Abstract
We propose a novel dependency-based hybrid tree model for semantic parsing, which converts natural language utterance into machine interpretable meaning representations. Unlike previous state-of-the-art models, the semantic information is interpreted as the latent dependency between the natural language words in our joint representation. Such dependency information can capture the interactions between the semantics and natural language words. We integrate a neural component into our model and propose an efficient dynamic-programming algorithm to perform tractable inference. Through extensive experiments on the standard multilingual GeoQuery dataset with eight languages, we demonstrate that our proposed approach is able to achieve state-of-the-art performance across several languages. Analysis also justifies the effectiveness of using our new dependency-based representation.
Introduction
Semantic parsing is the task of converting natural language (e.g., "What rivers do not run through Tennessee?") into machine-readable logical forms (e.g., answer(exclude(river(all), traverse(stateid('tennessee'))))).
The central challenge in this field is alignment: during training, we usually have the sentence and the logical form, but we do not know exactly which words correspond to which parts of the logic. Previous "hybrid tree" models used constituency-like structures to bridge this gap. However, these models often struggle with "non-isomorphic" structures—cases where the word order of a specific language doesn't neatly map to the nested structure of the logical form. This is especially problematic in multilingual settings where word orders vary significantly.
The Dependency-based Hybrid Tree Model
We propose a joint representation where the meaning representation (MR) is embedded directly into the dependency structure of the sentence. Instead of grouping words into spans, the model treats semantic units as labels on dependency arcs between words.
In this framework, a semantic unit like traverse might be represented as a dependency arc between the words "run" and "through." This approach allows the model to capture the interaction between words more naturally than span-based methods.
Technical Implementation
- Latent-Variable CRF: Since the "true" dependency structure isn't provided in the data, the model treats the tree as a latent variable. It uses a Conditional Random Field (CRF) to marginalize over all possible valid trees during training.
- Dynamic Programming: We developed an efficient $O(N^3M)$ inference algorithm (where $N$ is sentence length and $M$ is the number of semantic units) based on inside-outside style algorithms to make learning and decoding tractable.
- Neural Component: To improve feature learning, we integrated a neural bilinear layer that scores dependency arcs based on word embeddings, capturing the interaction between parent and child words more effectively than discrete features alone.
Key Insights: Flexibility Over Spans
Traditional models, such as the Relaxed Hybrid Tree (RHT), are "span-based." If a semantic unit needs to align with two words separated by other unrelated words, span-based models often fail.
As shown in the comparison above, the dependency approach allows the model to "skip" irrelevant words, making it much more robust to diverse word orders. By treating semantic units as relations between words, the model captures the inherent interaction between linguistic components, which is a more natural fit for how meaning is constructed in language.
Experimental Results
The model, DEPHT (Dependency-based Hybrid Tree), was evaluated on the multilingual GeoQuery dataset, a standard benchmark containing 880 queries about US geography across eight languages: English, German, Greek, Thai, Chinese, Indonesian, Swedish, and Farsi.
- State-of-the-Art: It achieved the highest accuracy/F1 scores on 7 out of 8 languages.
- Indonesian Breakthrough: The most dramatic improvement was seen in Indonesian, where the model outperformed the previous best system (RHT) by 11.8 absolute F1 points.
- Ablation Success: Experiments showed that "Bag of Words" features and the neural component were both critical, with the neural version (DEPHT + NN) consistently providing the best results.
Limitations and Future Work
While powerful, the model currently assumes "projective" trees (no crossing arcs). While rare in the GeoQuery dataset, this could be a hurdle for languages with extremely free word order. Additionally, to handle recursive semantic structures, the model allows "self-loops" on words but imposes a maximum depth to prevent infinite recursion during inference. Future directions include exploring transition-based parsing to handle non-projectivity and applying this dependency-based logic to other formalisms like SQL or Lambda Calculus.
Summary updated