Reference-Based Metrics Are Biased Against Blind and Low-Vision Users’ Image Description Preferences

“While many reference-based metrics have been proposed, we focus in this work on the three most commonly used metrics for image captioning: BLEU (Papineni et al., 2002), METEOR (Banerjee and Lavie, 2005), and ROUGE-L (Lin, 2004).”

“…from machine translation and summarization literature, reference-based metrics such as BLEU (Papineni et al., 2002), METEOR (Baner-jee and Lavie, 2005), ROUGE (Lin, 2004), CIDEr (Anderson et al., 2016), and SPICE (Vedantam et al., 2015) are the most common method for evaluating descriptions.”

“…from machine translation and summarization literature, reference-based metrics such as BLEU (Papineni et al., 2002), METEOR (Baner-jee and Lavie, 2005), ROUGE (Lin, 2004), CIDEr (Anderson et al., 2016), and SPICE (Vedantam et al., 2015) are the most common method for evaluating descriptions.”

“However, these systems often do not directly address the needs of BLV users (MacLeod et al., 2017), which recent work has started to partially attribute to the evaluation meth-ods used (Kreiss et al., 2022a).”

“Vision-language models have enabled the automation of image-to-text description generation, which can be used to generate alt-text descriptions; this could make visual content accessible to BLV users through, e.g., screen readers (Morris et al., 2016; Gleason et al., 2019).”

“While BLV user preferences for image descriptions are usually studied in isolation (Das et al., 2024; Muehlbradt and Kane, 2022; Stangl et al., 2020), recent work has started to investigate how those preferences compare to sighted user judgments (Kreiss et al., 2022a; Lundgard and Satyanarayan,…”

“For example, models should be able to indicate uncertainty over generated content (MacLeod et al., 2017), optimize for identity-respecting language (Bennett et al., 2021), and be severely sensitive to hallucinations (MacLeod et al., 2017).”

“As the internet becomes increasingly visual, long-standing accessibility issues blind and low-vision (BLV) users face remain largely unresolved (Gleason et al., 2019; Kreiss et al., 2022b).”

“While recent work suggests that image context shapes sighted and BLV participants’ preference ratings (Kreiss et al., 2022a; Stangl et al., 2021; Muehlbradt and Kane, 2022), these metrics tend to be used in context-agnostic settings.”

“Prior evidence suggests that context is an important signal for evaluating the usefulness of a description (Stangl et al., 2021; Muehlbradt and Kane, 2022; Kreiss et al., 2022a) and we should therefore expect the correlations to decrease due to the noisier reference signal.”

“While BLV user preferences for image descriptions are usually studied in isolation (Das et al., 2024; Muehlbradt and Kane, 2022; Stangl et al., 2020), recent work has started to investigate how those preferences compare to sighted user judgments (Kreiss et al., 2022a; Lundgard and Satyanarayan,…”

“While recent work suggests that image context shapes sighted and BLV participants’ preference ratings (Kreiss et al., 2022a; Stangl et al., 2021; Muehlbradt and Kane, 2022), these metrics tend to be used in context-agnostic settings.”

“Strikingly, BLV users tend to have strong preferences for description length (McCall and Chagnon, 2022).”

“…BLV user preferences for image descriptions are usually studied in isolation (Das et al., 2024; Muehlbradt and Kane, 2022; Stangl et al., 2020), recent work has started to investigate how those preferences compare to sighted user judgments (Kreiss et al., 2022a; Lundgard and Satyanarayan, 2021).”

“Reference-based metrics have also been shown to be biased toward sighted user ratings of alt text over BVI user ratings [KK24].”

“These inconsistencies can significantly impact benchmarking and the usability of downstream applications for BLV users [31, 50].”

“This dual evaluation approach overcomes the systematic bias that reference based metrics exhibit against BLV users’ preferences, as demonstrated by Ka-pur and Kreiss (Kapur and Kreiss, 2024).”

“As such, accessibility researchers within these areas have begun to develop various approaches to disability-centered model evaluation that involve prompting [47, 103], metric assessment [67], interviews [9, 120], and more.”

“Widely adopted evaluation metrics have been shown to be biased against their preferences (Kapur and Kreiss, 2024) and benchmark studies tend to pursue a larger audience first (Li et al., 2024a,d).”

“An alternative way is having an automatic BLV-preference-aligned metric (Kapur and Kreiss, 2024) that can be used as a criterion.”

“As such, accessibility researchers within these areas have begun to develop various approaches to disability-centered model evaluation that involve prompting [47, 103], metric assessment [67], interviews [9, 120], and more.”

“Reference-based metrics like BLEU (Papineni et al., 2002), ROUGE (Lin, 2004), and METEOR (Banerjee and Lavie, 2005) primarily assess similarity to human-written references but fail to capture human judgments of distinction (Kapur and Kreiss, 2024).”

“These recommendations build on previous works showing that BVI users value conversational interfaces [15] that automatic metrics do not necessarily reflect their preferences [25].”

“Standard evaluation metrics frequently fail to detect these harms, revealing bias against blind and low-vision users and disproportionate impacts on intersecting marginalized identities [56, 71], especially in domains such as healthcare and finance [70].”

“Text-to-image and image captioning models often default to narrow visual stereotypes, such as depicting people with disabilities as elderly or wheelchair-bound irrespective of context [ Kapur and Kreiss, 2024 ] .”