💻 computer science

Multi-View Dual-Contrastive Graph Learning with Adaptive Agreement for Semi-Supervised Text Classification

The paper proposes Multi-View Dual-Contrastive Graph Learning (MDCGA), a lightweight framework that integrates lexical, semantic, and keyword-based relational signals through enhanced dual contrastive objectives to achieve robust semi-supervised text classification with few labeled examples by mitigating false negatives and unreliable augmentations.

Original authors: Ibtissam Youb, Mohamed Hamlich

Published 2026-09-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ibtissam Youb, Mohamed Hamlich

Original paper licensed under CC BY 4.0 (https://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

In the vast landscape of artificial intelligence, teaching computers to understand human language often feels like trying to fill a library with books that have no titles. While machines can read millions of documents, they struggle to sort them into meaningful categories when they are given only a handful of examples to learn from. This is the central challenge of semi-supervised text classification: how to build a smart system when the teacher has very little time to explain the rules. Traditional methods often rely on massive, pre-trained models that require enormous computing power, or they treat documents as isolated lists of words, missing the subtle connections that link one piece of writing to another. To solve this, researchers have turned to graph-based learning, a technique that maps documents as nodes in a network, connecting them based on how similar they are. This allows the computer to learn by observing the relationships between texts, much like a person might understand a new concept by seeing how it relates to things they already know. However, these existing graph methods often stumble when they try to improve themselves by comparing documents, sometimes mistakenly treating two similar texts as opposites simply because they lack a label, which confuses the learning process.

A team of researchers has developed a new approach called Multi-View Dual-Contrastive Graph Learning with Adaptive Agreement, or MDCGA, designed to navigate these pitfalls without needing a supercomputer. Instead of relying on a single way of measuring similarity, the system builds three distinct maps of the same collection of documents. The first map connects texts that share deep semantic meaning, understanding the general idea behind the words. The second map links documents that use the same specific vocabulary, focusing on the exact words chosen by the authors. The third map ties together texts that share key concepts or phrases, capturing the core subjects they discuss. By constructing these three separate but related views, the system mimics how a human reader might approach a new topic: by grasping the overall meaning, noticing the specific language used, and identifying the central themes. The researchers then train the computer to ensure that a single document looks consistent across all three maps, reinforcing the idea that these different perspectives are describing the same reality.

The most significant innovation in this work is a new method for filtering out errors that naturally occur during this learning process. In standard training, the computer often assumes that any two documents it hasn't been told are similar must be different. This assumption frequently leads to mistakes, where the system tries to push apart two documents that are actually related, simply because it hasn't seen their label yet. To fix this, the researchers introduced an adaptive agreement mechanism. The system checks if two documents are connected in at least two of the three maps; if they are, the system treats them as likely related and stops trying to separate them. It also watches the computer's own internal understanding as it learns, pausing the separation of documents that the system has already learned to view as similar. This dual-check system acts as a safety net, ensuring the computer does not unlearn the connections it has already discovered.

When tested on four different real-world datasets ranging from movie reviews and news articles to medical abstracts, this new framework proved remarkably effective. In a fully supervised setting where the computer had access to all the labels, it achieved high accuracy, reaching 97.40% on a news classification task and 70.49% on a complex medical dataset with twenty-three different categories. More impressively, the system shone when labels were extremely scarce, a scenario where most other methods fail. In tests where the computer was given less than one percent of the available labels to learn from, MDCGA outperformed massive language models that are millions of times larger. On the medical dataset, for instance, the new method achieved 55% accuracy with very few examples, while the largest competing models managed only 42%. This was accomplished while updating fewer than 600,000 parameters, a tiny fraction of the resources required by the larger models, demonstrating that a well-structured, smaller system can often do the heavy lifting more efficiently than a giant one.

The researchers also explored how the different parts of their system contributed to this success. They found that using all three views of the documents together was essential; removing any one of them caused the accuracy to drop. They discovered that the system was robust, meaning it did not require precise tuning of its settings to work well, and that the specific way the documents were connected mattered more than the complexity of the underlying math. By combining multiple ways of seeing the data with a smart filter for errors, the team showed that it is possible to build a highly effective text classifier that is both lightweight and reliable. This work suggests that in the race to teach machines to understand language, the key may not always be building bigger models, but rather designing smarter ways to let them learn from the relationships that already exist within the data.

Problem Statement

Semi-supervised text classification faces a critical bottleneck: real-world applications often lack abundant labeled data due to the high cost of annotation and the need for domain expertise. While Graph Neural Networks (GNNs) have emerged as a principled approach to leverage unlabeled data by modeling document relations, existing Graph Contrastive Learning (GCL) methods for text suffer from two primary weaknesses when adapted from general graph learning:

Semantically Destructive Augmentation: Standard GCL relies on stochastic augmentations (e.g., edge dropping, node masking) that treat graphs as generic structures. In text graphs, where edges encode semantic or lexical relations, random perturbation destroys the very signals the encoder needs to preserve.
False Negative Sampling: Standard contrastive objectives treat all non-anchor nodes as negatives. In text corpora, documents often share topical or conceptual relatedness even if they have different labels. Forcing the model to separate these "false negatives" distorts the embedding space and degrades classification performance.

Current methods often address these issues in isolation (e.g., improving view construction or filtering negatives via single-view clustering) but fail to exploit the consistency of document relatedness across independently constructed relational views.

