Section: New Results

Evaluation of AI algorithms

Machine learning algorithms are typically evaluated in terms of end-to-end tasks, but it is very often difficult to get a grasp of how they achieve these tasks, what could be their break point, and more generally, how they would compare to the algorithms used by humans to do the same tasks. This is especially true of Deep Learning systems which are particularly opaque. The team develops evaluation methods based on psycholinguistic/linguistic criteria, and deploy them for systematic comparison of systems.

• Recurrent neural networks (RNNs) can learn continuous vector representations of symbolic structures such as sequences and sentences; these representations often exhibit linear regularities (analogies). Such regularities motivate our hypothesis that RNNs that show such regularities implicitly compile symbolic structures into tensor product representations (TPRs; Smolensky, 1990), which additively combine tensor products of vectors representing roles (e.g., sequence positions) and vectors representing fillers (e.g., particular words). To test this hypothesis, we introduce Tensor Product Decomposition Networks (TPDNs), which use TPRs to approximate existing vector representations. We demonstrate using synthetic data that TPDNs can successfully approximate linear and tree-based RNN autoencoder representations, suggesting that these representations exhibit interpretable compositional structure; we explore the settings that lead RNNs to induce such structure-sensitive representations. By contrast, further TPDN experiments show that the representations of four models trained to encode naturally-occurring sentences can be largely approximated with a bag of words, with only marginal improvements from more sophisticated structures. We conclude that TPDNs provide a powerful method for interpreting vector representations, and that standard RNNs can induce compositional sequence representations that are remarkably well approximated by TPRs; at the same time, existing training tasks for sentence representation learning may not be sufficient for inducing robust structural representations.

• LSTMs have proven very successful at language modeling. However, it remains unclear to what extent they are able to capture complex morphosyntactic structures. In [29], we examine whether LSTMs are sensitive to verb argument structures. We introduce a German grammaticality dataset in which ungrammatical sentences are constructed by manipulating case assignments (eg substituting nominative by accusative or dative). We find that LSTMs are better than chance in detecting incorrect argument structures and slightly worse than humans tested on the same dataset. Surprisingly, LSTMs are contaminated by heuristics not found in humans like a preference toward nominative noun phrases. In other respects they show human-similar results like biases for particular orders of case assignments.

• Pater (2019) proposes to use neural networks to model learning within existing grammatical frameworks. In [16] we argue that there is a fundamental gap to be bridged that does not receive enough attention : how can we use neural networks to examine whether it is possible to learn some linguistic representation (a tree, for example) when, after learning is finished, we cannot even tell if this is the type of representation that has been learned (all we see is a sequence of numbers)? Drawing a correspondence between an abstract linguistic representational system and an opaque parameter vector that can (or perhaps cannot) be seen as an instance of such a representation is an implementational mapping problem. Rather than relying on existing frameworks that propose partial solutions to this problem, such as harmonic grammar, we suggest that fusional research of the kind proposed needs to directly address how to ‘find’ linguistic representations in neural network representations.