Learn to Use Future Information in Simultaneous Translation

Xueqing Wu, Yingce Xia, Lijun Wu, Shufang Xie, Weiqing Liu, Jiang Bian, Tao Qin, Tie-Yan Liu

Abstract

Sequence learning has attracted much research attention from the machine learning community in recent years. In many applications, a sequence learning task is usually associated with multiple temporally correlated auxiliary tasks, which are different in terms of how much input information to use or which future step to predict. For example, in simultaneous machine translation, one can conduct translation under different latency constraints. While it is clear that those temporally correlated tasks can help each other, there is a very limited exploration on how to better leverage multiple auxiliary tasks to boost the performance of the main task. In this work, we introduce a learnable scheduler to sequence learning, which can adaptively select auxiliary tasks for training depending on the model status and the current training data. The scheduler and the model for the main task are jointly trained through bi-level optimization. Experiments show that our method significantly improves the performance of simultaneous machine translation and stock trend forecasting.

Introduction

In sequence learning, research often focuses on optimizing a single "main task." For instance, in simultaneous machine translation, the objective might be a "wait-3" task, where the model translates while looking only three words ahead. However, this task is surrounded by a family of temporally correlated tasks—such as wait-1, wait-5, or full-sentence translation—that share the same input/output space but differ in their latency or access to future information.

While Multi-Task Learning (MTL) suggests that training on these related tasks can be beneficial, standard approaches often treat all tasks equally or follow a rigid Curriculum Learning (CL) schedule. The core challenge is that the utility of specific auxiliary tasks varies throughout the training process. This paper addresses this by proposing an intelligent "meta-controller" that learns to dynamically select the most beneficial tasks to maximize performance on the primary goal.

The TCTS Framework

The authors introduce Temporally Correlated Task Scheduling (TCTS), a framework consisting of two primary components:

  1. The Model ($f$): The core sequence learner (e.g., a Transformer for translation or a GRU for stock forecasting).
  2. The Scheduler ($\phi$): A lightweight neural network that acts as a teacher. It monitors the model's current state—including training loss, validation performance, and training step—to decide which auxiliary task the model should prioritize next.
The TCTS framework architecture illustrating the bi-level optimization between the model parameters (theta) and the learnable scheduler (omega).

The system utilizes a bi-level optimization strategy. In the inner loop, the model is trained to minimize loss on tasks selected by the scheduler. In the outer loop, the scheduler is optimized via Reinforcement Learning (REINFORCE), receiving rewards based on the model's performance on a validation set specifically for the main task. This ensures the scheduler learns to pick tasks that directly improve the final objective.

Experimental Setup

The TCTS framework was evaluated across two distinct sequence learning domains:

  • Simultaneous Machine Translation: Benchmarked on four major datasets: IWSLT’14 (En→De), IWSLT’15 (En→Vi), IWSLT’17 (En→Zh), and WMT’15 (En→De).
  • Stock Trend Forecasting: Evaluated using historical transaction data for 300 stocks from the CSI300 index (Chinese A-share market) from 2008 to 2020.

Key Results and Analysis

TCTS demonstrated consistent improvements over standard baselines. In translation tasks, it improved BLEU scores by 1 to 3 points across various languages and latency settings. In stock forecasting, the method achieved superior RankIC (correlation between predicted and actual returns) and lower Mean Square Error compared to standard GRUs and even complex Graph Attention Networks (GAT).

Performance comparison of TCTS (Ours) against baselines (Wait-k, CL, MTL) across different language pairs and latency metrics (AP and AL).

Insights into Task Scheduling

Analysis of the scheduler's behavior revealed several key insights:

  • Dynamic Adaptation: The scheduler does not favor a single task. It typically begins with uniform sampling to explore the task space before converging on a specific mix that provides optimal "noised augmentation" for the main task.
  • The Value of "Easier" Tasks: For a low-latency "wait-3" translation task, the scheduler often preferred training on "wait-12" tasks. This suggests that learning high-quality representations from tasks with more future information helps the model perform better even when that information is restricted during inference.
  • Efficiency: By implementing "incremental encoding" (unidirectional attention), the authors achieved inference speeds 8x faster than previous state-of-the-art models with negligible accuracy loss.

Conclusion

While TCTS introduces a training overhead of approximately 20% to 30% due to the scheduler's computations, the performance gains are significant and more robust than traditional Curriculum Learning. Future work may explore dynamic latency trade-offs during training and apply the framework to other domains such as weather forecasting or streaming speech recognition.

Summary updated

Abstract

Sequence learning has attracted much research attention from the machine learning community in recent years. In many applications, a sequence learning task is usually associated with multiple temporally correlated auxiliary tasks, which are different in terms of how much input information to use or which future step to predict. For example, (i) in simultaneous machine translation, one can conduct translation under different latency (i.e., how many input words to read/wait before translation); (ii) in stock trend forecasting, one can predict the price of a stock in different future days (e.g., tomorrow, the day after tomorrow). While it is clear that those temporally correlated tasks can help each other, there is a very limited exploration on how to better leverage multiple auxiliary tasks to boost the performance of the main task. In this work, we introduce a learnable scheduler to sequence learning, which can adaptively select auxiliary tasks for training depending on the model status and the current training data. The scheduler and the model for the main task are jointly trained through bi-level optimization. Experiments show that our method significantly improves the performance of simultaneous machine translation and stock trend forecasting.
BibTeX
@inproceedings{wu2021learn,
  author = {Xueqing Wu and Yingce Xia and Lijun Wu and Shufang Xie and Weiqing Liu and Jiang Bian and Tao Qin and Tie-Yan Liu},
  title = {{Learn to Use Future Information in Simultaneous Translation}},
  booktitle = {International Conference on Machine Learning},
  year = {2021},
  url = {https://openreview.net/forum?id=FynX8He2_uL},
}
Figure