First draft of flexible character accuracy

qurator/dinglehopper/flexible_character_accuracy.py

@@ -0,0 +1,394 @@
+"""
+Implementation of the flexible character accuracy
+
+Citation:
+    Flexible character accuracy measure for reading-order-independent evaluation
+    C. Clausner, S. Pletschacher, A. Antonacopoulos
+    Pattern Recognition Letters, Volume 131, March 2020, Pages 390-397
+Link: http://www.primaresearch.org/publications/PRL_Clausner_FlexibleCharacterAccuracy
+DOI: https://doi.org/10.1016/j.patrec.2020.02.003
+
+Note that we deviated from the original algorithm at some places.
+"""
+
+from collections import Counter
+from functools import lru_cache, reduce
+from itertools import product, takewhile
+from typing import List, NamedTuple, Tuple, Optional
+
+from . import editops
+
+
+def flexible_character_accuracy(gt: str, ocr: str) -> Tuple[float, List["Match"]]:
+    """Calculate the flexible character accuracy.
+
+    Reference: contains steps 1-7 of the flexible character accuracy algorithm.
+
+    :param gt: The ground truth text.
+    :param ocr: The text to compare the ground truth with.
+    :return: Score between 0 and 1 and match objects.
+    """
+
+    best_score = -float('inf')
+    best_matches = []
+    # TODO: this should be configurable
+    combinations = product(range(15, 31, 5),
+                           range(0, 24, 3),
+                           range(0, 4, 1),
+                           range(0, 6, 1))
+    # TODO: place to parallelize the algorithm
+    for (edit_dist, length_diff, offset, length) in combinations:
+        coef = Coefficients(
+            edit_dist=edit_dist,
+            length_diff=length_diff,
+            offset=offset,
+            length=length
+        )
+        # Steps 1 - 6 of the flexible character accuracy algorithm.
+        matches = match_with_coefficients(gt, ocr, coef)
+        # Step 7 of the flexible character accuracy algorithm.
+        score = character_accuracy_for_matches(matches)
+        if score > best_score:
+            best_score = score
+            best_matches = matches
+        # early breaking: we only need one perfect fit
+        if best_score >= 1:
+            break
+    return best_score, best_matches
+
+
+def match_with_coefficients(gt: str, ocr: str, coef: "Coefficients") -> List["Match"]:
+    """Match ground truth with ocr and considers a given set of coefficients.
+
+    Reference: contains steps 1 - 6 of the flexible character accuracy algorithm.
+
+    :return: A list of match objects to score and align the texts.
+    """
+    # Steps 1 and 2 of the flexible character accuracy algorithm.
+    ocr_lines = initialize_lines(ocr)
+    gt_lines = initialize_lines(gt)
+
+    matches = []
+
+    # Step 5 of the flexible character accuracy algorithm.
+    while len(gt_lines) != 0 and len(ocr_lines) != 0:
+        # Steps 3 and 4 of the flexible character accuracy algorithm.
+        match = match_longest_gt_lines(gt_lines, ocr_lines, coef)
+        if match:
+            matches.append(match)
+
+    # Step 6 of the flexible character accuracy algorithm.
+    # remaining lines are considered as deletes and inserts
+    deletes = [distance(line, Part(text="", line=line.line, start=line.start))
+               for line in gt_lines]
+    inserts = [distance(Part(text="", line=line.line, start=line.start), line)
+               for line in ocr_lines]
+
+    return [*matches, *deletes, *inserts]
+
+
+def match_longest_gt_lines(gt_lines: List["Part"],
+                           ocr_lines: List["Part"],
+                           coef: "Coefficients") -> Optional["Match"]:
+    """Find the best match for the longest line(s) in ground truth.
+
+    The longest lines in ground truth are matched against lines in ocr to find the
+    best matching pair. This pair is then either considered a match on full line
+
+    Reference: contains steps 3 and 4 of the flexible character accuracy algorithm.
+
+    :return: Possible match object.
