Первичная версия: стилометрия реплик, LSA и протокол экспериментов

lab1/stylometry/metadata.py

@@ -0,0 +1,99 @@
+"""Анализ ошибок по метаданным персонажей."""
+
+from __future__ import annotations
+
+from pathlib import Path
+from typing import Sequence
+
+import pandas as pd
+
+from .evaluation import per_class_report
+
+DEFAULT_METADATA_PATH = Path(__file__).resolve().parents[1] / "data" / "metadata" / "characters.csv"
+
+METADATA_COLUMNS = ("gender", "age_group", "species", "role")
+
+
+def load_character_metadata(path: str | Path | None = None) -> pd.DataFrame:
+    """Читает таблицу пола, возраста, вида и роли персонажей."""
+    path = Path(path) if path is not None else DEFAULT_METADATA_PATH
+    if not path.exists():
+        raise FileNotFoundError(f"Нет файла метаданных: {path}")
+    frame = pd.read_csv(path)
+    required = {"character", *METADATA_COLUMNS}
+    missing = required - set(frame.columns)
+    if missing:
+        raise ValueError(f"В метаданных нет колонок: {sorted(missing)}")
+    frame = frame.copy()
+    frame["character"] = frame["character"].astype(str).str.strip()
+    return frame
+
+
+def subgroup_report(
+    y_true: Sequence,
+    y_pred: Sequence,
+    labels: Sequence,
+    metadata: pd.DataFrame,
+    group_column: str = "gender",
+) -> pd.DataFrame:
+    """Сводка F-меры по подгруппам персонажей."""
+    if group_column not in metadata.columns:
+        raise ValueError(f"Нет колонки {group_column!r} в метаданных.")
+    per_class = per_class_report(y_true, y_pred, labels)
+    per_class = per_class.merge(
+        metadata[["character", group_column]],
+        left_on="class",
+        right_on="character",
+        how="left",
+    )
+    rows = []
+    for value, group in per_class.groupby(group_column, dropna=False):
+        support = int(group["support"].sum())
+        weighted = float((group["f1"] * group["support"]).sum() / max(support, 1))
+        rows.append(
+            {
+                "group_column": group_column,
+                "group": str(value),
+                "n_characters": int(len(group)),
+                "n_samples": support,
+                "weighted_f1": weighted,
+                "mean_class_f1": float(group["f1"].mean()),
+                "min_class_f1": float(group["f1"].min()),
+                "max_class_f1": float(group["f1"].max()),
+            }
+        )
+    return pd.DataFrame(rows).sort_values(["weighted_f1"], ascending=False)
+
+
+def group_confusions(
+    y_true: Sequence,
+    y_pred: Sequence,
+    metadata: pd.DataFrame,
+    group_column: str = "gender",
+) -> pd.DataFrame:
+    """Считает, как часто ошибки происходят внутри и между группами."""
+    frame = pd.DataFrame({"true_character": list(y_true), "pred_character": list(y_pred)})
+    frame = frame.merge(
+        metadata[["character", group_column]].rename(
+            columns={"character": "true_character", group_column: "true_group"}
+        ),
+        on="true_character",
+        how="left",
+    ).merge(
+        metadata[["character", group_column]].rename(
+            columns={"character": "pred_character", group_column: "pred_group"}
+        ),
+        on="pred_character",
+        how="left",
+    )
+    errors = frame[frame["true_character"] != frame["pred_character"]]
+    table = (
+        errors.groupby(["true_group", "pred_group"], dropna=False)
+        .size()
+        .reset_index(name="count")
+        .sort_values("count", ascending=False)
+    )
+    total = max(int(table["count"].sum()), 1)
+    table["share_of_errors"] = table["count"] / total
+    table["same_group"] = table["true_group"] == table["pred_group"]
+    return table