| Zero-shot multi-label text classification assigns all relevant labels to an input text when some labels have no training instances. This task is important in application domains where the set of labels is constantly enriched and it is not feasible to obtain adequate training data for new categories, e.g., news, legal documents, healthcare, and e-commerce. A common approach in this area is to attempt to represent both documents and label descriptions in a common embedding space to reveal the most relevant labels per document. To this end, very short label descriptions, often a single word or a few words, are used. Moreover, predicting the appropriate number of labels per document is inherently difficult. In this paper, we consider the use of LLM-generated label descriptions, allowing existing zero-shot multi-label text classification methods to better represent labels and estimate their relevance to documents. We introduce a heterogeneous encoder-decoder approach combining a dual pre-trained encoder, frozen on the label side to preserve zero-shot capability, with a T5 decoder and a cardinality head that adapts the number of predicted labels to each input document. In the presented experiments, we used two datasets from the legal and healthcare domains to demonstrate how the performance of existing methods is improved when LLM-generated label descriptions are used. We also test various encoder-decoder models and report improved performance results over strong baselines. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.