Abstract
Existing studies on semantic parsing mainly focus on the in-domain setting. We formulate cross-domain semantic parsing as a domain adaptation problem: train a semantic parser on some source domains and then adapt it to the target domain. Due to the diversity of logical forms in different domains, this problem presents unique and intriguing challenges. By converting logical forms into canonical utterances in natural language, we reduce semantic parsing to paraphrasing, and develop an attentive sequence-to-sequence paraphrase model that is general and flexible to adapt to different domains. We discover two problems, small micro variance and large macro variance, of pre-trained word embeddings that hinder their direct use in neural networks, and propose standardization techniques as a remedy. On the popular Overnight dataset, which contains eight domains, we show that both cross-domain training and standardized pre-trained word embeddings can bring significant improvement.
Introduction
Semantic parsing—the task of mapping natural language utterances into machine-readable logical forms—is a cornerstone of natural language interfaces. However, traditional systems are often "in-domain," requiring retraining for every new application area. This paper, presented at EMNLP 2017, addresses the "symbol grounding problem": the difficulty of transferring knowledge when logical predicates change entirely between domains (e.g., from basketball to social media).
Methodology: Parsing as Paraphrasing
The authors propose a framework that treats semantic parsing as a paraphrasing task. Instead of mapping a user's input directly to a rigid logical form, they use a deterministic grammar to generate "canonical utterances." These are simple, natural language sentences that represent the logical form.
By transforming the problem into paraphrasing (translating natural input to canonical output), the model can leverage general linguistic structures that are shared across domains. The core architecture is an attentive sequence-to-sequence (Seq2Seq) model using Gated Recurrent Units (GRUs).
Embedding Standardization
A key contribution of this work is identifying why pre-trained word embeddings (like Word2Vec) often fail to improve neural semantic parsers. The authors identify two issues:
- Small Micro Variance: Individual word vectors have low variance, which can slow down neural network optimization.
- Large Macro Variance: High variance across the entire vocabulary hinders generalization to new, unseen words.
They introduce unit variance standardization to align these embeddings with standard neural network initialization (e.g., Xavier initialization), making them much more compatible with the learning process.
Key Results
The model was evaluated on the OVERNIGHT dataset, which spans eight distinct domains including Basketball, Social, and Recipes. The results demonstrate that natural language serves as an effective universal bridge for cross-domain transfer.
Key findings include:
- State-of-the-Art Performance: The model achieved a new best-in-domain average accuracy of 78.2%, with the best configuration reaching 80.6%, outperforming previous methods across most domains.
- Cross-Domain Transfer: Training on source domains consistently boosted performance on target domains, particularly when target data was scarce. For instance, in the "Social" domain, the model achieved near-peak accuracy using only 30% of the target-domain data.
- Standardization Impact: Standardizing Word2Vec embeddings led to an approximate 10% absolute improvement in accuracy compared to using raw embeddings or standard normalization.
Conclusion
This research highlights that while logical symbols are domain-specific, the underlying linguistic patterns used to describe them are relatively universal. By reducing semantic parsing to paraphrasing and addressing the statistical mismatches in pre-trained embeddings, the authors provide a robust path forward for scalable, cross-domain natural language interfaces. Future work will explore scaling these methods to open-domain knowledge bases and transferring across different formal query languages like SQL.
Summary updated