Semantics-Guided Data Augmentation for AMR Parsing

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing techniques for domain adaptation of Abstract Meaning Representation (AMR) parsers rely on general, domain-neutral data augmentation, leading to reduced performance, particularly when adapting to new domains with unique vocabulary and semantic patterns, as they require large, unannotated datasets and manual curation, which is expensive and time-consuming.

Innovation Solution

The implementation of a semantics-guided domain-specific data augmentation method that uses a domain lexicon legend to apply composition and substitution operations on existing annotated training datasets, leveraging unique domain-specific vocabulary and semantic patterns to automatically generate new training data, thereby enhancing the parser's performance in specific domains.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If general domain-neutral data augmentation is used, then the parser can be trained on larger datasets, but the performance deteriorates when adapting to new domains with unique vocabulary and semantic patterns

Engineering Contradiction:

Improvedataset sizeVSAvoidparser performance

Core Design Contradiction:

Quantity of substanceVSReliability

Solution Approach 1:

The patent applies local quality by transforming generic domain-neutral augmentation into domain-specific augmentation. The system uses a domain lexicon legend to identify domain-specific vocabulary and semantic patterns, then applies composition and substitution operations tailored to that specific domain. This ensures the augmented data maintains domain-specific characteristics rather than applying uniform transformations across all domains.

Inventive Principle:

Principle #3Local quality

Solution Approach 2:

The patent changes the parameters of data augmentation by introducing semantics-guided composition and substitution operations. Instead of random or generic transformations, the system modifies the semantic structure of AMR graphs by composing subgraphs based on domain lexicon relationships and substituting equivalent semantic units. This transforms the augmentation process from parameter-agnostic to parameter-specific, improving domain adaptation performance.

Inventive Principle:

Principle #35Parameter changes

2Manufacturing precision

If manual curation of domain-specific training data is performed, then the training data quality improves, but the time and cost increase significantly

Engineering Contradiction:

Improvetraining data qualityVSAvoiddata curation time

Core Design Contradiction:

Manufacturing precisionVSLoss of time

Solution Approach 1:

The patent implements self-service by enabling the system to automatically generate domain-specific training data without human intervention. The domain lexicon legend serves as a self-contained knowledge base that guides the composition and substitution operations. The system autonomously identifies domain-specific patterns, generates augmented AMR graphs, and creates training examples, eliminating the need for manual curation while maintaining high data quality.

Inventive Principle:

Principle #25Self-service

Solution Approach 2:

The patent introduces a domain lexicon legend as an intermediary between the augmentation process and the training data generation. This intermediary component stores domain-specific vocabulary and semantic relationships, mediating the transformation of generic AMR graphs into domain-specific training examples. The lexicon legend acts as a bridge that enables automatic generation of high-quality domain-specific data without manual involvement.

Inventive Principle:

Principle #24Intermediary (Mediator)

3Adaptability or versatility

If existing annotated training datasets are used without domain-specific augmentation, then the data availability is high, but the adaptability to new domains is poor

Engineering Contradiction:

Improvedomain adaptabilityVSAvoiddomain-specific training data

Core Design Contradiction:

Adaptability or versatilityVSQuantity of substance

Solution Approach 1:

The patent applies preliminary action by pre-processing the domain knowledge into a structured lexicon legend before the augmentation process. The domain lexicon legend is constructed in advance with domain-specific vocabulary, semantic relationships, and composition rules. This preliminary preparation enables the subsequent automatic generation of domain-specific training data without requiring manual curation for each new domain, thus improving adaptability while maintaining data quantity.

Inventive Principle:

Principle #10Preliminary action

Data Source

PatentUS20250005287A1Semantics-guided domain-specific data augmentation for text-to-graph parsing

Publication Date: 2025.01.02 INTERNATIONAL BUSINESS MACHINE CORPORATION

AI summary

Systems and techniques that facilitate semantics-guided domain-specific data augmentation for text-to-graph parsing are provided. In various embodiments, a system can access an annotated training dataset, wherein the annotated training dataset can comprise a set of abstract meaning representation graphs respectively corresponding to a set of natural language sentences. In various aspects, the system can generate an augmented version of the annotated training dataset, based on applying semantics-guided composition operations or semantics-guided substitution operations to the set of abstract meaning representation graphs. In various instances, a lexicon legend can comprise domain-specific graphs respectively representing discrete tokens unique to a domain of the annotated training dataset. In some cases, various of the domain-specific graphs can be composed or substituted onto or into various of the set of abstract meaning representation graphs, in response to semantic determinations, such as semantic-type-based determinations, argument-structure-based determinations, or incoming-semantic-relation-based determinations.