[Paper Review] Modularity Improves Out-of-Domain Instruction Following.

Rodolfo Corona, Daniel Fried+3 more|arXiv (Cornell University)|Oct 24, 2020

Topic Modeling28 references4 citations

TL;DR

This paper proposes a modular architecture for following natural language instructions by decomposing tasks into subgoal types, where each subgoal is handled by a dedicated module. By learning to segment instructions and predict subgoal types, the model improves zero-shot generalization to out-of-distribution environments and novel tasks on the ALFRED benchmark compared to standard sequence-to-sequence models.

ABSTRACT

We propose a modular architecture for following natural language instructions that describe sequences of diverse subgoals, such as navigating to landmarks or picking up objects. Standard, non-modular, architectures used in instruction following do not exploit subgoal compositionality and often struggle on out-of-distribution tasks and environments. In our approach, subgoal modules each carry out natural language instructions for a specific subgoal type. A sequence of modules to execute is chosen by learning to segment the instructions and predicting a subgoal type for each segment. When compared to standard sequence-to-sequence approaches on ALFRED, a challenging instruction following benchmark, we find that modularization improves generalization to environments unseen in training and to novel tasks.

Motivation & Objective

• To address the poor generalization of non-modular instruction-following models on unseen environments and novel tasks.
• To exploit the compositional structure of natural language instructions by decomposing them into reusable subgoal types.
• To improve zero-shot generalization in instruction-following by learning to segment instructions and predict subgoal types.
• To demonstrate that modular architectures outperform standard sequence-to-sequence models on out-of-distribution tasks in the ALFRED benchmark.

Proposed method

• The model uses a modular architecture where each subgoal type (e.g., navigate to a landmark, pick up an object) is handled by a dedicated neural module.
• Instruction segmentation is performed by a segmentation head that divides input instructions into subgoal segments.
• A subgoal type predictor assigns a specific subgoal type to each segmented instruction segment.
• Each subgoal module executes its corresponding task using its own policy, enabling compositional execution.
• The system is trained to optimize instruction-following performance on the ALFRED benchmark using end-to-end learning.
• The modular design allows for better zero-shot transfer to unseen environments and novel task combinations.

Experimental results

Research questions

• RQ1 Can a modular architecture improve generalization to out-of-distribution environments in instruction-following tasks?
• RQ2 Does subgoal compositionality in natural language instructions improve zero-shot transfer to novel tasks?
• RQ3 How does modular instruction following compare to standard sequence-to-sequence models in terms of out-of-domain performance?
• RQ4 Can learning to segment instructions and predict subgoal types lead to better generalization than end-to-end modeling?
• RQ5 To what extent does modularity enhance transfer to unseen environments and task combinations?

Key findings

• The modular architecture achieves superior generalization to environments not seen during training compared to standard sequence-to-sequence models.
• The model generalizes effectively to novel tasks that combine subgoals in ways not seen during training.
• Improved performance is observed in zero-shot evaluation on out-of-distribution environments in the ALFRED benchmark.
• The approach leverages subgoal compositionality, enabling better transfer than non-modular alternatives.
• Segmentation and subgoal type prediction are critical components that enable the model to generalize across diverse instruction types.
• The modular design leads to more robust and interpretable instruction following in complex, unseen environments.