notebooks/03_NER_Analysis.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "2dae8368-f5e8-4ea5-8836-3579b97302b0",
   "metadata": {},
   "outputs": [],
   "source": [
    "import sys\n",
    "sys.path.append('../util')\n",
    "sys.path.append('../experiments')\n",
    "\n",
    "import os\n",
    "# Disable weights and biases (if installed)\n",
    "os.environ[\"WANDB_DISABLED\"] = \"true\"\n",
    "os.environ[\"TOKENIZERS_PARALLELISM\"] = \"false\"\n",
    "os.environ[\"CUDA_DEVICE_ORDER\"] = \"PCI_BUS_ID\"\n",
    "os.environ[\"CUDA_VISIBLE_DEVICES\"] = \"0\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "f671255e-bc4d-4e49-888d-1063100cd603",
   "metadata": {},
   "outputs": [],
   "source": [
    "from pathlib import Path\n",
    "import re\n",
    "import logging\n",
    "from convert_annotations import read_webanno, webanno_to_iob_df, join_and_resolve, EMPTY_REGEX, VALUE_REGEX, log\n",
    "import pandas as pd\n",
    "import transformers, datasets\n",
    "from transformers import AutoModelForTokenClassification, AutoTokenizer, Trainer, TrainingArguments, pipeline, DataCollatorForTokenClassification, EarlyStoppingCallback, trainer_utils\n",
    "from huggingface_utils import load_custom_dataset, LabelAligner, compute_metrics, eval_on_test_set\n",
    "from run_experiment import get_train_args\n",
    "import hydra\n",
    "from hydra import compose, initialize\n",
    "\n",
    "datasets.logging.set_verbosity_error()\n",
    "transformers.logging.disable_default_handler()\n",
    "log.setLevel(logging.ERROR)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "509dddc7-1afb-4595-b066-b2af076b17a9",
   "metadata": {},
   "source": [
    "## Error Analysis"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "7a9b2675-bc7d-4a70-8ecc-3fd8311a5bd4",
   "metadata": {},
   "source": [
    "### Load the best model for the span length / granularity settings and the corresponding dataset"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "id": "8d194511-7042-4183-ba3a-4b8853bb888f",
   "metadata": {},
   "outputs": [],
   "source": [
    "def load_pipeline_and_data(setting):\n",
    "    cfg_file = f'{setting}.yaml'\n",
    "    model_checkpoint = f'../data/models/huggingface/{setting}/'\n",
    "\n",
    "    hydra.core.global_hydra.GlobalHydra.instance().clear()\n",
    "    initialize(config_path=Path('..') / 'experiments')\n",
    "    config = compose(config_name=cfg_file, overrides=['cuda=0', 'link=false'])\n",
    "    \n",
    "    tokenizer = AutoTokenizer.from_pretrained(config['base_model_checkpoint'])\n",
    "    \n",
    "    dataset, tags = load_custom_dataset(train=config['train_dataset'], dev=config['dev_dataset'], test=config['test_dataset'], tag_strings=config['task'])\n",
    "    \n",
    "    assert isinstance(tokenizer, transformers.PreTrainedTokenizerFast)\n",
    "    \n",
    "    label_aligner = LabelAligner(tokenizer)\n",
    "    dataset = dataset.map(lambda e: label_aligner.tokenize_and_align_labels(e, config['label_all_tokens']), batched=True)\n",
    "    \n",
    "    training_args = get_train_args(cp_path='../ner_results', run_name='error_analysis', report_to=[], **config, resume_from_checkpoint=model_checkpoint)\n",
    "    \n",
    "    model = AutoModelForTokenClassification.from_pretrained(\n",
    "        model_checkpoint\n",
    "    )\n",
    "\n",
    "    data_collator = DataCollatorForTokenClassification(tokenizer)\n",
    "    tr = Trainer(\n",
    "        args=training_args,\n",
    "        model=model,\n",
    "        train_dataset=dataset[\"train\"],\n",
    "        eval_dataset=dataset[\"dev\"],\n",
    "        data_collator=data_collator,\n",
    "        tokenizer=tokenizer,\n",
    "        compute_metrics=compute_metrics(tags, True),\n",
    "    )\n",
    "    \n",
    "    return tr, model, dataset[\"test\"], tokenizer"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "id": "cc6aa250-a2fb-4417-af3b-a8c93def2388",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "6c3ab631b4de40b1a17c612978a2aa11",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "  0%|          | 0/3 [00:00<?, ?it/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "e3b0e3f8066b48ae95b4f4100244f2c1",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "  0%|          | 0/59515 [00:00<?, ?ex/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "f042f6cec54f4834adc4840bd5c0f6b7",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "  0%|          | 0/12770 [00:00<?, ?ex/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "155212b364d54bbab185709ad22cee35",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "  0%|          | 0/13714 [00:00<?, ?ex/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "cbfb66fac78144dfaf4c43d13c5a266a",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "  0%|          | 0/6 [00:00<?, ?ba/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "d95d5ce8b4a64efdafabf5c4bc5a0f2e",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "  0%|          | 0/2 [00:00<?, ?ba/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "63cedd1aad9c46a2aac985f17a311ff0",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "  0%|          | 0/2 [00:00<?, ?ba/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "271cd8ff57374b18a02a2886b8b15fdc",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "  0%|          | 0/60 [00:00<?, ?ba/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "398c2325a93e4630b85ea35275fa9ee6",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "  0%|          | 0/13 [00:00<?, ?ba/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "73cc478683b94c7e8dca1e7dc259c251",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "  0%|          | 0/14 [00:00<?, ?ba/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "INFO:run_experiment:error_analysis\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "9887cf1272c44d0189836fc1278bfe3e",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "  0%|          | 0/3 [00:00<?, ?it/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "fdcf9ba03bda4747bc7ec441c705e796",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "  0%|          | 0/60 [00:00<?, ?ba/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "76156eab9ce247e4b493c02ee30a20ba",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "  0%|          | 0/13 [00:00<?, ?ba/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "d09ee54a313b4b399c96dbda00bbc5d5",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "  0%|          | 0/14 [00:00<?, ?ba/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "INFO:run_experiment:error_analysis\n"
