觀眾評論中的續集線索:結合面向情緒分析與結構變數之電影續集預測研究

No Thumbnail Available

Date

2025

Journal Title

Journal ISSN

Volume Title

Publisher

Abstract

在影視產業中,電影續集是延續票房表現並降低投資風險的重要策略之一。傳統評估電影續集潛力的方法多以票房、成本及評分等結構性指標為主,較少關注觀眾的情緒反應與留言內容對續集決策的影響。本研究以 IMDb 平台上的英文觀眾評論與電影結構資料為基礎,結合語意分析和機器學習方法,探討觀眾的語意參與是否能有效提升對電影續集推出的預測能力。研究設計上,本研究建立三組不同的特徵資料,包含純結構資料、純語意資料以及兩者的綜合資料,同時搭配羅吉斯迴歸、隨機森林和支持向量機三種分類模型進行交叉比較分析。在語意資料處理部分,本研究透過斷句、面向分類以及利用 BERT 模型進行情緒標註,將觀眾的非結構化留言轉化成可量化的面向情緒變數,並使用 AUC、F1-score 與 Balanced Accuracy 等多項指標來評估模型預測效能。實證結果發現,單獨使用語意資料的預測能力雖然略低於傳統的結構性指標,但將語意資料與結構資料結合後能夠顯著提升整體預測效果,證實語意資料與結構資料之間具有互補的關係。此外,三種演算法之中,隨機森林模型整體表現最為穩定且出色,尤其適合用於高維與複雜資料情境。整合語意與結構資料的模型在樣本內與樣本外測試中皆展現最佳的效能,進一步證明觀眾語意參與可以作為評估續集潛力的有效指標。本研究的結果不僅驗證了觀眾留言在續集決策中的實質價值,也提供電影產業一個結合自然語言處理與機器學習的創新工具,協助業者更準確地掌握觀眾期望,進而有效規劃內容生產與投資決策。
Sequels play a crucial role in the film industry by extending box office performance and reducing investment risks. Traditional methods for predicting sequel production primarily rely on structural indicators such as box office revenue, production costs, and ratings. This study examines whether audience-generated content—specifically user reviews on IMDb—can improve sequel prediction when analyzed through semantic features. Three datasets were constructed: structural-only, semantic-only, and a combined dataset. Sentiment labeling was conducted using BERT after sentence segmentation and aspect classification. Logistic regression, random forest, and support vector machines were applied to evaluate predictive performance using AUC, F1-score, and balanced accuracy.Results show that while semantic features alone yield slightly lower performance than structural features, combining both significantly enhances model accuracy. Among classifiers, the random forest model performs most robustly, particularly in complex data settings. The findings suggest that audience sentiment offers complementary value to traditional metrics and can serve as an effective indicator for sequel decision-making.

Description

Keywords

續集預測, 語意分析, IMDb, BERT, 觀眾評論, 機器學習, sequel prediction, sentiment analysis, IMDb, BERT, user reviews, machine learning

Citation

Collections

Endorsement

Review

Supplemented By

Referenced By