness, Agreeableness and Conscientiousness. This research compares two methods. Method one is key vocabulary prediction by using SINICA BOW-WordNet. Method two is machine learning prediction by using compact Bayes theorem. The results show that the accuracy of method two (80%) is better than method one ( 61%). The accuracy of method two will be better when the sample is enough. The result could be used to extend vocabularies of method one and improvements accuracy from 66.67% to 73.33%. This research demonstrates a different way to analyze personality by analyzing posted contents on Facebook from traditional questionnaire and the contribution of this research can provide helpful reference to HR of enterprise when recruiting employees.