     ]
    }
   ],
   "source": [
    "tr_short, model_short, test_ds_short, tokenizer = load_pipeline_and_data('02_ggponc_fine_short')\n",
    "tr_long, model_long, test_ds_long, _ = load_pipeline_and_data('04_ggponc_fine_long')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "id": "6d4d3b71-313b-4409-a6f4-8607fc25b121",
   "metadata": {},
   "outputs": [],
   "source": [
    "from transformers.pipelines.token_classification import AggregationStrategy\n",
    "pipe_short = pipeline(\"ner\", model_short, tokenizer=tokenizer, device=0, aggregation_strategy=AggregationStrategy.FIRST)\n",
    "pipe_long = pipeline(\"ner\", model_long, tokenizer=tokenizer, device=0, aggregation_strategy=AggregationStrategy.FIRST)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "c2daaee6-b02c-4751-a279-9b49bf104567",
   "metadata": {},
   "source": [
    "### Look at predictions on a few example sentences"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "id": "4793d55e-bd7a-4f1c-922a-8a4bb8a76678",
   "metadata": {},
   "outputs": [],
   "source": [
    "wiki_sentence = \"\"\"\"Cetuximab ist ein monoklonaler Antikörper, der gegen den epidermalen Wachstumsfaktorrezeptor (EGFR) gerichtet ist und \n",
    "dient zur Therapie des fortgeschrittenen kolorektalen Karzinoms zusammen mit Irinotecan oder in Kombination mit FOLFOX bzw. \n",
    "allein nach Versagen einer Behandlung mit Oxaliplatin und Irinotecan.\"\"\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "id": "a69c68fa-79da-47c9-87fd-7fe32de706cc",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[{'entity_group': 'Clinical_Drug',\n",
       "  'score': 0.962233,\n",
       "  'word': 'Cetuximab',\n",
       "  'start': 1,\n",
       "  'end': 10},\n",
       " {'entity_group': 'Clinical_Drug',\n",
       "  'score': 0.9626528,\n",
       "  'word': 'Antikörper',\n",
       "  'start': 32,\n",
       "  'end': 42},\n",
       " {'entity_group': 'Nutrient_or_Body_Substance',\n",
       "  'score': 0.9637979,\n",
       "  'word': 'Wachstumsfaktorrezeptor',\n",
       "  'start': 70,\n",
       "  'end': 93},\n",
       " {'entity_group': 'Nutrient_or_Body_Substance',\n",
       "  'score': 0.9641919,\n",
       "  'word': 'EGFR',\n",
       "  'start': 95,\n",
       "  'end': 99},\n",
       " {'entity_group': 'Therapeutic',\n",
       "  'score': 0.9628338,\n",
       "  'word': 'Therapie',\n",
       "  'start': 130,\n",
       "  'end': 138},\n",
       " {'entity_group': 'Diagnosis_or_Pathology',\n",
       "  'score': 0.9629651,\n",
       "  'word': 'Karzinoms',\n",
       "  'start': 174,\n",
       "  'end': 183},\n",
       " {'entity_group': 'Clinical_Drug',\n",
       "  'score': 0.9623752,\n",
       "  'word': 'Irinotecan',\n",
       "  'start': 197,\n",
       "  'end': 207},\n",
       " {'entity_group': 'Therapeutic',\n",
       "  'score': 0.9629323,\n",
       "  'word': 'FOLFOX',\n",
       "  'start': 232,\n",
       "  'end': 238},\n",
       " {'entity_group': 'Diagnosis_or_Pathology',\n",
       "  'score': 0.96274424,\n",
       "  'word': 'Versagen',\n",
       "  'start': 257,\n",
       "  'end': 265},\n",
       " {'entity_group': 'Therapeutic',\n",
       "  'score': 0.9627209,\n",
       "  'word': 'Behandlung',\n",
       "  'start': 272,\n",
       "  'end': 282},\n",
       " {'entity_group': 'Clinical_Drug',\n",
       "  'score': 0.96242136,\n",
       "  'word': 'Oxaliplatin',\n",
       "  'start': 287,\n",
       "  'end': 298},\n",
       " {'entity_group': 'Clinical_Drug',\n",
       "  'score': 0.96254873,\n",
       "  'word': 'Irinotecan',\n",
       "  'start': 303,\n",
       "  'end': 313}]"
      ]
     },
     "execution_count": 7,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "pipe_short(wiki_sentence)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "id": "4b63d634-fa8c-4b50-b6fd-6456e48f4cc1",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[{'entity_group': 'Clinical_Drug',\n",
       "  'score': 0.92421424,\n",
       "  'word': 'Cetuximab',\n",
       "  'start': 1,\n",
       "  'end': 10},\n",
       " {'entity_group': 'Clinical_Drug',\n",
       "  'score': 0.9235098,\n",
       "  'word': 'monoklonaler Antikörper',\n",
       "  'start': 19,\n",
       "  'end': 42},\n",
       " {'entity_group': 'Nutrient_or_Body_Substance',\n",
       "  'score': 0.92463005,\n",
       "  'word': 'epidermalen Wachstumsfaktorrezeptor',\n",
       "  'start': 58,\n",
       "  'end': 93},\n",
       " {'entity_group': 'Nutrient_or_Body_Substance',\n",
       "  'score': 0.9239702,\n",
       "  'word': 'EGFR',\n",
       "  'start': 95,\n",
       "  'end': 99},\n",
       " {'entity_group': 'Therapeutic',\n",
       "  'score': 0.92163867,\n",
       "  'word': 'Therapie des fortgeschrittenen kolorektalen Karzinoms',\n",
       "  'start': 130,\n",
       "  'end': 183},\n",
       " {'entity_group': 'Clinical_Drug',\n",
       "  'score': 0.9246952,\n",
       "  'word': 'Irinotecan',\n",
       "  'start': 197,\n",
       "  'end': 207},\n",
       " {'entity_group': 'Therapeutic',\n",
       "  'score': 0.9218072,\n",
       "  'word': 'FOLFOX',\n",
       "  'start': 232,\n",
       "  'end': 238},\n",
       " {'entity_group': 'Diagnosis_or_Pathology',\n",
       "  'score': 0.9211978,\n",
       "  'word': 'Versagen einer Behandlung mit Oxaliplatin und Irinotecan',\n",
       "  'start': 257,\n",
       "  'end': 313}]"
      ]
     },
     "execution_count": 8,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "pipe_long(wiki_sentence)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "id": "1899a7b2-27e8-4762-ba6c-bef50adc5822",
