可調適廣告迷因標題生成模型

Abstract

越來越多品牌利用迷因行銷，讓他們的品牌更貼近目標客戶，引起他們的共鳴。許多迷因標題生成模型利用特定的模板產生迷因標題，但這限制創作的自由度，無法產生具原創性且有影響力的迷因標題。因此，在本研究中，我們提出一個可調適廣告迷因標題生成模型GAMC幫某個品牌的貼文產生迷因標題，我們提出的模型包含五個模組，首先，我們透過視覺特徵提取模組和情感、情緒與幽默特徵擷取模組，從圖片中提取視覺與情緒相關特徵；接著，我們運用共注意力模組學習不同模態特徵間的關係；然後，我們利用大型語言模型生成迷因標題，並利用主資料集訓練模型以增加生成標題的幽默感，其中主資料集包含許多的迷因資料；最後，我們利用品牌資料與可適應模組微調已訓練好的模型，讓生成的迷因標題更加契合品牌形象。實驗結果顯示，我們提出的模型在幽默度、友善性及流暢度等評分指標上均優於比較模型。我們的廣告迷因標題生成模型，可幫助品牌展現其幽默感，提升品牌形象與曝光度，讓它們的貼文更具病毒式擴散能力。

Many companies have used meme marketing to make them more relatable and approachable to their target audience. Many meme caption generation models use custom meme templates to generate meme captions; however, they can only generate meme captions on the pre-trained classes (or topics). These constraints on creative freedom can significantly hinder the ability to produce original and impactful meme captions. Therefore, in this study, we propose an adaptable model to Generate Advertising Meme Captions, called GAMC, for the posts of a brand. The proposed model contains five modules: the visual feature extraction module, the emotion-sentiment-humor (ESH) module, the co-attention module, caption generation module, and adaptation module. First, we apply the visual feature extraction module to extract the visual features and the ESH module to derive the emotion, sentiment, and humor features from the photo of the input post. Next, we use the co-attention module to learn the inter-relationships between features of different modalities. Fourth, we employ the Large Language Model (LLM) to generate the meme caption for the input post in the caption generation module, and train the proposed model by the main dataset to increase the sense of humor of generated captions, where the main dataset contains a large number of meme captions. Finally, we adapt the trained model to the brand by using the adapters and the dataset collected from the brand’s posts to fine-tune the trained model in the adaptation module. The experimental results show that the proposed model outperforms the compared models in terms of humor, benign, and fluency scores. Our model can help businesses reveal their humorous side, enhance their brand image, promote effective communication with their customers, and spark positive word-of-mouth.