Abstracts of the Presentations:

Jed Dobson (English, https://faculty-directory.dartmouth.edu/james-e-dobson)

Title: Messing with Mistral: Humanistic Modes of Evaluating Language Models

Abstract: In this presentation, I’ll report on a series of comparisons of the Mistral 7B base (“foundation”) model with the instruction fine-tuned (“instruct”) model. This comparative project makes use of interpretive methods from the emergent computational humanities in order to understand the transmission of harmful inputs and bias from a foundation model to fine-tuned models and other potential downstream uses. The comparative analyses include next token prediction on the base and fine-tune models (with and without instruction prompting) with identity categories (race, gender, class, sexuality, etc.); automatic evaluation of imposed guard-railing behavior on fine-tuned model using toxicity scores of responses to toxic red team prompts from RLHF training data; and scoring of identity-based prompts for the writing of fictional medical discharge summaries using lexicon-based methods. These multiple experiments suggest that while fine tuning for instruction reduces some possible harms to users of recent open language models, it preserves biases, stereotypes, and ideological assumptions found in the foundation model and readily generate, without complicated prompting, harmful output.

Praveen Kopalle (Tuck, https://www.tuck.dartmouth.edu/faculty/faculty-directory/praveen-k-kopalle) and Prasad Vana (Tuck, https://www.tuck.dartmouth.edu/faculty/faculty-directory/prasad-vana)

Title: Generating “Accurate” Online Reviews: Augmenting a Transformer-Based Approach with Structured Predictions

Abstract: A particular challenge with Generative Artificial Intelligence (GenAI) relates to the “hallucination” problem, wherein the generated content is factually incorrect. This is of particular concern for typical generative tasks in marketing. Here, we propose a two-step approach to address this issue. Our empirical context of an experience good (wines) where information about the taste of the product is important to the readers of the review but crucially, this data is unavailable a priori. Consequently, typical generative AI models may hallucinate this attribute in the generated review. Our approach of augmenting a transformer model with structured predictions results in a precision of .866 and a recall of .768 for the taste of wines, vastly outperforming popular benchmarks: transformer (precision .316, recall .250) and ChatGPT (precision .394, recall .243). We conduct an experimental study where respondents rated the similarity of reviews generated by our approach (versus those generated by ChatGPT) to those written by human wine experts. We find our reviews to be significantly more similar to human-expert reviews than those generated by ChatGPT. Apart from our app implementation, our main contribution in this work is to offer one approach towards more accurate GenAI, particularly towards marketing-related tasks.