   "metadata": {},
   "outputs": [],
   "source": [
    "guideline_sentence = \"\"\"Als Alternative empfiehlt die ASCCP bei zytologischem Verdacht auf CIN 1/2 die sofortige Kolposkopie.\"\"\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "id": "37dc1a0c-5c61-49a2-93d8-caf59e88b4a0",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[{'entity_group': 'Other_Finding',\n",
       "  'score': 0.96294,\n",
       "  'word': 'Verdacht',\n",
       "  'start': 54,\n",
       "  'end': 62},\n",
       " {'entity_group': 'Diagnosis_or_Pathology',\n",
       "  'score': 0.96257114,\n",
       "  'word': 'CIN',\n",
       "  'start': 67,\n",
       "  'end': 70},\n",
       " {'entity_group': 'Diagnostic',\n",
       "  'score': 0.9636781,\n",
       "  'word': 'Kolposkopie',\n",
       "  'start': 89,\n",
       "  'end': 100}]"
      ]
     },
     "execution_count": 10,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "pipe_short(guideline_sentence)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "id": "3551790f-039f-488b-bea3-b519af737f2c",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[{'entity_group': 'Other_Finding',\n",
       "  'score': 0.9217645,\n",
       "  'word': 'zytologischem Verdacht auf CIN 1 / 2',\n",
       "  'start': 40,\n",
       "  'end': 74},\n",
       " {'entity_group': 'Diagnostic',\n",
       "  'score': 0.9219378,\n",
       "  'word': 'sofortige Kolposkopie',\n",
       "  'start': 79,\n",
       "  'end': 100}]"
      ]
     },
     "execution_count": 11,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "pipe_long(guideline_sentence)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "id": "e9fd6f45-5ce2-4831-a65c-2dd7f565a12c",
   "metadata": {},
   "outputs": [],
   "source": [
    "clinical_note_sentence  = \"\"\"Antibiose fortsetzen (s. o.), Abstrich erfragen, ggf. Umstellung der Antibiose. Thromboseprophylaxe bis zur sicheren Mobilität.\"\"\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "id": "c99df041-2840-45dc-a4b9-9bbafa68a3e3",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[{'entity_group': 'Therapeutic',\n",
       "  'score': 0.96311814,\n",
       "  'word': 'Antibiose',\n",
       "  'start': 0,\n",
       "  'end': 9},\n",
       " {'entity_group': 'Therapeutic',\n",
       "  'score': 0.96376956,\n",
       "  'word': 'Umstellung',\n",
       "  'start': 54,\n",
       "  'end': 64},\n",
       " {'entity_group': 'Therapeutic',\n",
       "  'score': 0.96073705,\n",
       "  'word': 'Antibiose',\n",
       "  'start': 69,\n",
       "  'end': 78},\n",
       " {'entity_group': 'Therapeutic',\n",
       "  'score': 0.9622617,\n",
       "  'word': 'Thromboseprophylaxe',\n",
       "  'start': 80,\n",
       "  'end': 99},\n",
       " {'entity_group': 'Other_Finding',\n",
       "  'score': 0.96310234,\n",
       "  'word': 'Mobilität',\n",
       "  'start': 117,\n",
       "  'end': 126}]"
      ]
     },
     "execution_count": 13,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "pipe_short(clinical_note_sentence)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "id": "8b7ded95-551d-4e62-949d-70328291d475",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[{'entity_group': 'Therapeutic',\n",
       "  'score': 0.92327774,\n",
       "  'word': 'Antibiose',\n",
       "  'start': 0,\n",
       "  'end': 9},\n",
       " {'entity_group': 'Diagnostic',\n",
       "  'score': 0.9229312,\n",
       "  'word': 'Abstrich',\n",
       "  'start': 30,\n",
       "  'end': 38},\n",
       " {'entity_group': 'Therapeutic',\n",
       "  'score': 0.9218672,\n",
       "  'word': 'Umstellung der Antibiose',\n",
       "  'start': 54,\n",
       "  'end': 78},\n",
       " {'entity_group': 'Therapeutic',\n",
       "  'score': 0.9218916,\n",
       "  'word': 'Thromboseprophylaxe',\n",
       "  'start': 80,\n",
       "  'end': 99},\n",
       " {'entity_group': 'Other_Finding',\n",
       "  'score': 0.9219573,\n",
       "  'word': 'sicheren Mobilität',\n",
       "  'start': 108,\n",
       "  'end': 126}]"
      ]
     },
     "execution_count": 14,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "pipe_long(clinical_note_sentence)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "756d4da8-8d90-4a00-83e5-055bf343befb",
   "metadata": {},
   "source": [
    "#### Get metrics and error analysis on test set"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "id": "0702ad5e-b156-41b4-9942-3aa92af15fae",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "\n",
       "    <div>\n",
       "      \n",
       "      <progress value='470' max='429' style='width:300px; height:20px; vertical-align: middle;'></progress>\n",
       "      [429/429 1:00:43]\n",
       "    </div>\n",
       "    "
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "13714it [00:05, 2575.51it/s]\n"
     ]
    },
    {
     "data": {
      "text/html": [
       "\n",
       "    <div>\n",
       "      \n",
       "      <progress value='470' max='429' style='width:300px; height:20px; vertical-align: middle;'></progress>\n",
       "      [429/429 1:00:22]\n",
       "    </div>\n",
       "    "
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "13714it [00:05, 2585.25it/s]\n"
     ]
    }
   ],
   "source": [
    "test_set_metrics_short = eval_on_test_set(test_ds_short, tr_short, tokenizer, \"test_short_full\")\n",
    "test_set_metrics_long = eval_on_test_set(test_ds_long, tr_long, tokenizer, \"test_long_full\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "id": "56f6c53e-2686-4f74-a5ce-ee2b2c77764a",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "{'test_long_full/boundary_error': 2684,\n",
       " 'test_long_full/false_negative': 1695,\n",
       " 'test_long_full/false_positive': 1849,\n",
       " 'test_long_full/label_boundary_error': 1830,\n",
       " 'test_long_full/labeling_error': 1521,\n",
       " 'test_long_full/true_positive': 22894,\n",
       " 'test_long_full/loss': 0.9115415811538696,\n",
       " 'test_long_full/Clinical_Drug_precision': 0.707409155261916,\n",
