LLM-Guided Co-Training for Text Classification (2509.16516v2)

Abstract: In this paper, we introduce a novel weighted co-training approach that is guided by LLMs. Namely, in our co-training approach, we use LLM labels on unlabeled data as target labels and co-train two encoder-only based networks that train each other over multiple iterations: first, all samples are forwarded through each network and historical estimates of each network's confidence in the LLM label are recorded; second, a dynamic importance weight is derived for each sample according to each network's belief in the quality of the LLM label for that sample; finally, the two networks exchange importance weights with each other -- each network back-propagates all samples weighted with the importance weights coming from its peer network and updates its own parameters. By strategically utilizing LLM-generated guidance, our approach significantly outperforms conventional SSL methods, particularly in settings with abundant unlabeled data. Empirical results show that it achieves state-of-the-art performance on 4 out of 5 benchmark datasets and ranks first among 14 compared methods according to the Friedman test. Our results highlight a new direction in semi-supervised learning -- where LLMs serve as knowledge amplifiers, enabling backbone co-training models to achieve state-of-the-art performance efficiently.

• The paper presents a novel integration of LLMs as auxiliary teacher models to generate high-confidence pseudo-labels for semi-supervised co-training.
• It utilizes iterative refinement with multiple student models to boost accuracy, particularly on datasets with complex linguistic patterns.
• Empirical results show substantial improvements in macro F1 scores, indicating enhanced robustness and efficiency in text classification.

LLM-Guided Co-Training for Text Classification

The paper "LLM-Guided Co-Training for Text Classification" provides a comprehensive paper on leveraging LLMs to enhance co-training methods in text classification tasks. Co-training, a semi-supervised learning strategy, utilizes multiple models trained on distinct labeled data subsets to iteratively refine their predictions. In recent advancements, LLMs such as GPT-3, with their extensive pre-training and generalization capacities, offer new potentialities in guiding co-training processes.

Methodology

The core contribution of the paper lies in integrating LLMs into co-training by employing them as auxiliary teacher models. The principle mechanism involves utilizing the generative capabilities of LLMs to produce pseudo-labels for unlabeled data, which are then used alongside genuine labels to train distinct student models. This approach hinges on several computation strategies:

1. Pseudo-Label Generation: LLMs generate high-confidence pseudo-labels by operating on unlabeled datasets, exploiting their linguistic understanding of contexts to predict probable classifications.
2. Iterative Refinement via Student Models: Multiple student models undergo iterative training. Each model utilizes pseudo-labels generated by LLMs together with their own predictions to refine accuracy over successive iterations.
3. Multi-Modal Interaction: By facilitating interactions between student models via LLM-guided pseudo-labels, the system enhances model robustness, particularly across domains with linguistic complexities.

Implementation Details

In terms of practical implementation, integrating LLMs into existing co-training frameworks demanded specific attention to model architectures and computational resource management:

• Architecture Design: The implementation adopts a modular system where LLMs are distinct entities interfacing with student models, ensuring flexible integration which can exploit pre-trained models without extensive reconfiguration.
• Computational Considerations: The system leverages pre-trained LLM infrastructures, minimizing training time for generating pseudo-labels. The iterative student refinement is executed in parallel processes to handle large volumes of data efficiently.

Numerical Results

The empirical evaluation indicated that LLM-guided co-training improves text classification performance across diverse datasets. Notably, the approach achieved substantial accuracy gains in domains characterized by high variability in text structures and semantics:

• A noticeable improvement in macro F1 scores was observed when implementing LLM-guided pseudo-label generation compared to traditional co-training methods.
• Enhanced accuracy was most pronounced for complex datasets where LLMs could supplement interpretation of nuanced language patterns.

Implications and Future Directions

The theoretical implications underscore the efficacy of LLMs in augmenting semi-supervised learning, providing a viable pathway to mitigate label scarcity challenges. Additionally, the approach facilitates a deeper investigation into LLMs' potential in context-specific reasoning tasks beyond standard usages.

Future research could explore refining LLM-to-student interactions by improving pseudo-label accuracy further or investigating resilience in real-world noisy data settings. Furthermore, expanding this methodology to multilingual datasets could validate its scalability within universal language processing applications.

Conclusion

The integration of LLMs as guiding models in co-training for text classification embodies a substantial shift towards exploiting advanced LLM capacities in semi-supervised learning. By elevating model accuracy and robustness through iterative refinement, the proposed methodology presents a significant enhancement over traditional systems, unlocking new potentials for scalable and efficient classification across diverse text-based datasets.