Compositionality decomposed: how do neural networks generalise?

Abstract

Despite a multitude of empirical studies, little consensus exists on whether neural networks are able to generalise compositionally, a controversy that, in part, stems from a lack of agreement about what it means for a neural model to be compositional. As a response to this controversy, we present a set of tests that provide a bridge between, on the one hand, the vast amount of linguistic and philosophical theory about compositionality of language and, on the other, the successful neural models of language. We collect different interpretations of compositionality and translate them into five theoretically grounded tests for models that are formulated on a task-independent level. In particular, we provide tests to investigate (i) if models systematically recombine known parts and rules (ii) if models can extend their predictions beyond the length they have seen in the training data (iii) if models' composition operations are local or global (iv) if models' predictions are robust to synonym substitutions and (v) if models favour rules or exceptions during training. To demonstrate the usefulness of this evaluation paradigm, we instantiate these five tests on a highly compositional data set which we dub PCFG SET and apply the resulting tests to three popular sequence-to-sequence models: a recurrent, a convolution-based and a transformer model. We provide an in-depth analysis of the results, which uncover the strengths and weaknesses of these three architectures and point to potential areas of improvement.

References in corpus (22)

• Sequence to Sequence Learning with Neural Networks
• Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling
• An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling
• From Frequency to Meaning: Vector Space Models of Semantics
• Convolutional Sequence to Sequence Learning
• Exploring the Limits of Language Modeling
• A Convolutional Neural Network for Modelling Sentences
• Assessing BERT's Syntactic Abilities
• Extracting Automata from Recurrent Neural Networks Using Queries and Counterexamples
• Modelling, Visualising and Summarising Documents with a Single Convolutional Neural Network
• Character-Level Question Answering with Attention
• Analysing Mathematical Reasoning Abilities of Neural Models
• Memorize or generalize? Searching for a compositional RNN in a haystack
• BERT Rediscovers the Classical NLP Pipeline
• The emergence of number and syntax units in LSTM language models
• RNNs Implicitly Implement Tensor Product Representations
• Discovering the Compositional Structure of Vector Representations with Role Learning Networks
• Measuring Compositionality in Representation Learning
• Systematic Generalization: What Is Required and Can It Be Learned?
• Learning compositionally through attentive guidance
• Rearranging the Familiar: Testing Compositional Generalization in Recurrent Networks
• The Importance of Being Recurrent for Modeling Hierarchical Structure

Cited by in corpus (10)

• Scaling Laws for Autoregressive Generative Modeling
• CLOSURE: Assessing Systematic Generalization of CLEVR Models
• From Machine Learning to Robotics: Challenges and Opportunities for Embodied Intelligence
• On Compositionality in Neural Machine Translation
• Structured Reordering for Modeling Latent Alignments in Sequence Transduction
• Can RNNs learn Recursive Nested Subject-Verb Agreements?
• Do Neural Models Learn Systematicity of Monotonicity Inference in Natural Language?
• EXplainable Neural-Symbolic Learning (X-NeSyL) methodology to fuse deep learning representations with expert knowledge graphs: the MonuMAI cultural heritage use case
• Word Interdependence Exposes How LSTMs Compose Representations
• Relational Weight Priors in Neural Networks for Abstract Pattern Learning and Language Modelling