       " 'test_long_full/Clinical_Drug_recall': 0.7555443548387096,\n",
       " 'test_long_full/Clinical_Drug_f1': 0.7306848647331221,\n",
       " 'test_long_full/Clinical_Drug_number': 1984,\n",
       " 'test_long_full/Diagnosis_or_Pathology_precision': 0.7566739173422978,\n",
       " 'test_long_full/Diagnosis_or_Pathology_recall': 0.7842795654847304,\n",
       " 'test_long_full/Diagnosis_or_Pathology_f1': 0.7702294685990339,\n",
       " 'test_long_full/Diagnosis_or_Pathology_number': 9758,\n",
       " 'test_long_full/Diagnostic_precision': 0.6845344960246217,\n",
       " 'test_long_full/Diagnostic_recall': 0.7385168788046486,\n",
       " 'test_long_full/Diagnostic_f1': 0.7105017968853985,\n",
       " 'test_long_full/Diagnostic_number': 3614,\n",
       " 'test_long_full/External_Substance_precision': 0.45098039215686275,\n",
       " 'test_long_full/External_Substance_recall': 0.3357664233576642,\n",
       " 'test_long_full/External_Substance_f1': 0.38493723849372385,\n",
       " 'test_long_full/External_Substance_number': 137,\n",
       " 'test_long_full/Nutrient_or_Body_Substance_precision': 0.6070763500931099,\n",
       " 'test_long_full/Nutrient_or_Body_Substance_recall': 0.6379647749510763,\n",
       " 'test_long_full/Nutrient_or_Body_Substance_f1': 0.6221374045801527,\n",
       " 'test_long_full/Nutrient_or_Body_Substance_number': 511,\n",
       " 'test_long_full/Other_Finding_precision': 0.6374537379718727,\n",
       " 'test_long_full/Other_Finding_recall': 0.613740022805017,\n",
       " 'test_long_full/Other_Finding_f1': 0.6253721588846125,\n",
       " 'test_long_full/Other_Finding_number': 7016,\n",
       " 'test_long_full/Therapeutic_precision': 0.7562471078204535,\n",
       " 'test_long_full/Therapeutic_recall': 0.7631333177679197,\n",
       " 'test_long_full/Therapeutic_f1': 0.7596746077861708,\n",
       " 'test_long_full/Therapeutic_number': 8566,\n",
       " 'test_long_full/overall_precision': 0.716070873484613,\n",
       " 'test_long_full/overall_recall': 0.729310453998607,\n",
       " 'test_long_full/overall_f1': 0.7226300269778531,\n",
       " 'test_long_full/overall_accuracy': 0.9139037190323899,\n",
       " 'test_long_full/runtime': 16.0708,\n",
       " 'test_long_full/samples_per_second': 853.351,\n",
       " 'test_long_full/steps_per_second': 26.694}"
      ]
     },
     "execution_count": 16,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "test_set_metrics_long"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "d4f8c39e-1e18-4d59-9430-b58c75f414f2",
   "metadata": {},
   "source": [
    "# Coordination ellipses"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8747176f-2697-4f0b-b618-81d69556f739",
   "metadata": {},
   "source": [
    "Annotators were asked to annotate ellipses and their corresponding resolution using the `prefix` and `suffix` items as well as `fragment` relations for morpheme coordination (i.e., omission of subwords, in German commonly indicated by a Auslassungsstrich \"-\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "id": "2fe7b432-c145-4632-8512-62e30e77e4df",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "10193\n"
     ]
    }
   ],
   "source": [
    "# Unzip webanno raw annotations\n",
    "!unzip -q -n -d ../data/annotations/webannoo ../data/annotations/webanno.zip\n",
    "!ls ../data/annotations/webanno | wc -l"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "c42d1c97-12b2-4f50-8884-fa5a84a6c260",
   "metadata": {},
   "source": [
    "#### Read the raw data from WebAnno TSV files"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "id": "311a434c-0e9a-47b8-a56b-3341b01eda00",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 35 s, sys: 1.08 s, total: 36.1 s\n",
      "Wall time: 46.1 s\n"
     ]
    }
   ],
   "source": [
    "%%time\n",
    "tsv_folder = '../data/annotations/webanno'\n",
    "\n",
    "webanno_df, _ = read_webanno(Path(tsv_folder).glob('*.tsv'))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "id": "6c6f8441-cba4-4efb-80a6-67d812f1378e",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 85999/85999 [49:24<00:00, 29.01it/s]\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 49min 39s, sys: 22.7 s, total: 50min 1s\n",
      "Wall time: 49min 55s\n"
     ]
    }
   ],
   "source": [
    "%%time\n",
    "\n",
    "import warnings\n",
    "\n",
    "with warnings.catch_warnings():\n",
    "    warnings.simplefilter(\"ignore\")\n",
    "    iob_df = webanno_to_iob_df(webanno_df, 'detail', True, debug=False, collect_errors=False, skip_errors=True, all_columns=True)\n",
    "    iob_df['id'] = iob_df.entity_id.where(~iob_df.entity_id.isna(), iob_df.spec_id)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "id": "a25437f1-22d3-442b-8fc4-5746bbdf5722",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Anything containing a prefix, suffix or fragment is potentially an elliptical construct\n",
    "ellipses = iob_df[~iob_df.prefix.str.match(EMPTY_REGEX) | ~iob_df.suffix.str.match(EMPTY_REGEX) | ~iob_df.fragment.str.match(EMPTY_REGEX)].copy()"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "88331b1f-30ee-49e5-9c18-ab58f60fbf56",
   "metadata": {},
   "source": [
    "### Resolve ellipses and connect fragment parts"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "id": "5e9b3a28-3572-47e3-b35d-f31aecdda24d",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "<timed exec>:6: PerformanceWarning: indexing past lexsort depth may impact performance.\n",
      "<timed exec>:8: PerformanceWarning: indexing past lexsort depth may impact performance.\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 3min 43s, sys: 1.37 s, total: 3min 44s\n",
      "Wall time: 3min 45s\n"
     ]
    }
   ],
   "source": [
    "%%time\n",
    "results = []\n",
    "for _, r in ellipses.id.reset_index().drop_duplicates().iterrows():\n",
    "    has_fragment = False\n",
    "    idx = r.file, r.sentence_id\n",
    "    _id = r.id\n",
    "    sentence = ellipses.loc[idx]\n",
    "    span = sentence[sentence.id == _id]\n",
    "    full_sentence = iob_df.loc[idx]\n",
