The Revision of the Chinese Linguistic Inquiry and Word Count Dictionary 2015
以自動化日常語言分析展現心理特性的研究，近年來相當受到關注，語文探索與字詞計算（Linguistic Inquiry and Word Count，簡稱LIWC）就是一項廣受學者青睞的分析工具。LIWC歷經幾次的改版，近期對LIWC2007的詞典做了大幅的增刪，並在2015正式發布最新版本。本研究之目的，即在對LIWC2015詞典建立相對應的中文版詞典，進行信效度檢驗，並介紹相關應用文獻。研究一以中文版 LIWC2007詞典為基礎，對照 LIWC2015詞典的改版，進行了相對應的類別增刪及語詞的增補。研究二蒐集各類題材的部落格文本為材料，並將各文本以奇偶句分成兩文本分別進行 LIWC分析，並檢驗其在各類別使用率的相關性，進行信度分析。研究三以Ptt電子佈告欄上的Hate版與Sad版文章各50篇進行書寫差異比較，以檢驗CLIWC2015的效度。本文並對LIWC與大數據分析相關之文獻進行介紹，同時說明LIWC的優勢與限制。期待透過CLIWC2015的修訂完成，對研究華語文使用者的心理特性探討，提供一項研究利器。
Automated analysis of natural language in its daily use has shown to be effective in capturing psychological characteristics in the literature. Linguistic Inquiry and Word Count (LIWC), developed by Pennebaker and his research team, is one of the most commonly used text analysis tools in the social sciences. The essential assumption of LIWC is that the frequencies of word usage in certain categories serve as language markers that index individuals' inner thoughts and psychological processes. LIWC contains two parts, the computer software and the dictionary. The computer software is used to calculate the frequency of words in each category. The dictionary is the LIWC key classifying words into categories. LIWC2015 is the latest dictionary, and is based on a significant revision of its predecessor, the LIWC2007 dictionary. The aim of the current study is to develop a corresponding Chinese version of the LIWC2015 dictionary (CLIWC2015) and demonstrate its reliability and validity.
Based on the Chinese LIWC 2007 dictionary, we revised CLIWC2015 by adding and deleting corresponding categories of the LIWC2015 dictionary. We described the details of the process in Study 1. There is a total of 10,795 words belonging to 79 categories in CLIWC2015, including 25 linguistic process categories and 54 psychological process categories. Study 2 collected 100 texts from blog posts on various topics. The average total word count in each post was 1,290 in Study 2. To calculate the reliability, sentences in each text were ordered first, and then odd- and even-numbered sentences were grouped into two subtexts. LIWC indices were calculated for each subtext, and then correlation coefficients between the corresponding subtexts for each language category were used for reliability analyses. Results showed that all word categories demonstrated strong correlation effects except one punctuation category which calculated the frequency of the dashes usage. One possible explanation is that dashes is not a commonly used punctuation mark in the blog posts which could have lowered the reliability. To examine the validity, study 3 collected 100 posts from the Ptt bulletin board system, 50 of which were from the "hate" board, and the rest were from the "sad" board. The average total word count in each post was 164 in Study 3. The two sets' linguistic features were compared. Consistent with our hypotheses, "hate" board posts used significantly more anger, swear and netspeak words, and exclamation marks. In contrast, "sad" board posts used significantly more first-personal singular pronouns, sad, anxiety and cognitive words, and higher cognitive complexity words. Across studies 2 and 3, our findings supported the reliability and validity of the CLIWC2015 well.
Unlike traditional content analysis, which requires a great deal of time and effort, one of the most important strengths of LIWC is the ability to analyze huge text files rapidly. Recently, more and more research has applied LIWC to analyze big data. In the last part of this article, we also discussed the implications of using CLIWC2015 and its applications in Chinese culture and big data analytics.