Text embeddings have become an essential part of a variety of language applications. However, methods for interpreting, exploring and reversing embedding spaces are limited, reducing transparency and precluding potentially valuable generative use cases. In this work, we develop an open-source, domain-agnostic framework for aligning Large Language Models to embedding spaces using the recently reported Embedding Language Model (ELM) method. We demonstrate our framework by training models to recover, summarize, and compare clinical trial abstracts from embeddings alone. In addition to inverting embeddings back to text more reliably than existing methods, our models can decode novel, interpolated embeddings into new clinical trial abstracts that human experts cannot distinguish from real ones. We further show that these generated abstracts are responsive to moving embeddings along concept vectors for age and sex of study subjects. Our public ELM implementation and experimental results will aid the alignment of Large Language Models to embedding spaces in the biomedical domain and beyond.

Though online medical literature has made health information more available than ever, the barrier of complex medical jargon prevents the general public from understanding it. Though parallel and comparable corpora for Biomedical Text Simplification have been introduced, these conflate the many syntactic and lexical operations involved in simplification. To enable more targeted development and evaluation, we present a fine-grained lexical simplification task and dataset, Jargon Explanations for Biomedical Simplification (JEBS). The JEBS task involves identifying complex terms, classifying how to replace them, and generating replacement text. The JEBS dataset contains 21,595 replacements for 10,314 terms across 400 biomedical abstracts and their manually simplified versions. Additionally, we provide baseline results for a variety of rule-based and transformer-based systems for the three subtasks. The JEBS task, data, and baseline results pave the way for development and rigorous evaluation of systems for replacing or explaining complex biomedical terms.

The cloze training objective of Masked Language Models makes them a natural choice for generating plausible distractors for human cloze questions. However, distractors must also be both distinct and incorrect, neither of which is directly addressed by existing neural methods. Evaluation of recent models has also relied largely on automated metrics, which cannot demonstrate the reliability or validity of human comprehension tests. In this work, we first formulate the pedagogically motivated objectives of plausibility, incorrectness, and distinctiveness in terms of conditional distributions from language models. Second, we present an unsupervised, interpretable method that uses these objectives to jointly optimize sets of distractors. Third, we test the reliability and validity of the resulting cloze tests compared to other methods with human participants. We find our method has stronger correlation with teacher-created comprehension tests than the state-of-the-art neural method and is more internally consistent. Our implementation is freely available and can quickly create a multiple choice cloze test from any given passage.