+    """
+    best_score, best_match, best_gt, best_ocr = -float('inf'), None, None, None
+    if not ocr_lines:
+        return best_match
+
+    # Step 3 of the flexible character accuracy algorithm (variation).
+    # Instead of the longest line we take all longest lines with equal length.
+    length = min(gt_lines[0].length, ocr_lines[0].length)
+    for gt_line in takewhile(lambda line: line.length >= length, gt_lines):
+        match, ocr_line = match_gt_line(gt_line, ocr_lines, coef)
+        score = 0 if not match else character_accuracy(match.dist)
+        if score > best_score:
+            best_score, best_match, best_gt, best_ocr = score, match, gt_line, ocr_line
+
+    # Step 4 of the flexible character accuracy algorithm.
+    # Remove on full match or split.
+    if best_match and best_gt:
+        splitted = remove_or_split(best_gt, best_match.gt, gt_lines)
+        if splitted:
+            gt_lines.append(best_match.gt)
+            best_match = None
+    if best_match and best_ocr:
+        remove_or_split(best_ocr, best_match.ocr, ocr_lines)
+
+    return best_match
+
+
+def match_gt_line(gt_line: "Part",
+                  ocr_lines: List["Part"],
+                  coef: "Coefficients") -> Tuple[Optional["Match"],
+                                                 Optional["Part"]]:
+    """Match the given ground truth line against all the lines in ocr.
+
+    Reference: contains steps 3 of the flexible character accuracy algorithm.
+
+    TODO: Make penalty function configurable?
+    TODO: Add empty ocr line to avoid having nonesense one character alignments?
+
+    :return: Match object and the matched ocr line.
+    """
+    min_penalty = float('inf')
+    best_match, best_ocr = None, None
+    for ocr_line in ocr_lines:
+        match = match_lines(gt_line, ocr_line)
+        penalty = calculate_penalty(gt_line, ocr_line, match, coef)
+        if penalty < min_penalty:
+            min_penalty, best_match, best_ocr = penalty, match, ocr_line
+    return best_match, best_ocr
+
+
+def remove_or_split(original: "Part",
+                    match: "Part",
+                    lines: List["Part"]) -> bool:
+    """Removes the matched line or splits it into parts.
+
+    Reference: contains step 4 of the flexible character accuracy algorithm.
+
+    :return: True if line was splitted.
+    """
+    splitted = False
+    del lines[lines.index(original)]
+    if match.length < original.length:
+        lines.extend(original.split(match))
+        # sorting for ocr is not mentioned in the paper, but is used as tie breaking =)
+        lines.sort(key=lambda x: x.length, reverse=True)
+        splitted = True
+    return splitted
+
+
+@lru_cache(maxsize=1000000)
+def match_lines(gt_line: "Part", ocr_line: "Part") -> Optional["Match"]:
+    """Matches two lines searching for a local alignment.
+
+    The shorter line is moved along the longer line
+    until the editing distance is minimized.
+
+    Reference: see figure 2 in the paper.
+
+    TODO: make distance function configurable?
+
+    :return: Match object if one is found.
+    """
+    min_length = min(gt_line.length, ocr_line.length)
+    best_match = None
+    if min_length == 0:
+        return best_match
+    length_diff = gt_line.length - ocr_line.length
+    min_edit_dist = float('inf')
+    # TODO: handle deletes and replacements by extending the length.
+    for i in range(0, max(1, length_diff + 1)):
+        for j in range(0, max(1, -1 * length_diff + 1)):
+            match = distance(gt_line.substring(rel_start=i, rel_end=i + min_length),
+                             ocr_line.substring(rel_start=j, rel_end=j + min_length))
+            edit_dist = score_edit_distance(match)
+            if edit_dist < min_edit_dist:
+                min_edit_dist = edit_dist
+                best_match = match
+    return best_match
+
+
+@lru_cache(maxsize=1000000)
+def distance(gt: "Part", ocr: "Part") -> "Match":
+    """Calculate the editing distance between the two lines.
+
+    Using the already available `editops()` function with the Levenshtein distance.
