OCR and Multimodal Extraction — References

Recognition accuracy, its measurement, and the evidence about what multimodal models get wrong when asked to read a figure.

Recognition and its measurement

• Tesseract OCR — the long-standing open-source engine, and its documentation of confidence output and page-segmentation modes.
• ICDAR Robust Reading Competitions — the ICDAR competition series, running since 2003, which defines the text-detection and recognition protocols and the edit-distance-based scoring that character and word error rates are reported against. Cited by name.
• Unicode Standard Annex #9: Unicode Bidirectional Algorithm — the specification that determines what a correct transcription of mixed-direction text even is, which is a prerequisite for scoring one.

Speech as a retrievable artifact

• Robust Speech Recognition via Large-Scale Weak Supervision — Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey and Ilya Sutskever, ICML, 2023. The weakly supervised multilingual transcription model behind much of the current open tooling, and its reported behaviour on noisy audio. Cited by name.

Reading figures and charts

• ChartQA: A Benchmark for Question Answering about Charts with Visual and Logical Reasoning — Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty and Enamul Hoque, Findings of ACL, 2022. A benchmark specifically for extracting and reasoning over values in charts, and evidence of how far short of transcription this task sits. Cited by name.
• DocVQA: A Dataset for VQA on Document Images — Minesh Mathew, Dimosthenis Karatzas and C. V. Jawahar, WACV, 2021. Question answering over document images, which is the task shape behind using a multimodal model as an extractor. Cited by name.

Why generated descriptions need labelling

• Object Hallucination in Image Captioning — Anna Rohrbach, Lisa Anne Hendricks, Kaylee Burns, Trevor Darrell and Kate Saenko, EMNLP, 2018. The measurement that captioning models assert objects that are not present, which is the reason a generated description cannot be treated as source text. Cited by name.

No verified reference is recorded for text-layer quality scoring, for confidence-banded indexing thresholds, or for selective reprocessing policies under a fixed extraction budget. These are operational practice rather than published results.