Factual or Contextual? Disentangling Error Types in Entity Description Generation

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 8322–8340 July 9-14, 2023 ©2023 Association for Computational Linguistics Factual or Contextual? Disentangling Error Types in Entity Description Generation Navita Goyal University of Maryland navita@cs.umd.edu Ani Nenkova Adobe Research nenkova@adobe.com Hal Daumé III University of Maryland Microsoft Research me@hal3.name Abstract In the task of entity description generation, given a context and a specified entity, a model must describe that entity correctly and in a contextually-relevant way. In this task, as well as broader language generation tasks, the gener- ation of a nonfactual description (factual error) versus an incongruous description (contextual error) is fundamentally different, yet often con- flated. We develop an evaluation paradigm that enables us to disentangle these two types of errors in naturally occurring textual contexts. We find that factuality and congruity are often at odds, and that models specifically struggle with accurate descriptions of entities that are less familiar to people. This shortcoming of language models raises concerns around the trustworthiness of such models, since factual errors on less well-known entities are exactly those that a human reader will not recognize.1 1 Introduction Gricean maxims of effective communication (Grice, 1975) as they pertain to referring expres- sions (Dale and Reiter, 1995) posit that referring expressions should not convey false information and that they should be relevant to context. Factu- ality and congruity are thus the two main properties of pragmatically appropriate referring expressions. Following these maxims, human-written refer- ring expressions strongly adhere to the principles of factuality and congruity (Dale and Reiter, 1995; Kheirabadi and Aghagolzadeh, 2012). Standard evaluation practices for referring expression gener- ation (Belz et al., 2009; Kang et al., 2019; Cao and Cheung, 2019), however, only distinguish between model generated referring expressions being accu- rate (ground-truth) versus inaccurate (not ground- truth), without considering factuality and congruity of the model outputs. However, this distinction 1The code and data used in the paper is available at https: //github.com/navitagoyal/Factual-or-Contextual-E rrors-in-LM-Desc-Gen. MASK Thomas Bach MASK an- nounced protest zones in the Sochi Winter Olympics 2014. Ground-Truth International Olympic Committee President Nonfactual Russian Olympic Committee Chairman The event opened with a scene dubbed "Green and Pleasant", af- ter a line from MASK William Blake MASK . Ground-Truth the English Poet Incongruous the English Painter Figure 1: Two example contexts with MASK s represent- ing potential location of a target referring expression. In the top case, the inaccurate generation is nonfactual (not true) but contextually plausible (given “Sochi”); in the bottom, the inaccurate generation is factual (Blake was both a poet and a painter) but incongruous. between factual and contextual errors is important, as contextually-relevant factual errors are likely to be harder for people to identify; this concern is supported by evidence that human annotators trust translations that are fluent/coherent but inadequate over translations that are adequate but disfluent (Martindale and Carpuat, 2018; Popovi´c, 2020). In this work, we design an evaluation framework to study the distribution of factual and contextual errors in …

Content truncated to 500 words