KeySI Framework Tunes Text Embeddings with Keyword Feedback.

Pre-trained language models are widely used for text embedding in large-scale analysis, but they often struggle to capture domain-specific semantics. Adapting these models typically demands extensive labeled data and specialized technical expertise for training. While visual interactions with document projections have shown promise for capturing human feedback, these methods usually require users to review individual documents, which can be time-consuming and inefficient. To streamline this process, the KeySI interaction framework has been introduced. KeySI enables feature-level feedback through a keyword-based concept specification approach. Users provide feedback by grouping extracted keywords into meaningful concepts, which the system then translates into document-level supervision for subsequent model tuning. This keyword-centric interaction significantly reduces the need for manual document inspection and labeling. A prototype implementation of KeySI curates representative keywords, visualizes both keywords and document embeddings using dimensionality reduction, and supports iterative refinement. User studies and quantitative experiments demonstrate KeySI's effectiveness in accurately capturing user intent and improving the alignment of embedding models with domain-specific requirements, thereby lowering the barrier for adapting these powerful tools.