跨語言文本摘要是透過機器將一種語言的文章轉換成另 一種語言的摘要，先前的研究大多將該任務以兩步驟方法處 理──「先翻譯後摘要」或「先摘要後翻譯」。但是，這兩 種方法皆會有翻譯錯誤的問題，且其中的機器翻譯模型難以 隨著摘要任務繼續更新微調（fine-tune）。針對上述問題， 我們採用預訓練跨語言編碼器以向量表示（represent）不同 語言的輸入，將其映射至相同的向量空間。預訓練方法已被 廣泛應用在各種自然語言生成任務中，並取得優異的模型表 現。此編碼器使得模型在學習摘要能力的過程中，同時保有 跨語言能力。本研究中，我們實驗三種不同的微調方法， 證明預訓練跨語言編碼器可以學習單詞階層（word-level） 的語意特徵。在我們所有的模型組態裡，最優異的模型可 在ROUGE-1分數上，超越基準模型3分。 ;Cross-lingual text summarization (CLTS) is the task to generate a summary in one language given a document in a another language. Most of the previous work consider CLTS as two sub-tasks: translate-then-summarize and summarize-then-translate. Both of them are suffered from translation error and the translation system is hard to be fine-tuned with text summarization directly. To deal with the above problems, we utilize a pretrained cross-lingual encoder, which has been demonstrated the effectiveness in natural language generation, to represent text inputs from from different languages. We augment a standard sequence-to-sequence (Seq2Seq) network with our pretrained cross-lingual encoder so as to capture cross-lingual contextualized word representation. We show that the pretrained cross-lingual encoder can be fine-tuned on a text summarization dataset while keeping the cross-lingual ability. We experiment three different fine-tune strategies and show that the pretrained encoder can capture cross-lingual semantic features. The best of the proposed models obtains 42.08 Rouge-1 on ZH2ENSUM datasets [Zhu et al., 2019], significantly improving our baseline model by more than 3 Rouge-1.