+
+    TODO: replace with @cache annotation in Python 3.9
+
+    :return: Match object containing the lines and the editing operations.
+    """
+    ops = editops(gt.text, ocr.text)
+    edits = Counter([edit[0] for edit in ops])
+    edits["match"] = gt.length - edits["delete"] - edits["replace"]
+    return Match(gt=gt, ocr=ocr, dist=Distance(**edits), ops=ops)
+
+
+def score_edit_distance(match: "Match") -> int:
+    """Calculate edit distance for a match.
+
+    Formula: $deletes + inserts + 2 * replacements$
+
+    :return: Sum of deletes, inserts and replacements.
+    """
+    return match.dist.delete + match.dist.insert + 2 * match.dist.replace
+
+
+def calculate_penalty(gt: "Part", ocr: "Part", match: "Match",
+                      coef: "Coefficients") -> float:
+    """Calculate the penalty for a given match.
+
+    For details and discussion see Section 3 in doi:10.1016/j.patrec.2020.02.003.
+
+    :return: Penalty for the given match.
+    """
+    min_edit_dist = score_edit_distance(match)
+    length_diff = abs(gt.length - ocr.length)
+    substring_length = min(gt.length, ocr.length)
+    offset = 0.0
+    if length_diff > 1:
+        substring_pos = max(match.gt.start - gt.start, match.ocr.start - ocr.start)
+        offset = length_diff / 2 - abs(substring_pos - length_diff / 2)
+    return (min_edit_dist * coef.edit_dist
+            + length_diff * coef.length_diff
+            + offset * coef.offset
+            - substring_length * coef.length)
+
+
+def character_accuracy_for_matches(matches: List["Match"]) -> float:
+    """Character accuracy of a full text represented by a list of matches.
+
+    See other `character_accuracy` for details.
+
+    """
+    agg: Counter = reduce(lambda acc, match: acc + Counter(match.dist._asdict()),
+                          matches, Counter())
+
+    score = character_accuracy(Distance(**agg))
+    return score
+
+
+def character_accuracy(edits: "Distance") -> float:
+    """Character accuracy calculated by necessary edit operations.
+
+    Edit operations are needed edits to transform one text into another.
+
+    The character accuracy is given by $1 - errors / characters$.
+
+    Errors are replacements, deletes and inserts.
+
+    Note that is is possible to have more errors than characters in which case the
+    character accuracy turns negative.
+
+    Comparing two empty strings (having no edits) results in a character accuracy of 1.
+    """
+    errors = edits.replace + edits.delete + edits.insert
+    chars = edits.match + edits.replace + edits.delete
+    if not chars and not errors:
+        # comparison of empty strings is considered a full match
+        score = 1
+    else:
+        score = 1 - errors / chars
+    return score
+
+
+def initialize_lines(text: str) -> List["Part"]:
+    """Splits a text into lines and converts them to our line data object.
+
+    The line objects are sorted by their length descending.
+
+    Reference: contains steps 1 and 2 of the flexible character accuracy algorithm.
+
+    :param text: Text to split into lines.
+    :return: List of sorted line objects.
+    """
+    lines = [Part(text=line, line=i, start=0)
+             for i, line in enumerate(text.splitlines())
+             if len(line) > 0]
+    lines.sort(key=lambda x: x.length, reverse=True)
+    return lines
+
+
+def combine_lines(matches: List["Match"]) -> Tuple[str, str]:
+    """Combines the matches to aligned texts.
+
+    TODO: just hacked, needs tests and refinement. Also missing insert/delete marking.
+
+    :param matches: List of match objects.
+    :return: the aligned ground truth and ocr as texts.
+    """
+    matches.sort(key=lambda x: x.gt.line + x.gt.start / 10000)
+    line = 0
+    gt, ocr = "", ""
+    for match in matches:
+        if match.gt.line > line:
+            gt += "\n"
+            ocr += "\n"
+            line += 1
+        gt += match.gt.text
+        ocr += match.ocr.text
+    return gt, ocr
+
+
+class Part(NamedTuple):
+    """Represent a line or part of a line.
