53d0de3d71
Кандидаты уровня 3 (FAISS), которые прошли порог семантической схожести, но LLM не подтвердила заимствование, раньше молча отбрасывались. Теперь это отдельный блок "recommendations" в отчёте — не плагиат, но источники, полезные для раскрытия темы.
195 lines
7.9 KiB
Python
195 lines
7.9 KiB
Python
"""Юнит-тесты скоринга плагиата (итоговый процент, который видит студент)."""
|
||
|
||
from app.scoring import aggregate_results, is_cited
|
||
|
||
|
||
def _m(title: str, pos: int) -> dict:
|
||
return {"source_title": title, "position_start": pos}
|
||
|
||
|
||
def _mp(title: str, start: int, end: int) -> dict:
|
||
return {"source_title": title, "position_start": start, "position_end": end}
|
||
|
||
|
||
def test_empty_input_is_zero():
|
||
out = aggregate_results([], [], [], total_fragments=10)
|
||
assert out["overall_similarity"] == 0.0
|
||
assert out["flagged_fragments"] == 0
|
||
assert out["matches"] == []
|
||
assert out["by_method"] == {"exact": 0, "fuzzy": 0, "semantic_llm": 0}
|
||
|
||
|
||
def test_zero_fragments_does_not_divide_by_zero():
|
||
out = aggregate_results([_m("A", 0)], [], [], total_fragments=0)
|
||
assert out["overall_similarity"] == 0.0
|
||
|
||
|
||
def test_percentage_is_share_of_flagged_positions():
|
||
# 2 из 4 фрагментов помечены → 50%
|
||
level1 = [_m("A", 0), _m("B", 5)]
|
||
out = aggregate_results(level1, [], [], total_fragments=4)
|
||
assert out["overall_similarity"] == 50.0
|
||
assert out["flagged_fragments"] == 2
|
||
|
||
|
||
def test_same_position_multiple_sources_counts_once():
|
||
# Одна позиция совпала с двумя разными источниками — доля не раздувается
|
||
level1 = [_m("A", 0)]
|
||
semantic = [_m("B", 0)] # та же позиция 0, другой источник
|
||
out = aggregate_results(level1, [], semantic, total_fragments=2)
|
||
assert out["flagged_fragments"] == 1
|
||
assert out["overall_similarity"] == 50.0
|
||
# оба совпадения сохранены в matches (дедуп только по паре title:pos)
|
||
assert len(out["matches"]) == 2
|
||
|
||
|
||
def test_exact_duplicate_match_is_deduplicated():
|
||
# Один и тот же источник в одной позиции пришёл дважды — остаётся один
|
||
dup = _m("A", 3)
|
||
out = aggregate_results([dict(dup)], [dict(dup)], [], total_fragments=5)
|
||
assert len(out["matches"]) == 1
|
||
|
||
|
||
def test_similarity_capped_at_100():
|
||
# Больше уникальных позиций, чем фрагментов (пограничный случай) → не выше 100%
|
||
matches = [_m(f"S{i}", i) for i in range(5)]
|
||
out = aggregate_results(matches, [], [], total_fragments=3)
|
||
assert out["overall_similarity"] == 100.0
|
||
|
||
|
||
def test_by_method_counts_raw_matches_per_level():
|
||
out = aggregate_results(
|
||
[_m("A", 0), _m("B", 1)], # exact = 2
|
||
[_m("C", 2)], # fuzzy = 1
|
||
[_m("D", 3), _m("E", 4), _m("F", 5)], # semantic_llm = 3
|
||
total_fragments=10,
|
||
)
|
||
assert out["by_method"] == {"exact": 2, "fuzzy": 1, "semantic_llm": 3}
|
||
|
||
|
||
def test_percentage_rounded_to_two_decimals():
|
||
# 1 из 3 → 33.333... → 33.33
|
||
out = aggregate_results([_m("A", 0)], [], [], total_fragments=3)
|
||
assert out["overall_similarity"] == 33.33
|
||
|
||
|
||
# ─── is_cited ───────────────────────────────────────────────────────────────
|
||
|
||
def test_is_cited_quoted_fragment():
|
||
text = 'Автор пишет: «дословная цитата» и развивает мысль.'
|
||
start = text.index("дословная")
|
||
end = start + len("дословная цитата")
|
||
assert is_cited(text, start, end) is True
|
||
|
||
|
||
def test_is_cited_bracket_citation_with_year():
|
||
text = "Текст фрагмента без кавычек [Иванов, 2023] продолжение."
|
||
end = text.index(" [Иванов")
|
||
assert is_cited(text, 0, end) is True
|
||
|
||
|
||
def test_is_cited_apa_style_citation():
|
||
text = "Some borrowed sentence here (Smith, 2020) continues."
