method: TextMind + ERNIE-Layout2021-09-06

Authors: TextMind + ERNIE-Layout

Affiliation: Baidu

Description: 1. TextMind provides enhanced layout information.
2. Apply such information for layout enhanced pre-training based on ERNIE.
3. Following the same evaluation protocol as Applica.ai.

method: Linklogis_BeeAI2021-07-20

Authors: Juntao Zhang, Donglai Chen, Yonghong Chen,Fangchao Ji, Jiayu Chen,Jinyuan He,Zhiqi zhu, Yan Xu

Affiliation: Linklogis_BeeAI Team

Description: Our model is developed based on a multimodal framework that is capable of extracting both layout information and textual semantics from scanned receipts. Following the same evaluation protocol as others, we remove any OCR mismatches and fix any discrepancies found in total amount randomly prefixed by *RM.

Authors: Applica.ai research team

Affiliation: Applica.ai

Description: Following the same evaluation rules as others, the OCR mismatch errors are excluded in the submission.

Additionally, we have manually fixed annotation discrepancies in "total" entity in the test set.

Note:
1. We submitted the best solution out of 100 fine-tuned models
2. In this task there is an annotation discrepancy in "total" entity which caused unfair comparison between models (In train/test sets "total" entity was randomly prefixed by "RM"). Number of errors in the top solutions caused by this kind of annotation error:

Applica.ai Lambert 2.0 + Excluding OCR Errors + Fixing total entity = 0
LayoutLM 2.0 (single model) = 3 (example: 275)
Applica.ai Lambert 2.0 + Excluding OCR Mismatch = 8 (example: 77)
Tencent Youtu = 8 (example: 120)
VIE = 0
HIK_OCR_Exclude_ocr_mismatch = 0
LayoutLM + Excluding OCR Mismatch = 9 (example: 121)

Ranking Table

Description Paper Source Code
DateMethodRecallPrecisionHmean
2021-09-06TextMind + ERNIE-Layout97.26%99.48%98.36%
2021-07-20 Linklogis_BeeAI97.05%99.34%98.18%
2021-01-02Applica.ai Lambert 2.0 + Excluding OCR Errors + Fixing total entity96.83%99.56%98.17%
2021-06-02Multimodal Transformer for Information Extraction96.76%99.56%98.14%
2021-02-16Applica.ai TILT + Excluding OCR Errors + Fixing total entity96.83%99.41%98.10%
2021-01-01Applica.ai Lambert 2.0 + Excluding OCR Mismatch96.40%99.11%97.74%
2020-12-07Tencent Youtu96.47%98.89%97.67%
2020-05-07HIK_OCR_Exclude_ocr_mismatch96.33%98.38%97.34%
2020-04-18LayoutLM + Excluding OCR Mismatch96.04%98.16%97.09%
2020-04-15PICK-PAPCIC & XZMU95.46%96.79%96.12%
2020-03-26Applica.ai roberta-base-2D95.39%95.80%95.60%
2019-08-14PATech_AICenter94.02%94.02%94.02%
2021-02-21RoBERTa-base finetuned on business documents92.80%93.27%93.03%
2021-02-21RoBERTa-base92.22%92.55%92.39%
2020-07-07Taikang Insurance Group Research Institute91.79%91.99%91.89%
2019-08-05PATECH_CHENGDU_OCR_V291.21%91.21%91.21%
2020-02-20Character & Word BiLSTM Encoder90.85%90.85%90.85%
2021-09-13JENTI-baseline88.47%92.75%90.56%
2019-05-05Ping An Property & Casualty Insurance Company90.49%90.49%90.49%
2019-04-29Enetity detection89.70%89.70%89.70%
2019-05-04H&H Lab89.63%89.63%89.63%
2019-05-02CLOVA OCR89.05%89.05%89.05%
2019-05-04HeReceipt-withoutRM83.00%83.24%83.12%
2019-05-06BOE_IOT_AIBD82.71%82.71%82.71%
2019-05-05PATECH_CHENGDU_OCR81.70%82.29%82.00%
2020-05-28SROIE LSTM - Axel Alejandro Ramos García81.99%81.99%81.99%
2020-04-28BERT-MRC81.05%81.05%81.05%
2019-04-30NER with spaCy model78.96%79.02%78.99%
2020-12-28Custom Named Entity Recognition77.59%77.59%77.59%
2019-05-05CITlab Argus Information Extraction (positional & line features, enhanced gt)77.38%77.38%77.38%
2019-04-28A Simple Method for Key Information Extraction as Character-wise Classification with LSTM75.58%75.58%75.58%
2019-05-05Location-aware BERT model for Text Information Extraction74.42%74.42%74.42%
2019-04-30BERT with Multi-task Confidence Prediction66.14%66.14%66.14%
2019-05-02With receipt framing63.04%63.54%63.29%
2019-05-05IFLYTEK-textNLP_v261.24%61.24%61.24%
2019-05-05SituTech_OCR59.01%62.38%60.64%
2019-04-30Key Information Extraction from Scanned Receipts28.75%36.31%32.09%

Ranking Graphic