I'm working with a project supervisor on a deep learning project. The project involves extracting keywords or catchphrases from legal documents so that they can then be used for semantic search of legal documents for lawyers and in the future, predict verdicts to many cases and help streamline the work of the judiciary.
The problem lies not in the aforementioned model. The model being used (and improved potentially) is D2V-BiGRU-CRF, which is a supervised model. The problem lies in the dataset that we have. The model was trained on a dataset that it was trained on was labelled and foreign (although it did bear some similarities to ours). Our dataset lacks labels (due to the ineptness of our judicial body), which is why we cannot train it and we cannot derive metrics for it.
In short, there are no labels for it, and we have been trying hard to generate some catchphrases (our labels) from other models to be used here. But this has its own issues and inaccuracies. I looked around and found Snorkel, but that would require a very robust labeling function. So, my question is, is there anyway we can derive labeled data from the unsupervised models' results? Or is there some other way we can go about it? Do note that asking a person to manually label the data is not an option as we have resource constraints.