+
+    This data object is maintained to be able to reproduce the original text.
+    """
+    text: str = ""
+    line: int = 0
+    start: int = 0
+
+    @property
+    def end(self) -> int:
+        return self.start + self.length
+
+    @property
+    def length(self) -> int:
+        return len(self.text)
+
+    def split(self, split: "Part") -> List["Part"]:
+        """Split the line part by another and returns the remaining parts.
+
+        `abc.split("b")` will return ´["a", "c"]`.
+
+        :param split: The line part we want to use to split.
+        :return: The parts before and after the split.
+        """
+        rest = []
+        if self.start < split.start:
+            rest.append(self.substring(rel_end=split.start - self.start))
+        if split.end < self.end:
+            rest.append(self.substring(rel_start=split.end - self.start))
+        return rest
+
+    def substring(self, rel_start: int = 0, rel_end: int = None) -> "Part":
+        """Get part of the given line.
+
+        Automatically handles the offset of the line.
+        Therefore `substring(rel_start=2)` will return `Part[start+rel_start:]`.
+
+        :param rel_start: start relative to the part of the line.
+        :param rel_end: end relative to the part of the line.
+        :return: Extracted part of the given part of the line.
+        """
+        text = self.text[rel_start:rel_end]
+        start = self.start + rel_start
+        return Part(text=text, line=self.line, start=start)
+
+
+class Distance(NamedTuple):
+    """Represent distance between two sequences."""
+    match: int = 0
+    replace: int = 0
+    delete: int = 0
+    insert: int = 0
+
+
+class Match(NamedTuple):
+    """Represent a calculated match between ground truth and the ocr result."""
+    gt: "Part"
+    ocr: "Part"
+    dist: "Distance"
+    ops: List
+
+
+class Coefficients(NamedTuple):
+    """Coefficients to calculate penalty for substrings.
+
+    See Section 3 in doi:10.1016/j.patrec.2020.02.003
+    """
+    edit_dist: int = 25
+    length_diff: int = 20
+    offset: int = 1
+    length: int = 4

qurator/dinglehopper/tests/test_flexible_character_accuracy.py

@@ -0,0 +1,291 @@
+"""
+Tests for the implementation of the flexible character accuracy
+
+Citation:
+    Flexible character accuracy measure for reading-order-independent evaluation
+    C. Clausner, S. Pletschacher, A. Antonacopoulos
+    Pattern Recognition Letters, Volume 131, March 2020, Pages 390-397
+Link: http://www.primaresearch.org/publications/PRL_Clausner_FlexibleCharacterAccuracy
+DOI: 10.1016/j.patrec.2020.02.003
+"""
+
+import pytest
+
+from ..flexible_character_accuracy import *
+
+CASE_ARGS = "gt,ocr,first_line_score,all_line_score"
+
+SIMPLE_CASES = [
+    ("a", "", 0, 0),
+    ("a", "a", 1, 1),
+    ("a\nb", "a\nb", 1, 1),
+    ("a\nb", "b\na", 1, 1),
+    ("aaa\nbbb\nccc", "ccc\naaa\nbbb", 1, 1),
+    ("aaa\nbbb\nccc", "aaa\nbbb", 1, 1 - 3 / 9),
+    ("bbb", "aaa\nbbb\nccc", 1, 1 - 6 / 3),
+    ("a", "a\nbb\nccc", 1, 1 - 5 / 1),
+    ("bb", "a\nbb\nccc", 1, 1 - 4 / 2),
+]
+
+COMPLEX_CASES = [
+    ("accc", "a\nbb\nccc", 0, 1 - 2 / 4),
+    ("aaa\nbbb\nccc", "bbb", 1, 1 - 6 / 9),
+]
+
+EXTENDED_CASES = [
+    # A: No errors
+    ((0, 1, 2, 3, 4, 11, 5, 6, 7, 8, 9),
+     (0, 1, 2, 3, 4, 11, 5, 6, 7, 8, 9),
+     1, 1),
+    # B: Different ordering of text blocks
+    ((0, 1, 2, 3, 4, 11, 5, 6, 7, 8, 9),
+     (5, 6, 7, 8, 9, 11, 0, 1, 2, 3, 4),
+     1, 1),
+    # C: Merge across columns
+    ((0, 1, 2, 11, 3, 4, 11, 5, 6, 7, 11, 8, 9),
+     (0, 1, 2, 5, 6, 7, 11, 3, 4, 8, 9),
+     1, 0.964),
+    # D: Over-segmentation
+    ((0, 1, 2, 3, 4, 11, 5, 6, 7, 8, 9),
+     (0, 1, 2, 11, 5, 6, 7, 11, 3, 4, 11, 8, 9),
+     1, 0.966),
+    # E: Part missing
+    ((0, 1, 2, 3, 4, 11, 5, 6, 7, 8, 9),
+     (0, 1, 2, 3, 4),
+     1, 0.50),
+    # E.2: Part missing
+    ((0, 1, 2, 3, 4, 11, 5, 6, 7, 8, 9),
+     (5, 6, 7, 8, 9),
+     1, 0.50),
+    # F: All missing
+    ((0, 1, 2, 3, 4, 11, 5, 6, 7, 8, 9),
+     (),
+     1, 0),
+    # G: Added parts
+    ((0, 1, 2, 3, 4),
+     (0, 1, 2, 3, 4, 11, 5, 6),
+     1, 0.621),
+]
+
+EDIT_ARGS = "gt,ocr,expected_dist"
+
+SIMPLE_EDITS = [
+    (Part(text="a").substring(), Part(text="a"), Distance(match=1)),
+    (Part(text="a").substring(), Part(text="b"), Distance(replace=1)),
+    (Part(text="abcd").substring(), Part(text="beed"),
+     Distance(match=2, replace=1, insert=1, delete=1)),
+]
+
+
+def extended_case_to_text(gt, ocr):
+    sentence = ("Eight", "happy", "frogs", "scuba", "dived",
+                "Jenny", "chick", "flaps", "white", "wings",
+                "", "\n")
+
+    gt_sentence = " ".join(sentence[i] for i in gt).replace(" \n ", "\n")
+    ocr_sentence = " ".join(sentence[i] for i in ocr).replace(" \n ", "\n")
+    return gt_sentence, ocr_sentence
+
+
+@pytest.mark.parametrize(CASE_ARGS, [*SIMPLE_CASES, *COMPLEX_CASES])
+def test_flexible_character_accuracy_simple(gt, ocr, first_line_score, all_line_score):
+    score, _ = flexible_character_accuracy(gt, ocr)
+    assert score == pytest.approx(all_line_score)
+
+
+@pytest.mark.xfail
+@pytest.mark.parametrize("ocr", [
+    "1 hav\nnospecial\ntalents.\nI am one\npassionate\ncuriousity.\"\nAlberto\nEmstein",
+    "1 hav\nnospecial\ntalents. Alberto\nI am one Emstein\npassionate\ncuriousity.\"",
+    "Alberto\nEmstein\n1 hav\nnospecial\ntalents.\nI am one\npassionate\ncuriousity.\""
+])
+def test_flexible_character_accuracy(ocr):
+    """Tests from figure 3 in the paper.
+
+    TODO: We have a 2 percent deviation from the original because of redistributed
+          one character alignments (e.g. the y-insert replaces the y-delete).
+    """
+    gt = """"I have
+no special
+talent.
+I am only
+passionately
+curious."
+Albert
+Einstein
+"""
+    replacements = 3
+    inserts = 5
+    deletes = 7
+    chars = len(gt) - gt.count("\n")
+    assert replacements + inserts + deletes == 15
+    edits = Distance(match=chars - deletes - replacements, replace=replacements,
+                     insert=inserts, delete=deletes)
+    expected = character_accuracy(edits)
+    assert expected == pytest.approx(0.779, abs=0.0005)
+    result, matches = flexible_character_accuracy(gt, ocr)
+    assert result == pytest.approx(expected, abs=0.0005)
+
+
+@pytest.mark.parametrize(CASE_ARGS, EXTENDED_CASES)
+def test_flexible_character_accuracy_extended(gt, ocr, first_line_score,
+                                              all_line_score):
+    """Tests from figure 4 in the paper."""
