1: \begin{abstract}

2: The language-independency of encoded representations within multilingual neural machine translation (MNMT) models is crucial for their generalization ability on zero-shot translation. Neural interlingua representations have been shown as an effective method for achieving this. However, fixed-length neural interlingua representations introduced in previous work can limit its flexibility and representation ability. In this study, we introduce a novel method to enhance neural interlingua representations by making their length variable, thereby overcoming the constraint of fixed-length neural interlingua representations. Our empirical results on zero-shot translation on OPUS, IWSLT, and Europarl datasets demonstrate stable model convergence and superior zero-shot translation results compared to fixed-length neural interlingua representations. However, our analysis reveals the suboptimal efficacy of our approach in translating from certain source languages, wherein we pinpoint the defective model component in our proposed method.

3: %Variable length neural interlingua can serve as flexible and implicit pivots for zero-shot translation in the MNMT model, potentially enhancing the representation ability of fixed-length neural interlingua across different languages.

4: % However, we observe the inferior performance of our method when translating from particular source languages, for which we identify the problematic model component in our proposed method.

5: % However, we find that the inaccurate length of neural interlingua during inference hinders performance for non-English source languages, for which we identify the problematic model component.

6: \end{abstract}

7: