Abstract:The web serves as a global repository of knowledge, used by billions of people to search for information. Ensuring that users receive the most relevant and up-to-date information, especially in the presence of multiple versions of web content from different time points remains a critical challenge for information retrieval. This challenge has recently been compounded by the increased use of question answering tools trained on Wikipedia or web content and powered by large language models (LLMs) which have been found to make up information (or hallucinate), and in addition have been shown to struggle with the temporal dimensions of information. Even Retriever Augmented Language Models (RALMs) which incorporate a document database to reduce LLM hallucination are unable to handle temporal queries correctly. This leads to instances where RALMs respond to queries such as "Who won the Wimbledon Championship?", by retrieving document passages related to Wimbledon but without the ability to differentiate between them based on how recent they are. In this paper, we propose and evaluate, TempRALM, a temporally-aware Retriever Augmented Language Model (RALM) with few-shot learning extensions, which takes into account both semantically and temporally relevant documents relative to a given query, rather than relying on semantic similarity alone. We show that our approach results in up to 74% improvement in performance over the baseline RALM model, without requiring model pre-training, recalculating or replacing the RALM document index, or adding other computationally intensive elements.
Abstract:For middle-school math students, interactive question-answering (QA) with tutors is an effective way to learn. The flexibility and emergent capabilities of generative large language models (LLMs) has led to a surge of interest in automating portions of the tutoring process - including interactive QA to support conceptual discussion of mathematical concepts. However, LLM responses to math questions can be incorrect or mismatched to the educational context - such as being misaligned with a school's curriculum. One potential solution is retrieval-augmented generation (RAG), which involves incorporating a vetted external knowledge source in the LLM prompt to increase response quality. In this paper, we designed prompts that retrieve and use content from a high-quality open-source math textbook to generate responses to real student questions. We evaluate the efficacy of this RAG system for middle-school algebra and geometry QA by administering a multi-condition survey, finding that humans prefer responses generated using RAG, but not when responses are too grounded in the textbook content. We argue that while RAG is able to improve response quality, designers of math QA systems must consider trade-offs between generating responses preferred by students and responses closely matched to specific educational resources.