+    gt_sentence, ocr_sentence = extended_case_to_text(gt, ocr)
+    result, _ = flexible_character_accuracy(gt_sentence, ocr_sentence)
+    assert result == pytest.approx(all_line_score, abs=0.001)
+
+
+@pytest.mark.parametrize(CASE_ARGS, [*SIMPLE_CASES, *COMPLEX_CASES, *EXTENDED_CASES])
+def test_match_with_coefficients(gt, ocr, first_line_score, all_line_score):
+    coef = Coefficients()
+    if not isinstance(gt, str):
+        gt, ocr = extended_case_to_text(gt, ocr)
+    matches = match_with_coefficients(gt, ocr, coef)
+    score = character_accuracy_for_matches(matches)
+    assert score == pytest.approx(all_line_score, abs=0.001)
+
+
+@pytest.mark.parametrize(CASE_ARGS, [*SIMPLE_CASES, *COMPLEX_CASES])
+def test_match_longest_gt_lines(gt, ocr, first_line_score, all_line_score):
+    coef = Coefficients()
+    gt_lines = initialize_lines(gt)
+    ocr_lines = initialize_lines(ocr)
+    match = match_longest_gt_lines(gt_lines, ocr_lines, coef)
+    score = 0
+    if match:
+        score = character_accuracy(match.dist)
+    assert score == pytest.approx(first_line_score)
+
+
+@pytest.mark.parametrize(CASE_ARGS, [
+    *SIMPLE_CASES,
+    ("accc", "a\nbb\nccc", 1.0, 1.0),
+])
+def test_match_gt_line(gt, ocr, first_line_score, all_line_score):
+    coef = Coefficients()
+    gt_lines = initialize_lines(gt)
+    ocr_lines = initialize_lines(ocr)
+    match, _ = match_gt_line(gt_lines[0], ocr_lines, coef)
+    score = 0
+    if match:
+        score = character_accuracy(match.dist)
+    assert score == pytest.approx(first_line_score)
+
+
+@pytest.mark.parametrize("original,match,expected_lines", [
+    (Part(), Part(), []),
+    (Part(text="abc"), Part(), [Part(text="abc")]),
+    (Part(text="abc"), Part("d"), [Part(text="bc", start=1)]),
+    (Part(text="abc"), Part("a", start=100), [Part(text="abc")]),
+    (Part(text="abc"), Part("a"), [Part(text="bc", start=1)]),
+    (Part(text="abc"), Part("b", start=1), [Part(text="a"), Part(text="c", start=2)]),
+    (Part(text="abc"), Part("c", start=2), [Part(text="ab")]),
+])
+def test_remove_or_split(original, match, expected_lines):
+    lines = [original]
+    splitted = remove_or_split(original, match, lines)
+    assert splitted == (len(lines) > 0)
+    assert lines == expected_lines
+
+
+@pytest.mark.parametrize(EDIT_ARGS, [
+    *SIMPLE_EDITS,
+    (Part(text="aaabbbaaa"), Part(text="bbb"), Distance(match=3)),
+    (Part(text="bbb"), Part(text="aaabbbaaa"), Distance(match=3)),
+    (Part(text=""), Part(text=""), None)
+])
+def test_match_lines(gt, ocr, expected_dist):
+    match = match_lines(gt, ocr)
+    if not expected_dist:
+        assert match is None
+    else:
+        assert match.gt.text in gt.text
+        assert match.ocr.text in ocr.text
+        assert match.dist == expected_dist
+
+
+@pytest.mark.parametrize(EDIT_ARGS, [
+    *SIMPLE_EDITS,
+    (Part(text="").substring(), Part(text=""), Distance()),
+    (Part(text="ab").substring(), Part("a"), Distance(match=1, delete=1)),
