Publications

Detailed Information

머신러닝 기법을 적용한 게임 산업 내 트위터 구전 효과 분석 : 구전의 양과 방향성 및 정보 불일치의 효과를 중심으로 : A Study of Twitter Effect in the Game Industry Using Machine Learning : Focusing on Volume, Valence, and Information Inconsistency

Cited 0 time in Web of Science Cited 0 time in Scopus
Authors

이현경; 김상훈; 이지수

Issue Date
2020-04
Publisher
대한경영학회
Citation
대한경영학회지, Vol.33 No.4, pp.649-665
Abstract
오늘날 온라인을 통해 생성되는 구전은 소비자들의 의사결정에 많은 영향을 미치고 있는데, 그 중 마이크로블로그를 통한 구전은 제품 정보를 실시간으로 빠르게 확산시키는 역할을 한다. 본 연구에서는 대표적 마이크로블로그인 트위터를 통해 형성되는 구전을 중심으로 게임 산업 내 트위터 구전 효과를 검증하였다. 트위터구전의 양과 방향성 및 정보 불일치가 게임 판매 성과에 미치는 영향을 검증하기 위해 총 47개의 비디오게임의 주별 판매량과 이들 게임과 관련한 223,345개의 트위터 리뷰를 수집하였고, 텍스트 마이닝(text mining)을 통해 비정형 텍스트인 트위터 리뷰를 분석 가능한 형태로 가공하였다. 리뷰의 긍 부정 방향성을 도출하기위해 본 연구에서는 머신러닝 기법으로 감성 분석(sentiment analysis)을 실시하였다. 게임 카테고리 내 리뷰데이터를 활용하여 나이브 베이지안 분류 모형을 구축하였고, 이를 트위터 리뷰의 긍 부정 판별에 적용하였다.해당 모형의 분류 정확도는 87%로 매우 높은 수준을 나타냈다. 이러한 방식은 학습 과정에서 게임이라는제품군 특성을 반영함으로써 감성 판별의 정확도를 높일 수 있다는 장점이 있다. 한편 본 연구에서는 정보엔트로피의 개념을 통해 긍정적 의견과 부정적 의견이 대립하는 정도인 정보 불일치를 측정하였고, 구전의정보 불일치가 리뷰 효과를 조절하는지를 검증하고자 하였다. 패널 회귀분석 결과, 트위터 구전의 양과 방향성은 게임 판매량에 긍정적인 영향을 미치는 것으로 나타났다. 또한, 정보 불일치의 척도인 엔트로피가 클수록트위터 구전의 양과 방향성이 게임 판매량에 미치는 긍정적 효과가 감소함을 확인하였다. 본 연구는 트위터라는매체의 영향력을 검증함으로써 온라인 구전 연구를 확장하는 동시에 향후 연구에서 활용 가능한 텍스트 데이터분석 기법을 제안하고 있다. 실무적으로는 게임 산업에 있어 마이크로블로그를 통한 구전 관리의 중요성을시사하고 있다.
Microblogs (e.g., Twitter, Instagram, and Facebook) have become important informationsources for consumers today. Word-of-mouth shared through microblogs can affect potentialconsumers product adoption by immediately disseminating others product evaluations. This studyaims to verify the Twitter effect in the video game industry. Twitter, a representative microblog,may have a crucial influence on the adoption of video games considering that video games areexperiential products whose quality inferences much rely on others evaluations and distributionstrategy focuses on the early period of releases. This study suggests that the volume and the valenceof Twitter reviews are positively related to the financial performance of video games. Further, italso suggests that information inconsistency, the extent of a conflict between positive and negativereviews, moderates the impacts of the volume and valence on sales.To test the hypotheses, we collected the weekly sales of 47 video games and 223,345Twitter posts for the games. Twitter reviews are text-based, requiring a different approach than thereview ratings, which are typically used in previous literature. In this study, a text mining techniquewas applied to transform the unstructured text data into an analytical format. To judge the emotionalvalence of Twitter reviews, a machine learning technique was applied as a tool for sentiment analysis.Specifically, we used game reviews from a review platform as training data and constructed a NaïveBayesian classifier based on words in the reviews. The model shows a high level of accuracy, 87%,in the remaining test set. We then classified each Twitter post into either positive or negative sentimentbased on the classification probability of our classifier. This approach has the advantage of improvingthe accuracy of sentiment analysis by reflecting the characteristics of games in the learning process.After measured the valence of Twitter reviews, we then defined and measured the information inconsistency. Information inconsistency, here, is defined as the extent of a conflict between positiveand negative Twitter reviews. For example, a game in which the percentage of positive and negativeopinions is equal is called to have higher information inconsistent than a game whose reviews arebiased toward either positive or negative. Such bi-modality which results from opinion oppositioncannot be captured by variance, the common measure of divergence of a distribution. To measureinformation inconsistency, this study adopted the concept of entropy of information theory. Entropymeasures information inconsistency from the proportions of multiple categories (i.e., positive vs.negative). We argued that information inconsistency lowers the certainty of the information, thusdecreases the impacts of volume and valence of Twitter reviews.The analysis results showed that the volume of Twitter reviews has a positive effect ongames sales. As expected, the Twitter reviews valence also found to have positive impact. Also,high entropy or high information inconsistency of Twitter sentiment reduced the effects of the volumeand the valence to games sales, suggesting that the word-of-mouth effect can be moderated whenreviews are sharply divided. The contributions of this study are as follows. First, it extends the online word-of-mouthliterature by considering a relatively new channel, Twitter. Second, it demonstrates the specific processof machine learning-based sentiment analysis in detail. Our methodology accomplishes high accuracyby reflecting category-specific sentiment in the learning process. The approach is applicable to researchon opinion mining in other categories. Third, we suggested a new moderator of word-of-mouth effect,information inconsistency and measured it in a novel way. Considering that the extent of opinionconflicts has not been received much attention so far, our study contributes to online review literatureby suggesting an important distributional feature. Lastly, the findings of this research also providepractical implications for the game companies. This study emphasizes the significance of microblogreputation management. Based on the results, several management tactics to leverage the reputationeffect are also suggested.
ISSN
1226-2234
URI
https://hdl.handle.net/10371/198044
DOI
https://doi.org/10.18032/kaaba.2020.33.4.649
Files in This Item:
There are no files associated with this item.
Appears in Collections:

Altmetrics

Item View & Download Count

  • mendeley

Items in S-Space are protected by copyright, with all rights reserved, unless otherwise indicated.

Share