💬 NLP

SSDAU: Structured Semantic Data Augmentation for Joint Entity and Relation Extraction

This paper proposes SSDAU, a novel data augmentation method that preserves semantic structure through entity-based segmentation, context-aware restructuring, and topic filtering to significantly improve the generalization and robustness of Joint Entity and Relation Extraction models against weak training data.

Original authors: Jiawei He, Mengyu Shi, Chunrong Fang

Published 2026-05-25

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jiawei He, Mengyu Shi, Chunrong Fang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot to read a story and pull out specific facts, like "Who did what to whom?" (e.g., "Steve Jobs founded Apple"). This task is called Joint Entity and Relation Extraction. The problem is that the robot is a picky eater; it only learns well if it has a huge, high-quality menu of examples. If the menu is small or the examples are messy, the robot gets confused and makes mistakes.

To fix this, scientists usually try to "cook up" more examples using a technique called Data Augmentation. Think of this as trying to make more copies of a recipe to feed the robot. However, most existing methods are like a clumsy chef who just swaps ingredients randomly. They might replace "Steve Jobs" with "Elon Musk" but leave the rest of the sentence weird, or they might chop the sentence in the middle of a thought. This creates "fake" recipes that look okay at a glance but are actually nonsense to the robot, causing it to learn the wrong lessons.

Enter SSDAU (Structured Semantic Data Augmentation).

The authors of this paper propose a new, smarter way to cook up these extra examples. Instead of randomly chopping and swapping, SSDAU acts like a master librarian who understands the structure of the story. Here is how it works, broken down into simple steps:

1. The "Lego" Breakdown (Discretization)

Imagine a sentence is a complex Lego castle. Old methods might try to smash the castle and rebuild it with random bricks. SSDAU, however, carefully takes the castle apart into its specific, meaningful blocks: the "Head" (who), the "Relation" (what they did), and the "Tail" (to whom).

The Analogy: It separates the sentence into "Steve Jobs" (Head), "Founded" (Relation), and "Apple" (Tail), keeping the surrounding words (the context) as a safety net so the meaning doesn't get lost.

2. The "Look-Alike" Matchmaker (Semantic Matching)

Now that the sentence is broken into blocks, SSDAU goes to a massive library to find other blocks that are semantically similar.

The Analogy: If the original block is "Steve Jobs," the system doesn't just pick any famous person. It looks for someone who fits the same role and context. It uses a special "smart scanner" (a BERT encoder) to understand that "Steve Jobs" and "Bill Gates" are both tech founders, but it also checks the surrounding words to make sure they fit the sentence perfectly. It's like finding a twin for a specific Lego piece that fits the exact same spot in the castle.

3. The "Quality Control" Inspector (Consistency Filtering)

This is the most crucial step. After swapping the blocks to create a new sentence, the system runs it through a strict Quality Control Inspector (using a model called BERTTopic).

The Analogy: Imagine you swapped "Steve Jobs" with "Elon Musk." The inspector checks: "Does this new sentence still make sense? Is the topic still about tech founders?" If the swap creates a confusing topic (like mixing a tech founder with a chef), the inspector throws that new sentence in the trash. This ensures that only high-quality, logical examples are fed to the robot.

Why is this better?

The paper claims that SSDAU is like upgrading from a "random word shuffler" to a "structural architect."

Old Methods: Often break the story's logic. If you train the robot on these broken stories, it gets confused and performs poorly, especially when the text is tricky or ambiguous.
SSDAU: Keeps the story's skeleton intact. It creates new examples that are diverse but still make perfect sense.

The Results

The authors tested this on several "menus" (datasets) and compared SSDAU against seven other popular methods.

The Outcome: SSDAU consistently made the robot smarter. When the text was tricky or ambiguous, other methods caused the robot's performance to crash (dropping by over 30% in some cases). SSDAU, however, stayed stable, only dropping by about 8%.
The Takeaway: By respecting the structure of the sentence and filtering out the "bad copies," SSDAU creates a much better training diet for the robot, helping it learn faster and more accurately, even when there isn't a lot of original data to start with.

In short, SSDAU doesn't just make more data; it makes better data by ensuring every new example is a logical, structurally sound copy of the original, preventing the robot from learning nonsense.

Technical Summary: SSDAU for Joint Entity and Relation Extraction

Problem Statement

Joint Entity and Relation Extraction (JERE) is a critical NLP task for applications like information retrieval and question answering. However, JERE models suffer from weak generalization due to low-quality or scarce training data. While data augmentation is a standard strategy to address data scarcity, existing methods (e.g., MixUp, back-translation, word substitution) often introduce noise that disrupts the semantic structure and dependencies inherent in entity-relation triples. This disruption can lead to issues such as overlapping relations, cascading errors, and weakened entity relevance, ultimately degrading model performance rather than enhancing it.

