Standardize: Aligning Language Models with Expert-Defined Standards for Content Generation

TL;DR

Standardize, a retrieval-style in-context learning-based framework to guide large language models to align with expert-defined standards, is introduced, demonstrating that the use of knowledge artifacts extracted from standards and integrating them in the generation process can effectively guide models to produce better standard-aligned content.

Abstract

Domain experts across engineering, healthcare, and education follow strict standards for producing quality content such as technical manuals, medication instructions, and children's reading materials. However, current works in controllable text generation have yet to explore using these standards as references for control. Towards this end, we introduce Standardize, a retrieval-style in-context learning-based framework to guide large language models to align with expert-defined standards. Focusing on English language standards in the education domain as a use case, we consider the Common European Framework of Reference for Languages (CEFR) and Common Core Standards (CCS) for the task of open-ended content generation. Our findings show that models can gain a 45% to 100% increase in precise accuracy across open and commercial LLMs evaluated, demonstrating that the use of knowledge artifacts extracted from standards and integrating them in the generation process can effectively guide models to produce better standard-aligned content.

Figures (11)

• Figure 1: In contrast to the simple prompting method used by teachers, the proposed Standardize framework aims to improve the performance of generative models for content generation by using the fine-grained information found in expert-defined standards. The framework involves a three-part process starting with the (i) extraction of target specifications from the prompt, (ii) lookup and retrieval of information that matches the target specifications from the specified standard, and (iii) knowledge augmentation to produce artifacts that represent the standard itself for integration into the generation process with generative models.
• Figure 2: A standard contains recommended characteristics of content across one or more domain-specific aspects or criteria. This figure shows an example of the CEFR standard where the set of criteria includes depth of meaning, structure, and grammatical complexity.
• Figure 3: A standard contains aspect definition which can be represented by flags such as linguistic variables. Given the mean values from gold-standard data in the target level, the generative model can then be steered to push the property of its generated content using directional instructions such as increase or decrease.
• Figure 4: A standard contains recommended exemplars that serve as gold-standard reference. This figure shows an example of the CEFR standard where three well-known pieces of literature are provided as examples of content that conforms to the target level specified (B1).
• Figure 5: Overview of mean ratings of grammaticality or fluency, coherence, and grade complexity distinction from the human expert evaluations using the top-performing models for CEFR and CCS. All evaluation procedures obtain generally favorable results as well as acceptable inter-rater reliability scores (equal and above the threshold of $0.30$)