基於變換隱空間之風格化對話生成;Transfer Latent Spaces for Stylized Dialogue Generation

NCUIR > School of Management at National Central University > Graduate Institute of Information Management > Electronic Thesis & Dissertation > Item 987654321/86674

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/86674

Title:	基於變換隱空間之風格化對話生成;Transfer Latent Spaces for Stylized Dialogue Generation
Authors:	陳威良;Chen, Wei-Liang
Contributors:	資訊管理學系
Keywords:	對話生成;文字風格轉換;深度學習;多任務學習;Dialogue generation;text style transfer;deep neural network;multi-task learning
Date:	2021-08-23
Issue Date:	2021-12-07 13:06:24 (UTC+8)
Publisher:	國立中央大學
Abstract:	現今對話生成技術已經展現出強大的潛力，不過現今的對話系統通產出的回覆通常平淡且一般。直接讓對話系統產出附有風格的回覆，是一個讓對話系統能夠生成多元化回覆的解決方法。在這篇論文中我們提出附帶風格的對話生成方法，在生成對話的同時做風格轉換，可以達到一個問句有多種風格的回覆，目的是為了讓機器可以在不同對話場景用合適的風格給予回應，這項任務也可以說是有效得結合對話生成任務及風格轉換任務，於是我們不僅重視回覆句能夠適當回覆，更強調能夠展現出強的烈風格。由於資料特性的關係，對話資料集通常是成對的資料，一個上文有對應的下文，而風格文本的資料集通常是不成對的，於是我們利用監督式學習的方式建構對話模型，用非監督式學習的方式建構風格轉換模型，並且使兩個模型共用一個解譯器，組合成一個多工模型。我們提出使用輕量級的深度神經網路模型，將對話生成模型的潛在空間橋接到附帶風格的潛在空間，使模型插入不同風格的深度神經網路模型，就能生成對應風格的回覆，讓我們提出的模型生成出特別生動且令人印象深刻的回覆。文中我們展示成功將對話潛在空間橋接到風格轉換的潛在空間的成果，並且我們利用多個自動化評估指標搭配人工評估，多方面比較我們提出的模型與基準模型的成效，結果指出，我們的模型能夠生成附帶強烈風格的回覆，風格強度優於基準模型許多，語句的流暢度也優於基準模型，並且維持與基準模型相當的回覆適當性。不僅如此我們希望我的的模型能有廣泛的適用性，我們利用兩個外部的對話資料集，測試我們模型的適用性，結果顯示用於日常對話的文本有不錯的適用性。;Dialogue generation technology has shown great potential, but currently dialogue systems usually generate plain and general responses. Directly allowing the dialogue system to produce stylized responses a solution that allow the dialogue systems to generate diversified responses. In this study, we propose a dialogue generation method with styles. While generating dialogues, we can do style conversion to achieve multiple styles of responses to a question. The purpose is to allow the machine to respond with appropriate styles in different dialogue scenarios. This task can also be said to be an effective combination of dialogue generation tasks and style conversion tasks, so we not only attach importance to reply sentences to be able to respond appropriately, but also emphasize the ability to reply with high style intensity. With the data characteristics, the dialogue dataset is usually parallel data, one context has a corresponding response, and the style text dataset is usually non-parallel, so we use supervised learning to construct the dialogue generation model, using unsupervised learning method constructs a style transfer model, and makes the two models shared a decoder and combine them into a hybrid model. We propose using lightweight deep neural network models to bridge the latent spaces of dialogue response generation model and style transfer model. This structure allows the model to generate many different impressive style sentences. In chapter four, we show the results of successfully bridging the latent space of dialogue generation to the latent space of style transfer, and we use multiple auto evaluation metrics and human evaluation to compare the effectiveness of our proposed model with the benchmark model in many aspects. The results indicate that the style intensity and the fluency of sentences are much better than that of the benchmark model, and the appropriateness of the responses is maintained comparable to that of the benchmark model. Not only that, we use two external dialogue datasets to test the applicability of our model. The results show that the text used for daily dialogue has good applicability.
Appears in Collections:	[Graduate Institute of Information Management] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	80	View/Open

社群 sharing

Loading...