What do Neural Machine Translation Models Learn about Morphology?

TL;DR

This study analyzes what neural MT models learn about morphology by using MT representations as features for POS and morphological tagging across morphologically diverse languages. It compares word-based and character-based representations, investigates encoder depth, and examines the influence of target language, revealing that character-based inputs excel at capturing morphology—especially for rare words—while lower encoder layers encode word structure and higher layers encode meaning. The findings show that target language affects the learned representations, and that encoder/decoder representations are of similar quality, with attention mainly shaping encoder outputs. The work provides concrete guidelines for morphology-aware MT design and suggests that integrating morphology into joint training could further improve translation quality and linguistic representational richness.

Abstract

Neural machine translation (MT) models obtain state-of-the-art performance while maintaining a simple, end-to-end architecture. However, little is known about what these models learn about source and target languages during the training process. In this work, we analyze the representations learned by neural MT models at various levels of granularity and empirically evaluate the quality of the representations for learning morphology through extrinsic part-of-speech and morphological tagging tasks. We conduct a thorough investigation along several parameters: word-based vs. character-based representations, depth of the encoding layer, the identity of the target language, and encoder vs. decoder representations. Our data-driven, quantitative evaluation sheds light on important aspects in the neural MT system and its ability to capture word structure.

Figures (8)

• Figure 1: Illustration of our approach: (i) NMT system trained on parallel data; (ii) features extracted from pre-trained model; (iii) classifier trained using the extracted features. Here a POS tagging classifier is trained on features from the first hidden layer.
• Figure 2: POS and morphological tagging accuracy of word-based and character-based models per word frequency in the training data. Best viewed in color.
• Figure 3: Improvement in POS/morphology accuracy of character-based vs. word-based models for words unseen/seen in training, and for all words.
• Figure 4: Increase in POS accuracy with char- vs. word-based representations per tag frequency in the training set; larger bubbles reflect greater gaps.
• Figure 5: Confusion matrices for POS tagging using word-based and character-based representations.