Pre-Learning Environment Representations for Data-Efficient Neural Instruction Following

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1946–1956 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 1946 Pre-Learning Environment Representations for Data-Efficient Neural Instruction Following David Gaddy and Dan Klein Computer Science Division University of California, Berkeley {dgaddy,klein}@berkeley.edu Abstract We consider the problem of learning to map from natural language instructions to state transitions (actions) in a data-efficient man- ner. Our method takes inspiration from the idea that it should be easier to ground language to concepts that have already been formed through pre-linguistic observation. We aug- ment a baseline instruction-following learner with an initial environment-learning phase that uses observations of language-free state transitions to induce a suitable latent rep- resentation of actions before processing the instruction-following training data. We show that mapping to pre-learned representations substantially improves performance over sys- tems whose representations are learned from limited instructional data alone. 1 Introduction In the past several years, neural approaches have become increasingly central to the instruction fol- lowing literature (e.g. Misra et al., 2018; Chap- lot et al., 2018; Mei et al., 2016). However, neu- ral networks’ powerful abilities to induce complex representations have come at the cost of data effi- ciency. Indeed, compared to earlier logical form- based methods, neural networks can sometimes re- quire orders of magnitude more data. The data- hungriness of neural approaches is not surprising – starting with classic logical forms improves data efficiency by presenting a system with pre-made abstractions, where end-to-end neural approaches must do the hard work of inducing abstractions on their own. In this paper, we aim to com- bine the power of neural networks with the data- efficiency of logical forms by pre-learning abstrac- tions in a semi-supervised way, satiating part of the network’s data hunger on cheaper unlabeled data from the environment. When neural nets have only limited data that Figure 1: After seeing this transition, a neural net might generalize this action as stack red blocks to the right of blue blocks except for on brown blocks, but a general- ization like stack red blocks on orange blocks is more plausible and generally applicable. We aim to guide our model towards more plausible generalizations by pre-learning inductive biases from observations of the environment. pairs language with actions, they suffer from a lack of inductive bias, fitting the training data but generalizing in ways that seem nonsensical to hu- mans. For example, a neural network given the transition shown in Figure 1 might map the cor- responding instruction to an adequate but unlikely meaning that red blocks should be stacked to the right of blue blocks except for on brown blocks. The inspiration for this work comes from the idea that humans avoid spurious hypotheses like this example partly because they have already formed a set of useful concepts about their environment before learning language (Bloom, 2000; Hespos and Spelke, 2004). These pre-linguistic abstrac- tions then constrain language learning and help generalization. With this view in mind, …

Content truncated to 500 words