Cadena SER and Levante-EMV reported on October 15, 2026, that researchers in the PRHLT group at the Universitat Politècnica de València (UPV) had developed a research system to help identify sexism in TikTok videos and memes. The approach described in the reports combines content analysis with viewers’ eye movements, brain activity and heart-rate responses. (Cadena SER; Levante-EMV)
How the research approaches context
The reports describe a system that considers text and visual information alongside physiological responses recorded as people watched the content. That mix addresses a tricky classification problem: sexism may be wrapped in humor, irony, stereotypes or comments that seem harmless at first. PRHLT researcher Iván Arcos also described a distinction between direct sexism and content that depicts or judges sexist situations, where meaning depends on context and intent. (Cadena SER)
Eye tracking records where viewers look; electroencephalography (EEG) measures electrical activity in the brain. Those measurements describe viewers’ responses, while the system’s task is to classify the content. The distinction matters: a viewer’s reaction is an input to the reported approach, not a classification result by itself.
The content and the viewer studies
The news reports describe a corpus of about 4,000 memes and 2,500 TikTok videos in English and Spanish, drawn from EXIST datasets. Ten people took part in the video study and 16 in the meme study, with eye tracking, EEG and heart rate recorded as they viewed the material. (Cadena SER; Levante-EMV)
Cadena SER reported that the EEG setup used 16 electrodes or channels. The reports also described longer reaction times and more gaze fixations for sexist videos, along with heart-rate fluctuations and response patterns in the meme study. Those are observations about the participants, not a measure of the system’s classification performance. (Cadena SER)
The reports said the work had recently been presented at the XLII International Conference of the Spanish Society for Natural Language Processing.
What the reported improvement figures cover
Levante-EMV reported a 26.3% improvement for the hardest meme category, described as misogyny and non-sexual violence. The outlet also reported an improvement of about 4% in determining whether a video was sexist and more than 6% in distinguishing direct from evaluative sexism when physiological responses were added to text and image inputs. (Levante-EMV)
Those figures apply to different tasks: one concerns a meme category, while the other two concern video classification. They do not describe one overall result for all content types.
What the EXIST 2026 paper adds
A conference paper published by Springer Nature on March 24, 2026, describes six EXIST 2026 tasks covering sexism identification, source-intention detection and sexism categorization across English- and Spanish-language memes and TikTok videos. The paper also states that the UPV Research Ethics Committee approved data collection under Project ID P01_31-10-2025. (Springer Nature)
The paper’s contribution is the challenge framework and its ethics note; the October reports describe the UPV system and the task-specific performance figures attributed to Paolo Rosso.
