We introduce Semantic Echo Pathways (SEP), a new approach for modeling the cross-domain evolution of medical language. Using continual neural topic models (CoNTM) trained separately on scientific literature, clinical notes, and public health-related data, we track linguistic drift and identify points where concepts change meaning. We propose three novel metrics: Cross-Domain Drift Score, Temporal Echo Lag, and Semantic Mutation Patterns to quantify how medical language travels between the scientific, clinical, and public domain. Applications to evolving concepts such as “long COVID”, diagnostic category changes reveal previously undocumented patterns of medical-semantic evolution. Our results bridge computational modeling with the human-centered perspectives of medical humanities, offering clear, domain-aware maps of how medical language shifts across time and domains, and combining quantitative analysis with linguistic and clinical insight.

While retrieval-augmented generation (RAG) relies on retrieved evidence to mitigate hallucination, its reliability is constrained by retrieval quality and diagnosing where retrieval fails remains difficult. Existing benchmarks typically report a single aggregate score over heterogeneous query sets and are often dominated by queries whose relevant passages can be identified through simple lexical overlap, concealing the failure modes that matter most in practice. We introduce GutenBench, a fine-grained, zero-shot evaluation dataset built around a four-level retrieval taxonomy: lexical, semantic, explicit aggregation, and latent conceptual retrieval. It pairs queries with relevance judgments adjudicated by subject-matter experts, enabling precise and controlled comparisons of retrievers. We evaluate several off-the-shelf retrievers in an asymmetric, English-to-target cross-lingual setting across 272 Bible translations (141 languages) and find that performance is strong on lexical and semantic queries, but degrades when models must aggregate information across passages or recognize conceptually relevant passages with no surface overlap.

In continual learning, our aim is to learn a new task without forgetting what was learned previously. In topic models, this translates to learning new topic models without forgetting previously learned topics. Previous work either considered Dynamic Topic Models (DTMs), which learn the evolution of topics based on the entire training corpus at once, or Online Topic Models, which are updated continuously based on new data but do not have long-term memory. To fill this gap, we propose the Continual Neural Topic Model (CoNTM), which continuously learns topic models at subsequent time steps without forgetting what was previously learned. This is achieved using a global prior distribution that is continuously updated. In our experiments, CoNTM consistently outperformed the dynamic topic model in terms of topic quality and predictive perplexity while being able to capture topic changes online. The analysis reveals that CoNTM can learn more diverse topics and better capture temporal changes than existing methods.

Fine-tuning pretrained language models on task-specific data is a common practice in Natural Language Processing (NLP) applications. However, the number of pretrained models available to choose from can be very large, and it remains unclear how to select the optimal model without spending considerable amounts of computational resources, especially for the text domain. To address this problem, we introduce PsyMatrix, a novel framework designed to efficiently characterize text datasets. PsyMatrix evaluates multiple dimensions of text and discourse, producing interpretable, low-dimensional embeddings. Our framework has been tested using a meta-dataset repository that includes the performance of 24 pretrained large language models fine-tuned across 146 classification datasets. Using the proposed embeddings, we successfully developed a meta-learning system capable of recommending the most effective pretrained models (optimal and near-optimal) for fine-tuning on new datasets.