ModernBERT_JP_CAUSE-EFFECT.ipynb

cause-effect.ipynb

{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "provenance": [],
      "gpuType": "T4",
      "name": "ModernBERT_JP_CAUSE-EFFECT.ipynb",
      "authorship_tag": "ABX9TyOwl6lTU9stRHzbPBEXnsF4",
      "include_colab_link": true
    },
    "kernelspec": {
      "name": "python3",
      "display_name": "Python 3"
    },
    "language_info": {
      "name": "python"
    },
    "accelerator": "GPU"
  },
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "view-in-github",
        "colab_type": "text"
      },
      "source": [
        "<a href=\"https://colab.research.google.com/gist/chottokun/af798278c56d46109c16d27aa64383ca/cause-effect.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "f1c542d9"
      },
      "source": [
        "#  コンテキスト指向 NER パイプライン\n",
        "\n",
        "このノートブックでは、`cl-nagoya/ruri-v3-pt-30m` モデルを使用して、文脈に基づいた固有表現抽出（Context NER）を行う一連の流れを実装しています。具体的には、以下の3つのステップで構成されています。\n",
        "\n",
        "1.  **データセット準備 (Dataset Preparation)**:\n",
        "    「原因 (CAUSE)」と「結果 (EFFECT)」の関係を含む日本語のトイデータを作成し、BIOタグ形式でアノテーションを行います。\n",
        "\n",
        "2.  **ファインチューニング (Fine-tuning)**:\n",
        "    作成したデータセットを用いて Ruri-v3 モデルを学習させます。データ量が少ないため、パターンを確実に学習（過学習）させるために20エポック回します。\n",
        "\n",
        "3.  **推論 (Inference)**:\n",
        "    学習済みモデルを使用して、新しい文章（例：「システム障害によりサービスが停止した」）から原因と結果を抽出できるか検証します。"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "123acae6"
      },
      "source": [
        "# --- 0. 環境構築 (Setup) ---\n",
        "# 必要なライブラリをインストール（既にインストール済みの場合はスキップされます）\n",
        "!pip install -U \"transformers>=4.48.0\" accelerate datasets seqeval\n",
        "\n",
        "import os\n",
        "# W&B (Weights & Biases) を無効化\n",
        "os.environ[\"WANDB_DISABLED\"] = \"true\"\n",
        "\n",
        "import logging\n",
        "import warnings\n",
        "import torch\n",
        "from datasets import Dataset\n",
        "from transformers import (\n",
        "    AutoTokenizer,\n",
        "    AutoModelForTokenClassification,\n",
        "    TrainingArguments,\n",
        "    Trainer,\n",
        "    DataCollatorForTokenClassification,\n",
        "    pipeline,\n",
        "    logging as transformers_logging\n",
        ")\n",
        "\n",
        "# ログ出力を整理して見やすく設定\n",
        "warnings.filterwarnings(\"ignore\")\n",
        "transformers_logging.set_verbosity_error()\n",
        "logging.getLogger(\"huggingface_hub\").setLevel(logging.ERROR)\n",
        "\n",
        "# --- 1. データセットの準備 (Data Preparation) ---\n",
        "print(\"1. データセットを準備しています...\")\n",
        "\n",
        "# ラベルの定義 (0:その他, 1:原因開始, 2:原因継続, 3:結果開始, 4:結果継続)\n",
        "id2label = {0: 'O', 1: 'B-CAUSE', 2: 'I-CAUSE', 3: 'B-EFFECT', 4: 'I-EFFECT'}\n",
        "label2id = {v: k for k, v in id2label.items()}\n",
        "\n",
        "# サンプルデータ（ここを独自のデータに差し替えることで再学習可能）\n",
        "japanese_toy_data = [\n",
        "    {\"tokens\": [\"大雨\", \"により\", \"川\", \"が\", \"氾濫\", \"した\", \"。\"], \"ner_tags\": [1, 0, 3, 4, 4, 4, 0]},\n",
        "    {\"tokens\": [\"システム\", \"障害\", \"の\", \"ため\", \"、\", \"サービス\", \"が\", \"停止\", \"している\", \"。\"], \"ner_tags\": [1, 2, 0, 0, 0, 3, 4, 4, 4, 0]},\n",