+    (Part(text="a").substring(), Part("ab"), Distance(match=1, insert=1)),
+])
+def test_distance(gt, ocr, expected_dist):
+    match = distance(gt, ocr)
+    assert match.gt == gt
+    assert match.ocr == ocr
+    assert match.dist == expected_dist
+
+
+@pytest.mark.parametrize("matches,expected_dist", [
+    ([], 1),
+    ([Match(gt=Part(text=""), ocr=Part(text=""), dist=Distance(), ops=[])], 1),
+    ([Match(gt=Part(text="abee"), ocr=Part("ac"),
+            dist=Distance(match=1, replace=1, delete=2), ops=[]),
+      Match(gt=Part(text="cd"), ocr=Part("ceff"),
+            dist=Distance(match=1, replace=1, insert=2), ops=[])],
+     1 - 6 / 6),
+])
+def test_character_accuracy_matches(matches, expected_dist):
+    assert character_accuracy_for_matches(matches) == pytest.approx(expected_dist)
+
+
+@pytest.mark.parametrize("dist,expected_dist", [
+    (Distance(), 1),
+    (Distance(match=1), 1),
+    (Distance(replace=1), 0),
+    (Distance(match=1, insert=1), 0),
+    (Distance(match=1, insert=2), 1 - 2 / 1),
+    (Distance(match=2, insert=1), 0.5),
+    (Distance(match=1, delete=1), 0.5),
+])
+def test_character_accuracy_dist(dist, expected_dist):
+    assert character_accuracy(dist) == pytest.approx(expected_dist)
+
+
+@pytest.mark.parametrize("line,subline,expected_rest", [
+    (Part(), Part(), []),
+    (Part("aaa bbb"), Part("aaa bbb"), []),
+    (Part("aaa bbb"), Part("aaa"), [Part(" bbb", start=3)]),
+    (Part("aaa bbb"), Part("bbb", start=4), [Part("aaa ")]),
+    (Part("aaa bbb", start=3), Part("aaa", start=3), [Part(" bbb", start=6)]),
+    (Part("aaa bbb", start=3), Part("bbb", start=7), [Part("aaa ", start=3)]),
+    (Part("aaa bbb ccc"), Part("bbb", start=4), [Part("aaa "), Part(" ccc", start=7)]),
+    (Part("aaa bbb ccc", start=3), Part("bbb", start=7),
+     [Part("aaa ", start=3), Part(" ccc", start=10)]),
+    (Part("aaa bbb"), Part(" ", start=3), [Part("aaa"), Part("bbb", start=4)]),
+    (Part("aaa bbb", start=3), Part(" ", start=6),
+     [Part("aaa", start=3), Part("bbb", start=7)]),
+])
+def test_split_line(line, subline, expected_rest):
+    rest = line.split(subline)
+    assert len(rest) == len(expected_rest)
+    assert set(rest) == set(expected_rest)
+
+
+def test_initialize_lines():
+    lines = initialize_lines("")
+    assert lines == []
+
+    lines = initialize_lines("22\n1\n333")
+    line1 = Part(text="22", line=0, start=0)
+    line2 = Part("1", line=1, start=0)
+    line3 = Part("333", line=2, start=0)
+    assert lines == [line3, line1, line2]
+
+
+@pytest.mark.xfail
+def test_combine_lines():
+    assert False
+
+
+@pytest.mark.parametrize("line,start,end,expected", [
+    (Part(text=""), 0, None, Part(text="")),
+    (Part(text="a"), 0, None, Part(text="a")),
+    (Part(text="ab"), 0, 1, Part(text="a")),
+    (Part(text="abc"), 0, -1, Part(text="ab")),
+    (Part(text="ab"), 1, None, Part(text="b", start=1)),
+])
+def test_line_substring(line, start, end, expected):
+    assert line.substring(rel_start=start, rel_end=end) == expected