The KITMUS Test: Evaluating Knowledge Integration from Multiple Sources

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 15088–15108 July 9-14, 2023 ©2023 Association for Computational Linguistics The KITMUS Test: Evaluating Knowledge Integration from Multiple Sources Akshatha Arodi12∗, Martin Pömsl12∗, Kaheer Suleman3, Adam Trischler3, Alexandra Olteanu3, Jackie Chi Kit Cheung12 1 McGill University 2 Mila Quebec AI Institute 3 Microsoft Research {akshatha.arodi@mail, martin.pomsl@mail, jcheung@cs}.mcgill.ca {kaheer.s}@gmail.com {adam.trischler, alexandra.olteanu}@microsoft.com Abstract Many state-of-the-art natural language under- standing (NLU) models are based on pretrained neural language models. These models often make inferences using information from multi- ple sources. An important class of such infer- ences are those that require both background knowledge, presumably contained in a model’s pretrained parameters, and instance-specific in- formation that is supplied at inference time. However, the integration and reasoning abili- ties of NLU models in the presence of multiple knowledge sources have been largely under- studied. In this work, we propose a test suite of coreference resolution subtasks that require reasoning over multiple facts. These subtasks differ in terms of which knowledge sources contain the relevant facts. We also introduce subtasks where knowledge is present only at inference time using fictional knowledge. We evaluate state-of-the-art coreference resolution models on our dataset. Our results indicate that several models struggle to reason on-the- fly over knowledge observed both at pretrain time and at inference time. However, with task- specific training, a subset of models demon- strates the ability to integrate certain knowl- edge types from multiple sources. Still, even the best performing models seem to have dif- ficulties with reliably integrating knowledge presented only at inference time. 1 Introduction Progress on natural language understanding (NLU) benchmarks has recently been driven by pretrained large language models (LLMs), which can be adapted to specific tasks via finetuning (Peters et al., 2018; Devlin et al., 2019; Le Scao and Rush, 2021). These models draw on a variety of knowl- edge sources, such as knowledge given in inputs at inference time and knowledge contained in their parameters, usually acquired via pretraining. ∗ Equal contribution 1. Servin is a judge. Kea is a baker. Servin and Kea met at a park. After a long day at work deciding cases in a law court, he was happy to relax. [Answer: Servin] 2. Schwing is a gladiower. The work of a gladiower is inwaging ledmonly. The work of a popesmer is chodoling larely. Bate is a popesmer. At the coffee shop, Schwing and Bate connected. The coffee was excellent. She shared experiences from a career of chodoling larely. [Answer: Bate] Figure 1: Examples from KITMUS showing coreference cases that require real (1.) and fictional (2.) knowledge. To resolve the pronoun (in red), a model needs to draw on entity-specific knowledge about an entity’s occupa- tion as well as on background knowledge about what kind of work the occupation entails. Recent work suggests that models can use pretrain-time knowledge in tasks like translation and question answering to obtain performance gains (Brown et al., 2020; Roberts et al., 2020). However, natural …

Content truncated to 500 words