We formalize human language understanding as a structured prediction task where the output is a partially ordered set (poset). Current encoder-decoder architectures do not take the poset structure of semantics into account properly, thus suffering from poor compositional generalization ability. In this paper, we propose a novel hierarchical poset decoding paradigm for compositional generalization in language. Intuitively: (1) the proposed paradigm enforces partial permutation invariance in semantics, thus avoiding overfitting to bias ordering information; (2) the hierarchical mechanism allows to capture high-level structures of posets. We evaluate our proposed decoder on Compositional Freebase Questions (CFQ), a large and realistic natural language question answering dataset that is specifically designed to measure compositional generalization. Results show that it outperforms current decoders.

Weaknesses: First, I feel that their poset decoding framework is not really a general-purpose approach to achieve compositional generalization in language, but is kind of specialized for decoding logical forms that include a set of conjunctive/disjunctive clauses, especially for generating SparQL or SQL queries. In this space, some existing work already proposes techniques for synthesizing unordered SQL clauses, e.g., [1][2]. In particular, although [2] does not consider compositional generalization, the hierarchical decoding process in this paper shares some high-level similarities with the sketch-based query synthesis approach in [2]. This paper lacks a discussion of related work for natural language to SQL synthesis. Second, though the results on CFQ are impressive, some important details are missing.

The CFQ dataset designed for testing compositional generalization is really challenging. The presented results in this paper on CFQ are impressive. However, as pointed out by Reviewer #4, the proposed method relies on many dataset-specific design, and the technical novelty is incremental compared to prior work on semantic parsing. The work will be much more convincing if it can also be validated on another dataset.

We formalize human language understanding as a structured prediction task where the output is a partially ordered set (poset). Current encoder-decoder architectures do not take the poset structure of semantics into account properly, thus suffering from poor compositional generalization ability. In this paper, we propose a novel hierarchical poset decoding paradigm for compositional generalization in language. Intuitively: (1) the proposed paradigm enforces partial permutation invariance in semantics, thus avoiding overfitting to bias ordering information; (2) the hierarchical mechanism allows to capture high-level structures of posets. We evaluate our proposed decoder on Compositional Freebase Questions (CFQ), a large and realistic natural language question answering dataset that is specifically designed to measure compositional generalization.