The core difficulty is what machine learning researchers call zero-shot classification. Traditional supervised text classifiers learn from labeled examples of every category they are expected to recognize. That works well when the category list is fixed and each category has plenty of training data. Hotline systems violate both assumptions. A few frequent categories, such as housing or traffic management, accumulate thousands of samples, while many others sit in a long tail with only a handful of observations. Worse, entirely new categories can appear at any moment, leaving the classifier with no labeled examples at all. Zero-shot methods promise a way around this by transferring knowledge from categories the model has seen to categories it has not, using semantic descriptions as a bridge.

Hotline texts, however, are a hostile environment for such transfer. They are short, colloquial, and non-standardized, often compressed into a few fragmented sentences that carry weak semantic signals. Meanwhile, hotline categories themselves tend to be semantically similar: a complaint about urban sanitation can look superficially like one about waste classification or property management. The result is overlapping class representations and blurred decision boundaries in the embedding space. Class labels are usually concise and abstract, offering little semantic information to anchor the alignment between texts and categories. The authors’ answer is a method they call Attribute-Guided Latent Space Alignment, or AGLA, which builds richer descriptions of each category and then forces texts and those descriptions to meet inside a carefully engineered shared latent space.

The first step is attribute construction. Rather than relying on bare labels, the team designed domain-specific templates that convert each label into a structured natural language description. A template of the form “The topic of this hotline request is related to [TOPIC] and belongs to the [DOMAIN] domain” expands a terse label such as “rural management” into a declarative sentence covering villagers’ self-governance, homestead approvals, collective assets, and infrastructure gaps. These attribute descriptions are then encoded with the Chinese-language embedding model bge-large-zh-v1.5 into 1024-dimensional vectors. To make the attributes more robust, the framework applies augmentation: random noise is injected into the attribute means, and attribute vectors within a training batch are mixed together using coefficients drawn from a symmetric Beta distribution. Multiple perturbed samples are drawn and averaged, pulling each class vector toward a stable region of the latent space and sharpening the class boundaries that unseen categories will eventually inherit.

The second pillar is a dual-channel variational autoencoder, or DC-VAE. Variational autoencoders learn to compress data into probabilistic latent variables by maximizing the evidence lower bound, balancing reconstruction fidelity against a Kullback-Leibler regularization term that keeps the latent distribution close to a prior. In AGLA, one channel encodes hotline texts and the other encodes class attributes, and both are projected into a single shared latent space of modest dimensionality, where their semantic relationship can be modeled uniformly. But short, noisy hotline texts create a well-known pathology: posterior collapse. When inputs carry little information, the encoder’s posterior degenerates toward the prior, the latent variables stop encoding anything useful about the input, and the mutual information between observation and latent code collapses toward zero. A collapsed latent space is useless for alignment.

To prevent this, the authors introduce a Mutual Information Booster. Self-attention first extracts attention weights from the textual input, and a weighted pooling operation produces a global representation that a multilayer perceptron converts into a sample-dependent, probability-aware weight. For attributes, a simpler perceptron predicts the corresponding weight directly. This weight then scales the KL regularization term on a per-sample basis, relaxing the constraint for inputs that carry rich information and tightening it elsewhere. The result is a latent space that preserves far more input-dependent semantic content, which the ablation experiments confirm is essential: without the booster, adding alignment alone lifts accuracy only to 36.76 percent, whereas with it the full model reaches 62.78 percent.

Alignment itself proceeds on two levels. Globally, a latent matching loss explicitly minimizes the discrepancy between the Gaussian distributions of text and attribute latent variables, matching both means and variances. The authors deliberately avoid the KL divergence between the two Gaussians because it is asymmetric and overly sensitive to variance estimates; instead they use an explicit mean-and-variance matching objective. Locally, a contrastive learning scheme based on dynamic negative sampling treats each text latent variable as an anchor, its matching attribute as a positive sample, and the most similar attribute from a different class as a hard negative. A triplet loss with a margin of 0.5 pulls anchors toward their positives while pushing them away from the nearest competing class, carving out well-separated semantic neighborhoods. The loss weights for all objectives are not fixed hyperparameters but are dynamically adjusted through warm-up schedules and adaptive curves that rise, decay, or peak as training progresses.

Once texts and attributes are aligned, training the zero-shot classifier is straightforward: the attribute latent variables of unseen classes are fed to a simple softmax classifier trained on seen-class data, and unseen-class texts are then assigned by proximity in the shared space. The evaluation used a real-world dataset of 98,641 hotline texts collected between January 2021 and October 2022, spanning 47 classes, from which 20 classes with at least 500 samples were selected. Five experimental groups with different combinations of unseen classes were tested, each repeated with five random seeds. Accuracy mostly stayed above 60 percent, peaking at 62.78 percent, with F1 scores between 0.51 and 0.62. Performance varied with the semantic character of the unseen classes: one group containing categories such as waste sorting and traffic management scored lower because their textual topics diverged sharply from the training distribution, while the best-performing group benefited from classes more closely aligned with the training set.

The comparisons are striking. Among variational autoencoder-based zero-shot baselines, the best competitor, a bidirectional VAE, reached 53.49 percent accuracy, roughly ten points below AGLA. More surprising is how poorly general-purpose pre-trained language models fared: StructBERT and PromptCLUE-large managed only 16 to 17 percent accuracy on hotline texts, and even Qwen2-7B-Instruct, a large language model fine-tuned on the hotline data, reached 52.48 percent, still below the proposed method. The authors attribute this to the models’ insufficient domain knowledge, hallucination risks, and the confidentiality constraints that restrict sending sensitive citizen complaints to external large models. On the Yelp Review dataset, whose short, semantically sparse texts resemble hotline requests, AGLA again led with 55.77 percent accuracy, suggesting genuine cross-domain generalization rather than overfitting to one corpus.

The study also stress-tested temporal robustness. On complaint texts collected during 2020 to 2021, a period disjoint from the main dataset, the method achieved 73.41 percent accuracy and a macro-F1 of 70.96 percent on eight water-supply complaint classes split into seen and unseen groups, indicating that the learned alignment survives shifts in how citizens phrase their grievances over time. Diagnostic analyses revealed where the remaining errors concentrate: classes whose attributes share high cosine similarity, such as certain overlapping municipal categories, show elevated misclassification rates, confirming that fine-grained semantic overlap and blurred business boundaries, not model capacity, are the principal bottleneck. Visualization of the aligned distributions showed median differences between text and attribute features below 0.01 and interquartile ranges overlapping by more than 85 percent, evidence that the two modalities genuinely share one semantic geometry. The authors point toward hierarchical label structures and richer attribute construction as the next frontier. For city governments drowning in citizen requests, the message is concrete: a carefully regularized latent space, guided by well-crafted attribute descriptions, can recognize complaint categories that no annotator has ever labeled, turning tomorrow’s novel grievances into today’s actionable data.

Subject of Research: Zero-shot text classification of government hotline requests using attribute-guided latent space alignment

Article Title: Attribute-guided latent space alignment for zero-shot text classification of hotline requests

Article References: Fan, Q., Diao, Y., & Han, H. (2026). Attribute-guided latent space alignment for zero-shot text classification of hotline requests. Machine Learning with Applications, Article 101025. https://doi.org/10.1016/j.mlwa.2026.101025

Image Credits: AI Generated

DOI: Not provided

Keywords: zero-shot learning, text classification, variational autoencoder, latent space alignment, hotline requests, urban governance, natural language processing, contrastive learning, class attributes, short text, machine learning, public services