    "    full_span = full_sentence[full_sentence.id == _id]\n",
    "    for fragment in span.fragment: # Expand span to fragments\n",
    "        if fragment and not re.match(EMPTY_REGEX, fragment):\n",
    "            has_fragment = True\n",
    "            fragment = re.split('[|;]', fragment)[0]\n",
    "            m = re.match(VALUE_REGEX, fragment)\n",
    "            if not m:\n",
    "                print(span.fragment)\n",
    "            fragment_index = int(m.group(1).split('-')[1])\n",
    "            fragment_start = min(fragment_index, full_span.token_id.min())\n",
    "            fragment_end = max(fragment_index, full_span.token_id.max())\n",
    "            full_span = full_sentence[(full_sentence.token_id >= fragment_start) & (full_sentence.token_id <= fragment_end)]\n",
    "            expanded_context = full_sentence[full_sentence.id.isin(full_span.id.dropna().unique())].token_id\n",
    "            fragment_start = min(fragment_start, expanded_context.min())\n",
    "            fragment_end = max(fragment_end, expanded_context.max())\n",
    "            full_span = full_sentence[(full_sentence.token_id >= fragment_start) & (full_sentence.token_id <= fragment_end)]\n",
    "    fragment, resolution, missing_prefix, missing_suffix = join_and_resolve(full_span)\n",
    "    results.append({\n",
    "        'file' : idx[0],\n",
    "        'sentence_id' : idx[1],\n",
    "        'full_sentence' : list(full_sentence.token),\n",
    "        'span_index_start': full_span.token_id.min() - 1,\n",
    "        'span_index_end' : full_span.token_id.max() - 1,\n",
    "        'full_span' : fragment,\n",
    "        'resolution' : resolution,\n",
    "        'fragment' : has_fragment,\n",
    "        'missing_prefix' : missing_prefix,\n",
    "        'missing_suffix' : missing_suffix,\n",
    "    })\n",
    "results = pd.DataFrame(results)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "099021d3-dd91-4ba3-8e33-52f972ec0993",
   "metadata": {},
   "source": [
    "### Statistics"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "id": "1761c4c0-59f0-4730-b243-e658c0eb239f",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(4666, 78090, 0.05975156870277885)"
      ]
     },
     "execution_count": 22,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "ellipses_sentence_count = len(results[['file', 'sentence_id']].drop_duplicates())\n",
    "sentence_count = 78090\n",
    "ellipses_sentence_count, sentence_count, ellipses_sentence_count / sentence_count"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "id": "dd38aa1d-5bc0-4d04-8a64-736baf89c501",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Ellipses 5264\n",
      "Forward ellipses 1964 844\n",
      "Backward ellipses 2979 2739\n",
      "Complex ellipses 282 199\n"
     ]
    }
   ],
   "source": [
    "print('Ellipses', len(results))\n",
    "print('Forward ellipses', len(results[results.missing_prefix & ~results.missing_suffix]), len(results[results.fragment & results.missing_prefix & ~results.missing_suffix]))\n",
    "print('Backward ellipses', len(results[~results.missing_prefix & results.missing_suffix]), len(results[results.fragment & ~results.missing_prefix & results.missing_suffix]))\n",
    "print('Complex ellipses', len(results[results.missing_prefix & results.missing_suffix]), len(results[results.fragment & results.missing_prefix & results.missing_suffix]))"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "6b97695e-4371-4566-b344-241513a14c1a",
   "metadata": {},
   "source": [
    "### Browsing random examples and their resolution"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "id": "ecd2ea55-5250-4ede-aae1-fb1ce9af540a",
   "metadata": {},
   "outputs": [],
   "source": [
    "import random\n",
    "\n",
    "def show_random(df):\n",
    "    i = random.randint(0, len(df))\n",
    "    item = df.iloc[i]    \n",
    "    assert len(item.full_span) == len(item.resolution)\n",
    "    spans = []\n",
    "    solutions = []\n",
    "    marker = []\n",
    "    for span, res in zip(item.full_span, item.resolution):\n",
    "        if span == res:\n",
    "            marker.append(' ' * len(span))\n",
    "        else:\n",
    "            marker.append('^' * len(res))\n",
    "        start_index = res.find(span)\n",
    "        spans.append(span.rjust(start_index + len(span)).ljust(len(res) - start_index))\n",
    "        solutions.append(res)\n",
    "    print(' '.join(spans))\n",
    "    print(' '.join(solutions))\n",
    "    print(' '.join(marker))\n",
    "    print(item.file, item.sentence_id, 'fragment:', item.fragment)\n",
    "    print('...........')\n",
    "    print('')"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "317feabe-bce8-4ed5-8eec-4cc6b8e58048",
   "metadata": {},
   "source": [
    "Examples with fragment attribute (typically elliptical compounds)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "id": "731d926e-bfcb-4bf2-8bca-8cf4f078b252",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Lymph         - und Gefäßinvasion\n",
      "Lymphinvasion - und Gefäßinvasion\n",
      "^^^^^^^^^^^^^                    \n",
      "21_hodentumoren_0194.tsv 2 fragment: True\n",
      "...........\n",
      "\n",
      "signifikant besseres 1                 - und 2 - Jahresüberleben\n",
      "signifikant besseres 1-Jahresüberleben - und 2 - Jahresüberleben\n",
      "                     ^^^^^^^^^^^^^^^^^                          \n",
      "03_lungenkarzinom_0391.tsv 29 fragment: True\n",
      "...........\n",
      "\n",
      "en-bloc - und R0 - Resektionsraten\n",
      "en-bloc - und R0 - Resektionsraten\n",
      "                                  \n",
      "01_magenkarzinom_0073.tsv 27 fragment: True\n",
      "...........\n",
      "\n",
      "stationären und ambulanten Versorgungs          - und Therapieangeboten\n",
      "stationären und ambulanten Versorgungsangeboten - und Therapieangeboten\n",
      "                           ^^^^^^^^^^^^^^^^^^^^                        \n",
      "26_palliativmedizin_0089.tsv 5 fragment: True\n",
      "...........\n",
