基於預訓練語言模型的多面向電商產品文案生成

Abstract

本篇文章的研究目的為找出如何有效地透過預訓練語言模型生成商品文案，一篇成功的商品文案應具備不同敘述類型的多樣性、保留商品相關資訊的準確性以及提高與商品內容本身的相關性。

而在本研究中透過在GPT-2微調的過程中分別加入分類標籤、關鍵字組作為嵌入維度，使模型能夠學習如何根據輸入的內容生成相對應的內容。在以生成不同敘述類型為目標的實驗中，採用不同類型詞彙數量作為分類標準的主題辭典，在生成結果中獲得平均82.8%的準確率，能夠協助模型生成相對應敘述類型的文本。

以保留商品資訊為重點的實驗中，則是加入採用KeyBERT擷取出的關鍵字組能夠提高相關資訊出現的機率。最後在添加相關與非相關敘述類型的關鍵字組實驗中，結果顯示即使是添加非相關敘述類型的關鍵字組，模型也能考慮其內容並生成相對應的文本，同時也保持一定的敘述類型準確度；而添加相關敘述類型的關鍵字組則能夠提升敘述類型的準確度。

The research aims to explore effective methods for generating product copywriting using pre-trained language models. In this study, we fine-tuned the GPT-2 model by incorporating classification labels, keyword sets as embedding dimensions. This allowed the model to learn how to generate content based on the given input. Experimental results targeting different narrative styles achieved an average accuracy of 82.8% by utilizing topic dictionaries with varying vocabulary sizes as classification criteria, enabling the model to generate text corresponding to the desired narrative style.

In the experiment focusing on preserving product information, the inclusion of keyword phrases extracted using KeyBERT was found to enhance the probability of relevant information generated. Furthermore, in the experiment involving the addition of relevant and non-relevant descriptive keyword, the results demonstrated that the model was capable of considering the content and generating corresponding text even with the presence of non-relevant keywords, while also maintaining a certain level of accuracy in the descriptive style.