Active Learning for Interactive Neural Machine Translation of Data Streams

Abstract

We study the application of active learning techniques to the translation of unbounded data streams via interactive neural machine translation. The main idea is to select, from an unbounded stream of source sentences, those worth to be supervised by a human agent. The user will interactively translate those samples. Once validated, these data is useful for adapting the neural machine translation model.We propose two novel methods for selecting the samples to be validated. We exploit the information from the attention mechanism of a neural machine translation system. Our experiments show that the inclusion of active learning techniques into this pipeline allows to reduce the effort required during the process, while increasing the quality of the translation system. Moreover, it enables to balance the human effort required for achieving a certain translation quality. Moreover, our neural system outperforms classical approaches by a large margin.

Self-Regulated Interactive Sequence-to-Sequence Learning

“…Considering the -greedy-like strategy of the regulator and the strong role of the cost factor shown in Figure 4, the regulator module does not appear to choose individual actions based e.g., on the difficulty of inputs, but rather composes mini-batches with a feedback ratio according to the feedback type's statistics. This confirms the observations of Peris and Casacuberta (2018), who find that the subset of instances selected for labeling is secondaryit is rather the mixing ratio of feedback types that matters. This finding is also consistent with the mini-batch update regime that forces the regulator to take a higher-level perspective and optimize the expected improvement at the granularity of (minibatch) updates rather than at the input level.…”

Data-efficient Neural Text Compression with Interactive Learning

“…A key component of AL is the choice of the sampling strategy, which curates the samples in order to maximize the model's performance with a minimum amount of user interaction. Many AL sampling strategies have proven effective for human-supervised natural language processing tasks other than compression (Hahn et al, 2012;Peris and Casacuberta, 2018;Liu et al, 2018).…”

NMT-Keras: a Very Flexible Toolkit with a Focus on Interactive NMT and Online Learning

“…Following the spirit of the Keras library, we developed NMT-Keras, released under MIT license, that aims to provide a highly-modular and extensible framework to NMT. https://github.com/lvapeab/nmt-keras NMT-Keras supports advanced features, including support of interactive-predictive NMT (INMT) (Barrachina et al, 2009;Peris et al, 2017c) protocols, continuous adaptation (Peris et al, 2017a) and active learning (Peris and Casacuberta, 2018b) strategies. An additional goal, is to ease the usage of the library, but allowing the user to configure most of the options involving the NMT process.…”