      "\n",
      "Gefühl der Selbstverantwortung und -           kontrolle\n",
      "Gefühl der Selbstverantwortung und - der Selbstkontrolle\n",
      "                                     ^^^^^^^^^^^^^^^^^^^\n",
      "07_malignes-melanom_0281.tsv 22 fragment: True\n",
      "...........\n",
      "\n",
      "besseren progressionsfreien           und Gesamtüberleben\n",
      "besseren progressionsfreien Überleben und Gesamtüberleben\n",
      "         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^                    \n",
      "07_malignes-melanom_0141.tsv 35 fragment: True\n",
      "...........\n",
      "\n",
      "malignem Melanom in den Stadien I ,         II ,         III und 7972 Patienten im Stadium IV\n",
      "malignem Melanom in den Stadien I , Stadien II , Stadium III und 7972 Patienten im Stadium IV\n",
      "                                    ^^^^^^^^^^   ^^^^^^^^^^^                                 \n",
      "07_malignes-melanom_0002.tsv 6 fragment: True\n",
      "...........\n",
      "\n",
      "HPV31 ,    33 ,    45 und    51\n",
      "HPV31 , HPV33 , HPV45 und HPV51\n",
      "        ^^^^^   ^^^^^     ^^^^^\n",
      "04_praevention-zervixkarzinom_0035.tsv 17 fragment: True\n",
      "...........\n",
      "\n",
      "Sterbe        - / Todeswünsche\n",
      "Sterbewünsche - / Todeswünsche\n",
      "^^^^^^^^^^^^^                 \n",
      "26_palliativmedizin_0025.tsv 5 fragment: True\n",
      "...........\n",
      "\n",
      "sequentiellen und simultanen Chemo         - / Radiotherapie\n",
      "sequentiellen und simultanen Chemotherapie - / Radiotherapie\n",
      "                             ^^^^^^^^^^^^^                  \n",
      "03_lungenkarzinom_0348.tsv 2 fragment: True\n",
      "...........\n",
      "\n"
     ]
    }
   ],
   "source": [
    "for _ in range(0, 10):\n",
    "    show_random(results[results.fragment])"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "6d7dade1-0d4c-4621-9bc8-9642ecc96cfd",
   "metadata": {},
   "source": [
    "Examples without fragment attribute (often omitted prepositions)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "id": "58ca6070-b6c7-4b43-8b63-980d7f9e382c",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Prognosefaktor für die lokale Tumorkontrolle ,     das rezidivfreie Überleben , und     das Gesamtüberleben\n",
      "Prognosefaktor für die lokale Tumorkontrolle , für das rezidivfreie Überleben , und für das Gesamtüberleben\n",
      "                                               ^^^^^^^                              ^^^^^^^                \n",
      "29_adulte-weichgewebesarkome_0124.tsv 2 fragment: False\n",
      "...........\n",
      "\n",
      "im Bindungsstil begründete Ängste vor Ablehnung und     Abhängigkeit\n",
      "im Bindungsstil begründete Ängste vor Ablehnung und vor Abhängigkeit\n",
      "                                                    ^^^^^^^^^^^^^^^^\n",
      "26_palliativmedizin_0606.tsv 14 fragment: False\n",
      "...........\n",
      "\n",
      "Verlängerung des progressionsfreien Überlebens und des Gesamtüberlebens bei geringer Toxizität und     guter Lebensqualität\n",
      "Verlängerung des progressionsfreien Überlebens und des Gesamtüberlebens bei geringer Toxizität und bei guter Lebensqualität\n",
      "                                                                                                   ^^^^^^^^^               \n",
      "02_kolorektales-karzinom_0494.tsv 2 fragment: False\n",
      "...........\n",
      "\n",
      "Resultat von erhöhter Anstrengung bzw .     physischer Erschöpfung\n",
      "Resultat von erhöhter Anstrengung bzw . von physischer Erschöpfung\n",
      "                                        ^^^^^^^^^^^^^^            \n",
      "28_komplementaermedizin_0048.tsv 33 fragment: False\n",
      "...........\n",
      "\n",
      "Rezidiv oder Progress nach Hochdosischemotherapie und      autologer Stammzelltransplantation\n",
      "Rezidiv oder Progress nach Hochdosischemotherapie und nach autologer Stammzelltransplantation\n",
      "                                                      ^^^^^^^^^^^^^^                         \n",
      "18_hodgkin-lymphom_0164.tsv 4 fragment: False\n",
      "...........\n",
      "\n",
      "Klinisch sichtbare Karzinose des Peritoneums oder     Pleura\n",
      "Klinisch sichtbare Karzinose des Peritoneums oder der Pleura\n",
      "                                                  ^^^^^^^^^^\n",
      "01_magenkarzinom_0177.tsv 7 fragment: False\n",
      "...........\n",
      "\n",
      "palliativmedizinische Betreuung / Mitbetreuung unter Einbeziehen der Angehörigen und                       spezialisierten\n",
      "palliativmedizinische Betreuung / Mitbetreuung unter Einbeziehen der Angehörigen und unter Einbeziehen der spezialisierten\n",
      "                                                                                     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\n",
      "01_magenkarzinom_0165.tsv 1 fragment: False\n",
      "...........\n",
      "\n",
      "Entstehung von sekundären Zweitkarzinomen in der Lunge oder    dem Ösophagus\n",
      "Entstehung von sekundären Zweitkarzinomen in der Lunge oder in dem Ösophagus\n",
      "                                                            ^^^^^^          \n",
      "03_lungenkarzinom_0045.tsv 12 fragment: False\n",
      "...........\n",
      "\n",
      "Radiotherapie des Beckens mit einer Dosis von zunächst 10 x 3 Gy und     anschließend nochmals 10 x 2,5 Gy\n",
      "Radiotherapie des Beckens mit einer Dosis von zunächst 10 x 3 Gy und mit anschließend nochmals 10 x 2,5 Gy\n",
      "                                                                     ^^^^^^^^^^^^^^^^                     \n",
      "07_malignes-melanom_0230.tsv 7 fragment: False\n",
      "...........\n",
      "\n",
      "Präparaten komplementären             Medizin\n",
      "Präparaten komplementären Medizin der Medizin\n",
      "           ^^^^^^^^^^^^^^^^^^^^^^ ^^^^^^^^^^^\n",
      "07_malignes-melanom_0279.tsv 8 fragment: False\n",
