这里是差集:Docling 全量解析 text block 减去最终会进入正文的 block。优先人工检查 risk=high / medium。
layout_review.html excluded_blocks.json excluded_blocks.tsv final_body_blocks.tsv visual_assets.tsv original.pdf
{
"parsed_text_blocks": 125,
"final_body_blocks": 36,
"excluded_blocks": {
"count": 89,
"by_reason": {
"after_back_matter_stop": 33,
"docling_page_header": 13,
"body_heading": 9,
"outside_body_flow_caption": 8,
"first_page_metadata": 7,
"outside_body_flow_reference": 5,
"front_matter_heading": 3,
"docling_footnote": 2,
"inside_abstract": 2,
"non_body_heading": 2,
"abstract_heading": 1,
"back_matter_heading": 1,
"front_matter_author_line": 1,
"outside_body_flow": 1,
"outside_body_flow_metadata_line": 1
},
"by_role_guess": {
"reference": 30,
"page_header": 15,
"body_heading": 9,
"caption": 8,
"affiliation": 5,
"footnote": 5,
"front_matter_heading": 4,
"abstract_candidate": 2,
"back_matter_heading": 2,
"metadata": 2,
"page_footer": 2,
"title_candidate": 2,
"abstract_heading": 1,
"back_matter_text": 1,
"unknown_text": 1
},
"by_risk_level": {
"low": 85,
"medium": 3,
"high": 1
},
"high_risk_count": 1,
"medium_risk_count": 3
},
"char_counts": {
"parsed_text_chars": 29561,
"final_body_chars": 21680,
"excluded_chars": 7881
}
}
{
"truncated": true,
"message": "Body extraction stopped at page 7 block #/texts/82: References. 33 following text blocks were excluded as after_back_matter_stop.",
"stop_trigger": {
"ref": "#/texts/82",
"page": 7,
"layout_order": 91,
"role_guess": "back_matter_heading",
"body_decision_reason": "back_matter_heading",
"text_preview": "References"
},
"first_truncated_block": {
"ref": "#/texts/92",
"page": 7,
"layout_order": 92,
"role_guess": "reference",
"body_decision_reason": "after_back_matter_stop",
"text_preview": "X. Liao, Q. Huang, S. Mai, X. Wang, M. Xu, L. Xing, Y. Liao, W. Li, J. Power Sources 272 (2014) 501 e 507."
},
"truncated_block_count": 33,
"truncated_pages": [
7,
8
],
"by_role_guess": {
"reference": 25,
"affiliation": 4,
"page_header": 2,
"back_matter_heading": 1,
"back_matter_text": 1
}
}
这些框显示被排除块,同时叠加真实图表资产框。青色虚线表示该 text block 被图表资产 caption 吸收;红色 STOP 是截断触发点,红色框是截断后被排除的块。