Semantic parsing—the task of mapping natural language to structured representations like Abstract Meaning Representation (AMR), Semantic Role Labeling (SRL), or SQL—suffers from a severe "annotation bottleneck." While high-resource languages like English benefit from massive, human-curated datasets, creating similar resources for the world’s other 7,000 languages is prohibitively expensive. As Rexhina Blloshmi, Rocco Tripodi, and Roberto Navigli note in XL-AMR: Enabling Cross-Lingual AMR Parsing with Transfer Learning Techniques, this data gap prevents the deployment of intuitive database interfaces and meaning-aware systems globally.
To overcome this, researchers have developed techniques to transfer semantic knowledge from resource-rich languages to resource-poor ones, effectively "recycling" English annotations to train models in other languages.
Projecting Semantics Across the Language Gap
The most direct way to bypass the annotation bottleneck is annotation projection. If we have an English sentence paired with a semantic graph, and a translation of that sentence in Italian, we can attempt to "project" the semantic labels from the English words to their Italian counterparts.
Marco Damonte and Shay B Cohen demonstrated in Cross-Lingual Abstract Meaning Representation Parsing that English-centric formalisms like AMR can serve as a universal semantic backbone. By using machine translation and word alignment, they built parsers for languages like Chinese and German without requiring native gold-standard data.
However, simple projection is often noisy. To improve quality, Leonardo F. R. Ribeiro et al. proposed "smelting" in Smelting Gold and Silver for Improved Multilingual AMR-to-Text Generation, combining high-quality "gold" English data with "silver" synthetic data—sentences generated from semantic graphs—to help models learn the complex mappings between abstract meaning and diverse linguistic expressions.
Aligning Languages in a Shared Latent Space
Rather than relying on explicit word-to-word translation, a more robust approach involves forcing the model to learn a language-agnostic representation. The goal is to map a Chinese question and its English equivalent to the same point in a "latent semantic space," from which a single decoder can produce the correct SQL or AMR graph.
Mirella Lapata and Tom Sherborne pioneered this in Zero-Shot Cross-lingual Semantic Parsing with the ZX-PARSE framework. By using language-adversarial objectives, they trained models to ignore the specific language of the input and focus solely on the underlying meaning. Similarly, Bo Shao et al. introduced MASP in Multi-level Alignment Pretraining for Multi-lingual Semantic Parsing, which aligns languages at the vector, word, and sentence levels simultaneously, leveraging "cheap" data like dictionaries to bridge the gap.
Leveraging Universal Structural Pivots
If the surface forms of languages are too different, researchers look to structural invariants. Universal Dependencies (UD) and syntax trees provide a bridge because, while words change, the relationship between a verb and its subject is often structurally similar across languages.
Jingfeng Yang 0001 et al. showed in Frustratingly Simple but Surprisingly Strong: Using Language-Independent Features for Zero-shot Cross-lingual Semantic Parsing that simply adding UPOS tags and UD features to a model allows it to outperform much larger, computationally expensive models in zero-shot settings. For graph-based representations, Jonas Groschwitz proposed in Normalizing Compositional Structures Across Graphbanks a mathematical framework to normalize different semantic graphbanks, allowing models to leverage the collective information of disparate datasets.
The Role of In-Context Learning and Prompting
With the advent of Large Language Models (LLMs), the focus has shifted toward In-Context Learning (ICL), where the model learns to parse a new language by seeing a few examples in its prompt.
However, low-resource languages still struggle because the models have less pre-existing knowledge of their grammar. Peng Shi et al. addressed this in XRICL: Cross-lingual Retrieval-Augmented In-Context Learning for Cross-lingual Text-to-SQL Semantic Parsing by using a "translation as chain-of-thought" strategy. By asking the model to first translate a query into English before parsing it into SQL, they achieved state-of-the-art results without any task-specific fine-tuning. To further refine this, Mayank Kothyari et al. developed SCUD4ICL in Diverse In-Context Example Selection After Decomposing Programs and Aligned Utterances Improves Semantic Parsing, which breaks down complex examples into smaller, cleaner fragments to reduce "interference" and help the model focus on the core logical mapping.
By combining these strategies—projection, shared latent spaces, structural pivots, and clever prompting—the research community is steadily dismantling the semantic annotation bottleneck, moving toward a future where natural language interfaces are accessible in every language.
Key Foundational Work
Word Sense Disambiguation Incorporating Lexical and Structural Semantic Information (2007) — contributed 1 concept elements to this direction.
Shallow Semantic Parsing of Chinese (2004) — contributed 1 concept elements to this direction.