      "...........\n",
      "\n"
     ]
    }
   ],
   "source": [
    "for _ in range(0, 10):\n",
    "    show_random(results[~results.fragment])"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "7b3d091b-8ee2-48da-8beb-96ba39b172e0",
   "metadata": {},
   "source": [
    "### Evaluation of NER models on these instances"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "id": "399f7b28-2293-4775-aba0-06216477b681",
   "metadata": {},
   "outputs": [],
   "source": [
    "idx = set(results.set_index(['file', 'sentence_id']).index.unique().values)\n",
    "idx_frag = set(results[results.fragment].set_index(['file', 'sentence_id']).index.unique().values)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "id": "5ecc9574-2f2b-4fd6-ada2-073ac98d15d3",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "083811d3ada34b12af366fc10310c914",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "  0%|          | 0/14 [00:00<?, ?ba/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "54f8c235fa064e4bbba973fab908a09b",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "  0%|          | 0/14 [00:00<?, ?ba/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "8473e5273b284a438e846cfc6000c956",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "  0%|          | 0/14 [00:00<?, ?ba/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "0ce62607abe54c71b78b048a90a37fe7",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "  0%|          | 0/14 [00:00<?, ?ba/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "test_ds_short_ellipsis = test_ds_short.filter(lambda item: (item['fname'], item['sentence_id']) in idx)\n",
    "test_ds_short_fragment = test_ds_short.filter(lambda item: (item['fname'], item['sentence_id']) in idx_frag)\n",
    "test_ds_long_ellipsis = test_ds_long.filter(lambda item: (item['fname'], item['sentence_id']) in idx)\n",
    "test_ds_long_fragment = test_ds_long.filter(lambda item: (item['fname'], item['sentence_id']) in idx_frag)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "id": "71f539c2-3b0e-4505-b4cf-a29d468898e9",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "740it [00:00, 1867.93it/s]\n",
      "542it [00:00, 1926.76it/s]\n",
      "740it [00:00, 1821.35it/s]\n",
      "542it [00:00, 1887.95it/s]\n"
     ]
    }
   ],
   "source": [
    "test_set_metrics_short_ellipses = eval_on_test_set(test_ds_short_ellipsis, tr_short, tokenizer, \"test_short_ellipsis\")\n",
    "test_set_metrics_short_fragment = eval_on_test_set(test_ds_short_fragment, tr_short, tokenizer, \"test_short_fragment\")\n",
    "test_set_metrics_long_ellipses = eval_on_test_set(test_ds_long_ellipsis, tr_long, tokenizer, \"test_long_ellipsis\")\n",
    "test_set_metrics_long_fragment = eval_on_test_set(test_ds_long_fragment, tr_long, tokenizer, \"test_long_fragment\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "id": "bcf25bc8-0a4f-4a6a-992e-6824832a696c",
   "metadata": {},
   "outputs": [],
   "source": [
    "error_cols = ['false_negative', 'false_positive', 'labeling_error', 'boundary_error', 'label_boundary_error']"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "id": "8c8246cb-704d-4554-adbd-1f8fd2df0e7c",
   "metadata": {},
   "outputs": [],
   "source": [
    "def get_dataframe(metrics):\n",
    "    count = 0\n",
    "    for k, v in metrics.items():\n",
    "        if '_number' in k:\n",
    "            count += v\n",
    "    data = {}\n",
    "    key = None\n",
    "    for k_long, v in metrics.items():\n",
    "        subset, k = tuple(k_long.split('/'))\n",
    "        assert not key or key == subset\n",
    "        key = subset\n",
    "        if k in error_cols + ['true_positive']:\n",
    "            data[f\"{k}_abs\"] = v\n",
    "            data[f\"{k}\"] = v / count\n",
    "        if k == 'overall_f1':\n",
    "            data[k] = v\n",
    "    data['subset'] = subset\n",
    "    return pd.DataFrame([data]).set_index('subset')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "id": "7ff694ae-a5dd-4710-9af8-ba688c975c04",
   "metadata": {},
   "outputs": [],
   "source": [
    "stats = pd.concat([\n",
    "    get_dataframe(test_set_metrics_short),\n",
    "    get_dataframe(test_set_metrics_short_ellipses),\n",
    "    get_dataframe(test_set_metrics_short_fragment),\n",
    "    get_dataframe(test_set_metrics_long),\n",
    "    get_dataframe(test_set_metrics_long_ellipses),\n",
    "    get_dataframe(test_set_metrics_long_fragment),\n",
    "])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 33,
   "id": "2e200e64-ee5a-41a5-a1f6-73fba8085650",
   "metadata": {},
   "outputs": [],
   "source": [
    "import matplotlib\n",
    "from matplotlib import pyplot as plt\n",
    "from matplotlib import cm"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 34,
   "id": "c0eee753-1263-4e04-8649-1b0777850931",
   "metadata": {},
   "outputs": [],
   "source": [
    "matplotlib.rcParams.update({'font.size': 16})"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 38,
   "id": "f73ec690-dd09-4ca9-a176-b2641d3eb346",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 864x180 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "ax = plt.figure(figsize=(12,2.5)).add_subplot(111)\n",
    "\n",
    "stats[error_cols].T.rename(\n",
    "    {\n",
    "        'false_negative' : 'False\\nNegatives',\n",
    "        'false_positive' : 'False\\nPositives',\n",
    "        'labeling_error' : 'Labeling\\nErrors',\n",
    "        'boundary_error' : 'Boundary\\nErrors',\n",
    "        'label_boundary_error' : 'Label Boundary\\nErrors',\n",
    "    }).rename({\n",
    "        'test_short_full' : 'Short spans (test set)',\n",
    "        'test_short_ellipsis' : 'Short spans (ellipses)',\n",
    "        'test_short_fragment' : 'Short spans (fragments)',\n",
    "        'test_long_full' : 'Long spans (test set)',\n",
    "        'test_long_ellipsis' : 'Long spans (ellipses)',\n",
    "        'test_long_fragment' : 'Long spans (fragments)'\n",
    "    }, axis=1\n",
    "    ).plot.bar(rot=0, ax=ax\n",
    ")\n",
    "\n",
    "ax.grid(axis='y')\n",