        "    {\"tokens\": [\"寝不足\", \"は\", \"集中力\", \"低下\", \"を\", \"招く\", \"。\"], \"ner_tags\": [1, 0, 3, 4, 0, 0, 0]},\n",
        "    {\"tokens\": [\"インフレ\", \"によって\", \"物価\", \"が\", \"上昇\", \"した\", \"。\"], \"ner_tags\": [1, 0, 3, 4, 4, 4, 0]},\n",
        "    {\"tokens\": [\"渋滞\", \"の\", \"せいで\", \"到着\", \"が\", \"遅れた\", \"。\"], \"ner_tags\": [1, 0, 0, 3, 4, 4, 0]},\n",
        "    {\"tokens\": [\"火花\", \"が\", \"火災\", \"の\", \"原因\", \"と\", \"なった\", \"。\"], \"ner_tags\": [1, 0, 3, 4, 4, 0, 0, 0]},\n",
        "    {\"tokens\": [\"過労\", \"が\", \"健康\", \"悪化\", \"に\", \"つながる\", \"。\"], \"ner_tags\": [1, 0, 3, 4, 0, 0, 0]},\n",
        "    {\"tokens\": [\"メモリ\", \"リーク\", \"により\", \"アプリ\", \"が\", \"クラッシュ\", \"した\", \"。\"], \"ner_tags\": [1, 2, 0, 3, 4, 4, 4, 0]},\n",
        "    {\"tokens\": [\"地球\", \"温暖化\", \"は\", \"海面\", \"上昇\", \"を\", \"引き起こす\", \"。\"], \"ner_tags\": [1, 2, 0, 3, 4, 0, 0, 0]},\n",
        "    {\"tokens\": [\"ストレス\", \"で\", \"胃\", \"が\", \"痛く\", \"なる\", \"。\"], \"ner_tags\": [1, 0, 3, 4, 4, 4, 0]},\n",
        "    {\"tokens\": [\"構造\", \"的\", \"欠陥\", \"により\", \"橋\", \"が\", \"崩落\", \"した\", \"。\"], \"ner_tags\": [1, 2, 2, 0, 3, 4, 4, 4, 0]},\n",
        "    {\"tokens\": [\"猛暑\", \"のため\", \"作物\", \"が\", \"枯れた\", \"。\"], \"ner_tags\": [1, 0, 3, 4, 4, 0]},\n",
        "    {\"tokens\": [\"停電\", \"は\", \"落雷\", \"が\", \"原因\", \"だった\", \"。\"], \"ner_tags\": [3, 0, 1, 0, 0, 0, 0]},\n",
        "    {\"tokens\": [\"栄養\", \"不足\", \"は\", \"免疫力\", \"低下\", \"を\", \"招く\", \"。\"], \"ner_tags\": [1, 2, 0, 3, 4, 0, 0, 0]},\n",
        "    {\"tokens\": [\"古い\", \"ドライバ\", \"を\", \"使う\", \"と\", \"動作\", \"が\", \"不安定\", \"に\", \"なる\", \"。\"], \"ner_tags\": [1, 2, 0, 0, 0, 3, 4, 4, 0, 0, 0]},\n",
        "    {\"tokens\": [\"政治\", \"的\", \"不安\", \"が\", \"株価\", \"下落\", \"を\", \"招いた\", \"。\"], \"ner_tags\": [1, 2, 2, 0, 3, 4, 0, 0, 0]},\n",
        "    {\"tokens\": [\"ウイルス\", \"感染\", \"により\", \"データ\", \"が\", \"消失\", \"した\", \"。\"], \"ner_tags\": [1, 2, 0, 3, 4, 4, 4, 0]},\n",
        "    {\"tokens\": [\"需要\", \"急増\", \"が\", \"供給\", \"不足\", \"に\", \"つながった\", \"。\"], \"ner_tags\": [1, 2, 0, 3, 4, 0, 0, 0]},\n",
        "    {\"tokens\": [\"政策\", \"変更\", \"は\", \"抗議\", \"デモ\", \"を\", \"引き起こした\", \"。\"], \"ner_tags\": [1, 2, 0, 3, 4, 0, 0, 0]},\n",
        "    {\"tokens\": [\"大きな\", \"音\", \"に\", \"猫\", \"が\", \"驚いた\", \"。\"], \"ner_tags\": [1, 2, 0, 3, 4, 4, 0]}\n",
        "]\n",
        "\n",
        "dataset = Dataset.from_list(japanese_toy_data)\n",
        "\n",
        "# --- 2. モデルとトークナイザーの準備 (Model Loading) ---\n",
        "model_id = \"cl-nagoya/ruri-v3-pt-30m\"\n",
        "print(f\"{model_id} をロード中...\")\n",
        "\n",
        "tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)\n",
        "model = AutoModelForTokenClassification.from_pretrained(\n",
        "    model_id,\n",
        "    num_labels=len(id2label),\n",
        "    id2label=id2label,\n",
        "    label2id=label2id,\n",
        "    trust_remote_code=True\n",
        ")\n",
        "\n",
        "# データの前処理（トークン化とラベル位置合わせ）\n",
        "def tokenize_and_align_labels(examples):\n",
        "    tokenized_inputs = tokenizer(\n",
