NeoAnd/Docling
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: Establish confidence estimation for document and pages (#1313)

Browse Source 
* Establish confidence field, propagate layout confidence through

Signed-off-by: Christoph Auer <cau@zurich.ibm.com>

* Add OCR confidence and parse confidence (stub)

Signed-off-by: Christoph Auer <cau@zurich.ibm.com>

* Add parse quality rules, use 5% percentile for overall and parse scores

Signed-off-by: Christoph Auer <cau@zurich.ibm.com>

* Heuristic updates

Signed-off-by: Christoph Auer <cau@zurich.ibm.com>

* Fix garbage regex

Signed-off-by: Christoph Auer <cau@zurich.ibm.com>

* Move grade to page

Signed-off-by: Christoph Auer <cau@zurich.ibm.com>

* Introduce mean_score and low_score, consistent aggregate computations

Signed-off-by: Christoph Auer <cau@zurich.ibm.com>

* Add confidence test

Signed-off-by: Christoph Auer <cau@zurich.ibm.com>

---------

Signed-off-by: Christoph Auer <cau@zurich.ibm.com>
This commit is contained in:
Christoph Auer
2025-05-21 12:32:49 +02:00
committed by GitHub
parent 14d4f5b109
commit 90875247e5
7 changed files with 199 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

51
docling/models/page_preprocessing_model.py
View File

@@ -1,11 +1,13 @@
import re
from collections.abc import Iterable
from pathlib import Path
from typing import Optional
import numpy as np
from PIL import ImageDraw
from pydantic import BaseModel
from docling.datamodel.base_models import Page
from docling.datamodel.base_models import Page, ScoreValue
from docling.datamodel.document import ConversionResult
from docling.datamodel.settings import settings
from docling.models.base_model import BasePageModel
@@ -21,6 +23,14 @@ class PagePreprocessingModel(BasePageModel):
def __init__(self, options: PagePreprocessingOptions):
self.options = options
# Pre-compiled regex patterns for efficiency
self.GLYPH_RE = re.compile(r"GLYPH<[0-9A-Fa-f]+>")
self.SLASH_G_RE = re.compile(r"(?:/G\d+){2,}")
self.FRAG_RE = re.compile(r"\b[A-Za-z](?:/[a-z]{1,3}\.[a-z]{1,3}){2,}\b")
self.SLASH_NUMBER_GARBAGE_RE = re.compile(
r"(?:/\w+\s*){2,}"
) # Two or more "/token " sequences
def __call__(
self, conv_res: ConversionResult, page_batch: Iterable[Page]
) -> Iterable[Page]:
@@ -60,6 +70,18 @@ class PagePreprocessingModel(BasePageModel):
if self.options.create_parsed_page:
page.parsed_page = page._backend.get_segmented_page()
# Rate the text quality from the PDF parser, and aggregate on page
text_scores = []
for c in page.cells:
score = self.rate_text_quality(c.text)
text_scores.append(score)
conv_res.confidence.pages[page.page_no].parse_score = float(
np.nanquantile(
text_scores, q=0.10
) # To emphasise problems in the parse_score, we take the 10% percentile score of all text cells.
)
# DEBUG code:
def draw_text_boxes(image, cells, show: bool = False):
draw = ImageDraw.Draw(image)
@@ -88,3 +110,30 @@ class PagePreprocessingModel(BasePageModel):
draw_text_boxes(page.get_image(scale=1.0), page.cells)
return page
def rate_text_quality(self, text: str) -> float:
# Hard errors: if any of these patterns are found, return 0.0 immediately.
blacklist_chars = ["<EFBFBD>"]
if (
any(text.find(c) >= 0 for c in blacklist_chars)
or self.GLYPH_RE.search(text)
or self.SLASH_G_RE.search(text)
or self.SLASH_NUMBER_GARBAGE_RE.match(
text
) # Check if text is mostly slash-number pattern
):
return 0.0
penalty = 0.0
# Apply a penalty only if the fragmented words pattern occurs at least three times.
frag_matches = self.FRAG_RE.findall(text)
if len(frag_matches) >= 3:
penalty += 0.1 * len(frag_matches)
# Additional heuristic: if the average token length is below 2, add a penalty.
# tokens = text.split()
# if tokens and (sum(map(len, tokens)) / len(tokens)) < 2:
# penalty += 0.2
return max(1.0 - penalty, 0.0)