Evaluating Layers of Representation in Neural Machine Translation on Part-of-Speech and Semantic Tagging Tasks

By Yonatan Belinkov, Lluís Màrquez i Villodre, Hassan Sajjad, Nadir Durrani, Fahim Dalvi, James R. Glass. 2017

177 citations recorded by the source corpus.

While neural machine translation (NMT) models provide improved translation quality in an elegant framework, it is less clear what they learn about language. Recent work has started evaluating the quality of vector representations learned by NMT models on morphological and syntactic tasks. In this paper, we investigate the representations learned at different layers of NMT encoders. We train NMT systems on parallel data and use the models to extract features for training a classifier on two tasks: part-of-speech and semantic tagging. We then measure the performance of the classifier as a proxy to the quality of the original NMT model for the given task. Our quantitative analysis yields interesting insights regarding representation learning in NMT models. For instance, we find that higher layers are better at learning semantics while lower layers tend to be better for part-of-speech tagging. We also observe little effect of the target language on source-side representations, especially in higher quality models.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Natural Language Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Natural Language Processing

Citations in the corpus

The Bottom-up Evolution of Representations in the Transformer: A Study with Machine Translation and Language Modeling Objectives (2019, background): “…of linguistic analysis have been considered including morphology (Belinkov et al., 2017a; Dalvi et al., 2017; Bisazza and Tump, 2018), syntax (Shi et al., 2016; Tenney et al., 2019b) and semantics (Hill et al., 2017; Belinkov et al., 2017b; Raganato and Tiedemann, 2018; Tenney et al., 2019b).”
Compositionality Decomposed: How do Neural Networks Generalise? (2019, background): “A robust finding from such analyses is that features such as syntactic constituency, part-of-speech tags and dependency edges can be reliably predicted from the hidden representations of both recurrent neural networks (Belinkov et al., 2017; Blevins, Levy, & Zettlemoyer, 2018; Shi, Padhi, & Knight, ”
Analyzing Multi-Head Self-Attention: Specialized Heads Do the Heavy Lifting, the Rest Can Be Pruned (2019, background): “Different levels of linguistic analysis have been considered including morphology (Belinkov et al., 2017a; Dalvi et al., 2017; Bisazza and Tump, 2018), syntax (Shi et al., 2016) and semantics (Hill et al., 2017; Belinkov et al., 2017b; Raganato and Tiedemann, 2018).”
Probing Classifiers: Promises, Shortcomings, and Advances (2021, background): “7 “evaluate the quality of the trained classiﬁer on the given task as a proxy to the quality of the extracted representations” (Belinkov et al. 2017a).”
Analysis Methods in Neural Language Processing: A Survey (2018, background): “, 2016a) NMT encoder neurons Sentence length Regression (Belinkov et al., 2017b) NMT states POS, lexical semantics Classification (Belinkov et al.”
What do you learn from context? Probing for sentence structure in contextualized word representations (2019, methodology): “Recent work has explored many token-level properties of these representations, such as their ability to capture part-of-speech tags (Blevins et al., 2018; Belinkov et al., 2017b; Shi et al., 2016), morphology (Belinkov et al., 2017a;b), or word-sense disambiguation (Peters et al., 2018a).”
Designing and Interpreting Probes with Control Tasks (2019, methodology): “Applying dropout, the most popular probing regularization method (Adi et al., 2017; Belinkov and Glass, 2019; Şahin et al., 2019; Kim et al., 2019; Elloumi et al., 2018; Belinkov and Glass, 2017; Belinkov et al., 2018) does not consistently lead to high-accuracy, high-selectivity MLP probes across a”
Linguistic Knowledge and Transferability of Contextual Representations (2019, methodology): “We use the dataset of Bjerva et al. (2016); the tagset has since been developed as part of the Parallel Meaning Bank (Abzianidze et al.”
Deep RNNs Encode Soft Hierarchical Syntax (2018, background): “This result conﬁrms the ﬁndings of Shi et al. (2016) and Belinkov et al. (2017b), who demonstrate that neural machine translation encoders learn syntax, and shows that RNNs trained on other NLP tasks also induce syntax.”
Language Modeling Teaches You More than Translation Does: Lessons Learned Through Auxiliary Syntactic Task Analysis (2018, methodology): “Recently, researchers have found that deep LSTMs (Hochreiter and Schmidhuber, 1997) trained on tasks like machine translation learn substantial syntactic and semantic information about their input sentences, including part-of-speech (Belinkov et al., 2017a,b; Blevins et al., 2018).”
The Lazy Encoder: A Fine-Grained Analysis of the Role of Morphology in Neural Machine Translation (2018, methodology): “While this methodology is similar to that of previous work (K¨ohn, 2015; Belinkov et al., 2017a,b; Dalvi et al., 2017) we make sure that our results are not affected by overﬁtting by eliminating any vocabulary overlap between the classiﬁer’s training and test sets.”
Investigating Multilingual NMT Representations at Scale (2019, methodology): “Recent work on interpretability for NLU tasks uses methods such as diagnostic tasks (Belinkov et al., 2017; Tenney et al., 2019; Belinkov et al., 2018), attention based methods (Raganato and Tiedemann, 2018) or task analysis (Zhang and Bowman, 2018) and is primarily focused on understanding the…”
Information-Theoretic Probing for Linguistic Structure (2020, background): “Many speculate that these representations encode a continuous analogue of discrete linguistic properties, e.g., part-of-speech tags, due to the networks’ impressive performance on many NLP tasks (Belinkov et al., 2017).”
Edinburgh Research Explorer Understanding Learning Dynamics Of Language Models with SVCCA (unknown): “Diagnostic classiﬁers are widely used (Belinkov et al., 2018; Giulianelli et al., 2018 examples, then random embeddings as input representations often outperform any pretrained intermediate representation ( ?”