We introduce a curriculum learning approach to adapt generic neural machine translation models to a specific domain. Samples are grouped by their similarities to the domain of interest and each group is fed to the training algorithm with a particular schedule. This approach is simple to implement on top of any neural framework or architecture, and consistently outperforms both unadapted and adapted baselines in experiments with two distinct domains and two language pairs.
When translating diglossic languages such as Arabic, situations may arise where we would like to translate a text but do not know which dialect it is. A traditional approach to this problem is to design dialect identification systems and dialect-specific machine translation systems. However, under the recent paradigm of neural machine translation, shared multi-dialectal systems have become a natural alternative. Here we explore under which conditions it is beneficial to perform dialect identification for Arabic neural machine translation versus using a general system for all dialects.
Our submission to the MADAR shared task on Arabic dialect identification employed a language modeling technique called Prediction by Partial Matching, an ensemble of neural architectures, and sources of additional data for training word embeddings and auxiliary language models. We found several of these techniques provided small boosts in performance, though a simple character-level language model was a strong baseline, and a lower-order LM achieved best performance on Subtask 2. Interestingly, word embeddings provided no consistent benefit, and ensembling struggled to outperform the best component submodel. This suggests the variety of architectures are learning redundant information, and future work may focus on encouraging decorrelated learning.
Neural machine translation has achieved impressive results in the last few years, but its success has been limited to settings with large amounts of parallel data. One way to improve NMT for lower-resource settings is to initialize a word-based NMT model with pretrained word embeddings. However, rare words still suffer from lower quality word embeddings when trained with standard word-level objectives. We introduce word embeddings that utilize morphological resources, and compare to purely unsupervised alternatives. We work with Arabic, a morphologically rich language with available linguistic resources, and perform Ar-to-En MT experiments on a small corpus of TED subtitles. We find that word embeddings utilizing subword information consistently outperform standard word embeddings on a word similarity task and as initialization of the source word embeddings in a low-resource NMT system.
Character-level string-to-string transduction is an important component of various NLP tasks. The goal is to map an input string to an output string, where the strings may be of different lengths and have characters taken from different alphabets. Recent approaches have used sequence-to-sequence models with an attention mechanism to learn which parts of the input string the model should focus on during the generation of the output string. Both soft attention and hard monotonic attention have been used, but hard non-monotonic attention has only been used in other sequence modeling tasks and has required a stochastic approximation to compute the gradient. In this work, we introduce an exact, polynomial-time algorithm for marginalizing over the exponential number of non-monotonic alignments between two strings, showing that hard attention models can be viewed as neural reparameterizations of the classical IBM Model 1. We compare soft and hard non-monotonic attention experimentally and find that the exact algorithm significantly improves performance over the stochastic approximation and outperforms soft attention.