Methodology: Structured Semantic Data Augmentation (SSDAU)

The authors propose SSDAU, a novel augmentation framework designed to preserve the semantic structure of text during the data generation process. Unlike methods that perturb raw text, SSDAU aligns triplet text to maintain semantic integrity through three core components:

1. Data Discretization and Reconstruction

Encoder: The text is segmented based on triplet tags (head entity, relation, tail entity). A feature-based encoder processes the input sentences to extract specific text blocks ( $q_h, q_r, q_t$ ) while preserving context tokens ( $l$ ) and cut positions ( $p$ ). This generates three distinct text collections: head entities ( $Q_h$ ), tail entities ( $Q_t$ ), and relations ( $Q_r$ ).
Decoder: The decoder organizes these collections into groups ( $B$ ) based on relation types and label types, preparing them for semantic matching.

2. Structured Semantic Data Augmentation

Discrete Text Matching: The system employs a text matcher based on the Similarities tool. It evaluates semantic, syntactic, and lexical similarity between text blocks. Crucially, to distinguish between semantically similar but distinct entities, the method fuses contextualized [CLS] embeddings from a pretrained BERT encoder with traditional similarity scores using pretrained pooler weights.
Generation: Text blocks with high similarity scores (above a threshold $\epsilon$ ) are selected. The system replaces the original text content in specific blocks with high-similarity alternatives while maintaining the original segmentation positions, thereby generating augmented data that retains semantic coherence.

3. Scoring-Based Consistency Filtering

Topic-Aware Filtering: To mitigate topic ambiguity and information loss, the BERTTopic model is utilized. The system extracts entities and relations, encodes them, and computes a correlation score between head and tail entities.
Filtering Mechanism: Candidate triples are scored against gold-standard semantics. Triples inconsistent with the expected topic or semantic structure are filtered out. This step ensures that the augmented data maintains topic consistency and prevents error propagation during training.

Key Contributions

Structured Preservation: SSDAU introduces a paradigm that segments text based on entity labels and employs an encoder-decoder architecture to reorganize text while strictly preserving the underlying semantic structure, avoiding the noise injection common in perturbation-based methods.
Hybrid Similarity Metric: The paper proposes a method to fuse contextualized embeddings with traditional similarity scores, effectively distinguishing between semantically similar but distinct entities—a common failure point in existing augmentation techniques.
Consistency Filtering: The integration of a BERTTopic-based scoring mechanism ensures topic coherence in augmented data, significantly reducing the risk of generating semantically ambiguous or irrelevant training samples.
Comprehensive Evaluation: The method is evaluated across five representative JERE models (including PRGC, CasRel, TPLinker, SPN4RE, and OneRel) on four dataset variations (NYT and WebNLG, both fully and partially annotated).

Experimental Results

The authors report that SSDAU significantly outperforms seven popular baselines (including WS, BT, NI, SSN, GM, Mixup, and ChatIE) across all metrics:

Performance: SSDAU achieves an average precision of 92.03% and an F1 score of 91.96% across all datasets, substantially outperforming all baselines.
Robustness to Ambiguity: In scenarios involving semantically ambiguous data, SSDAU demonstrates superior stability. While baseline methods (e.g., Back Translation) suffered F1 score drops of up to 31.91%, SSDAU's performance degradation was limited to 8.26%.
Low-Resource Scenarios: The method proved particularly effective in low-resource settings and partially annotated datasets, showing improvements of up to 3% in precision for models like CasRel and SPN.
Ablation Studies: Experiments confirmed that each component (discretization, structured augmentation, and consistency filtering) contributes positively to the final performance. Specifically, removing the filtering mechanism or using non-pretrained initialization led to significant performance drops.

Significance and Claims

The paper claims that SSDAU offers a superior approach to data augmentation for JERE by prioritizing semantic structure preservation over simple text perturbation. The authors argue that their method effectively addresses the challenges of data scarcity and semantic ambiguity in low-resource scenarios. By ensuring that augmented data remains consistent with gold-standard semantics and topic coherence, SSDAU provides a robust solution for training more generalizable JERE models. The results suggest that preserving structured semantics through semantic tags and contextualized embeddings is a promising direction for future text data augmentation, particularly for tasks requiring high precision and disambiguation capabilities.

The authors acknowledge limitations, noting that while SSDAU filters low-quality data, the inclusion of more high-quality data could further enhance performance. Additionally, they suggest that the efficiency of the semantic text-matching component could be improved for long texts in future work.