Methodology: MDCGA

The authors propose MDCGA (Multi-View Dual-Contrastive Graph Learning with Adaptive Agreement), a framework designed to learn a shared document representation by integrating multiple relational signals while mitigating false negatives.

1. Multi-View Document Graph Construction

Instead of using random noise for augmentation, MDCGA constructs three distinct, domain-agnostic document graphs based on how humans identify relatedness:

Semantic View ( $G_S$ ): Derived from Sentence-T5 embeddings. Edges are formed between documents with high cosine similarity in their semantic vectors.
Lexical View ( $G_L$ ): Based on shared high TF-IDF terms. Edges connect documents sharing a sufficient number of salient terms.
Keyphrase View ( $G_K$ ): Based on overlapping keyphrases extracted via KeyBERT. Edges connect documents sharing salient multi-word concepts.

2. Graph Encoding and Fusion

A shared-parameter GraphSAGE encoder processes all three views independently. This forces the model to learn consistent mappings for the same document across different relational perspectives. The final fused embedding for a document is the average of its representations from the three views.

3. Dual-Contrastive Objectives

The framework employs two complementary contrastive losses:

Cross-View Alignment: Ensures that the embeddings of the same document across different views (e.g., Semantic vs. Lexical) are pulled together, preserving document identity.
Within-View Consistency: Ensures that a document remains close to its neighbors within a specific view, preserving local graph structure.

4. Adaptive Agreement Masking (False Negative Filtering)

To address the false negative problem, MDCGA introduces an Adaptive Agreement Mask that excludes likely positive pairs from the negative set. This mask combines two independent signals:

Static Structural Consensus: If two documents are connected in at least two of the three views, they are treated as likely related and masked out from negative sampling.
Dynamic Semantic Proximity: If the learned embeddings of two documents become sufficiently similar during training, they are also masked out.
This dual-cue approach ensures that the model does not force separation on pairs that are genuinely related, refining the geometry of the embedding space.

Key Contributions

Multi-View Construction without Destructive Augmentation: The paper introduces a method to construct three linguistically principled views (semantic, lexical, keyphrase) that capture distinct facets of document relatedness, avoiding the semantic destruction caused by random graph perturbations.
Cross-View Consensus for False Negative Filtering: The authors propose the first use of cross-view consensus as an explicit filter for false negatives in semi-supervised text graph contrastive learning. By requiring agreement across independent views (and dynamic embedding proximity), the method identifies unreliable negatives more robustly than single-view approaches.
Dual-Contrastive Objective: The framework jointly optimizes for cross-view identity alignment and within-view neighborhood coherence, creating a unified representation space.
Parameter Efficiency: The approach achieves competitive performance with a lightweight architecture (fewer than 600K parameters), contrasting with the massive parameter counts of Large Language Models (LLMs).

Experimental Results

The authors evaluated MDCGA on four benchmarks: MR (Movie Reviews), Ohsumed (Medical Abstracts), AG News, and R8 (Reuters).

Fully Supervised Setting: MDCGA achieved state-of-the-art or competitive results, with accuracies of 85.71% (MR), 70.49% (Ohsumed), 92.76% (AG News), and 97.40% (R8). It notably outperformed other graph-based methods and matched the performance of a 175B-parameter GPT-3.5 model on MR while using significantly fewer parameters.
Semi-Supervised (Low-Label) Setting: Under extreme label scarcity (e.g., 0.38% labeled data for MR, 6.22% for Ohsumed), MDCGA demonstrated superior robustness:
- It significantly outperformed fully fine-tuned BERT and RoBERTa, which collapsed in low-data regimes.
- It surpassed QLoRA-fine-tuned 7B–8B LLMs (Llama-2, Llama-3, Bloom) on Ohsumed and R8. For instance, on Ohsumed, MDCGA reached 55.01% accuracy compared to the best LLM's 42.16%.
- On AG News and MR, LLMs retained a moderate advantage, likely due to domain overlap with their pretraining data, but MDCGA remained highly competitive.
Parameter Efficiency: MDCGA trains fewer than 600,000 parameters, which is less than 2% of the parameters updated in QLoRA-tuned Llama-3-8B and less than 0.5% of fully fine-tuned RoBERTa-base, yet achieves comparable or better accuracy in low-label scenarios.
Ablation Studies:
- Views: The tri-view configuration consistently outperformed single or dual-view variants, confirming the complementarity of semantic, lexical, and keyphrase signals.
- Masking: The full adaptive mask (combining structural and dynamic cues) yielded the best results, particularly on datasets with high class counts (Ohsumed, AG News), validating the necessity of false negative filtering.
- Robustness: The method showed limited sensitivity to hyperparameter choices, with accuracy variations remaining within standard deviation ranges across different top-k and threshold settings.

Significance and Claims

The paper claims that MDCGA offers a practical trade-off for semi-supervised text classification: it substitutes the massive scale of Large Language Models with the structural exploitation of unlabelled data through a lightweight graph encoder.

The authors emphasize that in regimes of extreme label scarcity, particularly in specialized domains (like medical abstracts) or fine-grained topical structures, the ability to model document relations explicitly via multi-view graphs provides a more effective signal than the general pretraining of LLMs. The work demonstrates that a small, structurally aware model can match or surpass the performance of billion-parameter models while remaining computationally feasible on consumer hardware. The introduction of cross-view consensus as a false-negative filter is presented as a novel mechanism to improve the reliability of contrastive learning in text.