        "        examples[\"tokens\"],\n",
        "        truncation=True,\n",
        "        is_split_into_words=True,\n",
        "        max_length=8192 # Ruri-v3の長文脈対応能力\n",
        "    )\n",
        "    labels = []\n",
        "    for i, label in enumerate(examples[\"ner_tags\"]):\n",
        "        word_ids = tokenized_inputs.word_ids(batch_index=i)\n",
        "        previous_word_idx = None\n",
        "        label_ids = []\n",
        "        for word_idx in word_ids:\n",
        "            if word_idx is None:\n",
        "                label_ids.append(-100) # 特殊トークンは無視\n",
        "            elif word_idx != previous_word_idx:\n",
        "                label_ids.append(label[word_idx]) # 単語の先頭\n",
        "            else:\n",
        "                label_ids.append(-100) # 単語の途中\n",
        "            previous_word_idx = word_idx\n",
        "        labels.append(label_ids)\n",
        "    tokenized_inputs[\"labels\"] = labels\n",
        "    return tokenized_inputs\n",
        "\n",
        "tokenized_dataset = dataset.map(tokenize_and_align_labels, batched=True)\n",
        "\n",
        "# --- 3. ファインチューニング (Training) ---\n",
        "print(\"\\n2. モデルの学習を開始します...\")\n",
        "\n",
        "training_args = TrainingArguments(\n",
        "    output_dir=\"ruri_final_demo\",\n",
        "    learning_rate=5e-5,\n",
        "    per_device_train_batch_size=4,\n",
        "    num_train_epochs=20, # データが少ないため多めに設定（本番データでは3-5程度でOK）\n",
        "    weight_decay=0.01,\n",
        "    eval_strategy=\"epoch\",\n",
        "    save_strategy=\"epoch\",\n",
        "    logging_steps=10,\n",
        "    report_to=\"none\" # WandBを無効化\n",
        ")\n",
        "\n",
        "trainer = Trainer(\n",
        "    model=model,\n",
        "    args=training_args,\n",
        "    train_dataset=tokenized_dataset,\n",
        "    eval_dataset=tokenized_dataset,\n",
        "    processing_class=tokenizer,\n",
        "    data_collator=DataCollatorForTokenClassification(tokenizer)\n",
        ")\n",
        "\n",
        "trainer.train()\n",
        "trainer.save_model(\"./ruri_final_demo\")\n",
        "print(\"学習完了。モデルを保存しました。\")\n",
        "\n",
        "# --- 4. 推論の検証 (Inference) ---\n",
        "print(\"\\n3. 推論テストを実行します...\")\n",
        "\n",
        "device = 0 if torch.cuda.is_available() else -1\n",
        "ner_pipeline = pipeline(\n",
        "    \"ner\",\n",
        "    model=\"./ruri_final_demo\",\n",
        "    tokenizer=\"./ruri_final_demo\",\n",
        "    aggregation_strategy=\"simple\",\n",
        "    trust_remote_code=True,\n",
        "    device=device\n",
        ")\n",
        "\n",
        "# テスト用入力文\n",
        "text = \"システム障害によりサービスが停止した。\"\n",
        "results = ner_pipeline(text)\n",
        "\n",
        "print(f\"\\n入力文: {text}\")\n",
        "print(\"抽出結果:\")\n",
        "if not results:\n",
        "    print(\"エンティティが検出されませんでした。\")\n",
        "else:\n",
        "    for entity in results:\n",
        "        print(f\"  - [{entity['entity_group']}] {entity['word']} (信頼度: {entity['score']:.4f})\")"
      ],
      "execution_count": null,
      "outputs": []
    }
  ]
}