|
||
end = text.index(" (Smith")
|
||
assert is_cited(text, 0, end) is True
|
||
|
||
|
||
def test_is_cited_bare_plagiarism_is_false():
|
||
text = "Просто скопированный текст без всякого оформления далее."
|
||
assert is_cited(text, 0, 10) is False
|
||
|
||
|
||
def test_is_cited_bracket_without_year_is_false():
|
||
text = "Текст фрагмента [см. приложение] продолжение."
|
||
end = text.index(" [см")
|
||
assert is_cited(text, 0, end) is False
|
||
|
||
|
||
def test_is_cited_empty_full_text_is_false():
|
||
assert is_cited("", 0, 5) is False
|
||
|
||
|
||
def test_is_cited_out_of_range_position_is_false():
|
||
assert is_cited("короткий текст", 0, 999) is False
|
||
|
||
|
||
# ─── aggregate_results: cited/uncited split ──────────────────────────────────
|
||
|
||
def test_cited_match_excluded_from_uncited_similarity():
|
||
text = 'Вот «процитированный фрагмент» и текст дальше.'
|
||
start = text.index("процитированный")
|
||
end = start + len("процитированный фрагмент")
|
||
out = aggregate_results([_mp("A", start, end)], [], [], total_fragments=4, full_text=text)
|
||
assert out["matches"][0]["cited"] is True
|
||
assert out["overall_similarity"] == 25.0 # цитата всё ещё учтена в общем %
|
||
assert out["uncited_similarity"] == 0.0 # но не в "некорректных" заимствованиях
|
||
assert out["cited_fragments"] == 1
|
||
assert out["uncited_fragments"] == 0
|
||
|
||
|
||
def test_uncited_match_counts_in_both_percentages():
|
||
text = "Скопированный без указания источника текст фрагмента дальше."
|
||
out = aggregate_results([_mp("A", 0, 20)], [], [], total_fragments=4, full_text=text)
|
||
assert out["matches"][0]["cited"] is False
|
||
assert out["overall_similarity"] == out["uncited_similarity"] == 25.0
|
||
assert out["uncited_fragments"] == 1
|
||
|
||
|
||
def test_no_full_text_means_nothing_marked_cited():
|
||
out = aggregate_results([_m("A", 0)], [], [], total_fragments=2) # full_text не передан
|
||
assert out["matches"][0]["cited"] is False
|
||
assert out["overall_similarity"] == out["uncited_similarity"]
|
||
|
||
|
||
# ─── aggregate_results: recommendations ──────────────────────────────────────
|
||
|
||
def _cand(title: str, similarity: float, url: str | None = None) -> dict:
|
||
return {
|
||
"fragment": "фрагмент",
|
||
"position_start": 0,
|
||
"position_end": 8,
|
||
"similarity": similarity,
|
||
"source_title": title,
|
||
"source_url": url,
|
||
"source_db": "openalex",
|
||
}
|
||
|
||
|
||
def test_no_related_candidates_means_empty_recommendations():
|
||
out = aggregate_results([], [], [], total_fragments=10)
|
||
assert out["recommendations"] == []
|
||
|
||
|
||
def test_related_candidates_become_recommendations_sorted_by_similarity():
|
||
out = aggregate_results(
|
||
[], [], [], total_fragments=10,
|
||
related_candidates=[_cand("Low", 60.0), _cand("High", 90.0)],
|
||
)
|
||
titles = [r["source_title"] for r in out["recommendations"]]
|
||
assert titles == ["High", "Low"]
|
||
|
||
|
||
def test_recommendations_dedup_keeps_best_per_source():
|
||
out = aggregate_results(
|
||
[], [], [], total_fragments=10,
|
||
related_candidates=[_cand("A", 60.0, "url-a"), _cand("A", 85.0, "url-a")],
|
||
)
|
||
assert len(out["recommendations"]) == 1
|
||
assert out["recommendations"][0]["similarity"] == 85.0
|
||
|
||
|
||
def test_recommendations_capped_at_ten():
|
||
candidates = [_cand(f"S{i}", float(i), f"url-{i}") for i in range(15)]
|
||
out = aggregate_results([], [], [], total_fragments=10, related_candidates=candidates)
|
||
assert len(out["recommendations"]) == 10
|
||
|
||
|
||
def test_source_already_flagged_as_match_excluded_from_recommendations():
|
||
semantic = [{**_cand("A", 90.0, "url-a"), "method": "semantic+llm"}]
|
||
out = aggregate_results(
|
||
[], [], semantic, total_fragments=10,
|
||
related_candidates=[_cand("A", 60.0, "url-a"), _cand("B", 70.0, "url-b")],
|
||
)
|
||
titles = [r["source_title"] for r in out["recommendations"]]
|
||
assert titles == ["B"]
|