    "\n",
    "c1, c2, c3 = '#19439c', '#7a8e82', '#ffc67a'\n",
    "\n",
    "bars = ax.patches\n",
    "hatches = [None] * 15 + ['////'] * 15\n",
    "colors = ([c1] * 5 + [c2] * 5 + [c3] * 5) * 3\n",
    "\n",
    "for bar, hatch, color in zip(bars, hatches, colors):\n",
    "    if not hatch:\n",
    "        bar.set_color(color)\n",
    "    else:\n",
    "        bar.set_color('white')\n",
    "    bar.set_edgecolor(color)\n",
    "    bar.set_hatch(hatch)\n",
    "\n",
    "    \n",
    "ax.set_ylabel('Errors / mention')\n",
    "    \n",
    "plt.legend(bbox_to_anchor=(1.0, 1.05))\n",
    "plt.ylim([0.0,0.1])\n",
    "\n",
    "plt.savefig('error_analysis.png', dpi=600, bbox_inches=\"tight\")\n",
    "\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8631d688-e920-4bc4-b1eb-896a021c3aa3",
   "metadata": {},
   "source": [
    "### Manual inspection"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 39,
   "id": "5f21f4df-252c-4338-ba4f-7447a33e519b",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[{'entity_group': 'Diagnostic',\n",
       "  'score': 0.896983,\n",
       "  'word': 'Krebs',\n",
       "  'start': 0,\n",
       "  'end': 5},\n",
       " {'entity_group': 'Diagnostic',\n",
       "  'score': 0.50938445,\n",
       "  'word': 'Vorsorge',\n",
       "  'start': 6,\n",
       "  'end': 14},\n",
       " {'entity_group': 'Diagnostic',\n",
       "  'score': 0.9128919,\n",
       "  'word': 'Früherkennung',\n",
       "  'start': 18,\n",
       "  'end': 31}]"
      ]
     },
     "execution_count": 39,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "pipe_long('Krebs-Vorsorge / -Früherkennung')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 40,
   "id": "011f79ed-6fad-42b4-ba11-6eac1f0ba1b7",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[{'entity_group': 'Diagnosis_or_Pathology',\n",
       "  'score': 0.92164946,\n",
       "  'word': 'HPV31, 33, 45 und 51',\n",
       "  'start': 0,\n",
       "  'end': 20}]"
      ]
     },
     "execution_count": 40,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "pipe_long('HPV31, 33, 45 und 51')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 41,
   "id": "de70ed60-2450-4588-8a56-31a6f1f6b953",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[{'entity_group': 'Nutrient_or_Body_Substance',\n",
       "  'score': 0.92097646,\n",
       "  'word': 'Vitamin C',\n",
       "  'start': 0,\n",
       "  'end': 9},\n",
       " {'entity_group': 'Nutrient_or_Body_Substance',\n",
       "  'score': 0.91745543,\n",
       "  'word': 'E',\n",
       "  'start': 11,\n",
       "  'end': 12},\n",
       " {'entity_group': 'Nutrient_or_Body_Substance',\n",
       "  'score': 0.8318585,\n",
       "  'word': 'A1',\n",
       "  'start': 17,\n",
       "  'end': 19}]"
      ]
     },
     "execution_count": 41,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "pipe_long('Vitamin C, E und A1')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 42,
   "id": "4a5d7566-2c77-4a5c-b806-a33d4304e9fa",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[{'entity_group': 'Therapeutic',\n",
       "  'score': 0.9219,\n",
       "  'word': 'Chemo',\n",
       "  'start': 0,\n",
       "  'end': 5},\n",
       " {'entity_group': 'Therapeutic',\n",
       "  'score': 0.9225101,\n",
       "  'word': 'Strahlentherapie',\n",
       "  'start': 11,\n",
       "  'end': 27}]"
      ]
     },
     "execution_count": 42,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "pipe_long('Chemo- und Strahlentherapie')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 43,
   "id": "fad41fdd-3056-4712-bc75-f2c275c0346c",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[{'entity_group': 'Clinical_Drug',\n",
       "  'score': 0.92278975,\n",
       "  'word': 'BRAF',\n",
       "  'start': 0,\n",
       "  'end': 4},\n",
       " {'entity_group': 'Clinical_Drug',\n",
       "  'score': 0.92444754,\n",
       "  'word': 'MEK - Inhibitor',\n",
       "  'start': 10,\n",
       "  'end': 23}]"
      ]
     },
     "execution_count": 43,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "pipe_long('BRAF- und MEK-Inhibitor')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 44,
   "id": "d262d767-440d-4c2a-8250-971a83d2f60f",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[{'entity_group': 'Therapeutic',\n",
       "  'score': 0.6532479,\n",
       "  'word': 'zielgerichtete',\n",
       "  'start': 0,\n",
       "  'end': 14},\n",
       " {'entity_group': 'Therapeutic',\n",
       "  'score': 0.91695094,\n",
       "  'word': 'Immuntherapien',\n",
       "  'start': 19,\n",
       "  'end': 33}]"
      ]
     },
     "execution_count": 44,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "pipe_long('zielgerichtete und Immuntherapien')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 45,
   "id": "6d2ac817-4f39-453c-8c7e-6b8a4d5f2b79",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[{'entity_group': 'Diagnosis_or_Pathology',\n",
       "  'score': 0.918407,\n",
       "  'word': 'HPV - 16',\n",
       "  'start': 0,\n",
       "  'end': 6},\n",
       " {'entity_group': 'Diagnosis_or_Pathology',\n",
       "  'score': 0.9107298,\n",
       "  'word': '18 - Positivität',\n",
       "  'start': 18,\n",
       "  'end': 32}]"
      ]
     },
     "execution_count": 45,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "pipe_long('HPV-16- und/oder -18-Positivität')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 46,
   "id": "ff03c406-a1ea-46b1-8794-796fea38b7c9",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/home/Florian.Borchert/miniconda3/envs/ggponc/lib/python3.8/site-packages/transformers/pipelines/base.py:908: UserWarning: You seem to be using the pipelines sequentially on GPU. In order to maximize efficiency please use a dataset\n",
      "  warnings.warn(\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "[{'entity_group': 'Diagnosis_or_Pathology',\n",
       "  'score': 0.92019343,\n",
       "  'word': 'BRCA1',\n",
       "  'start': 0,\n",
       "  'end': 5},\n",
       " {'entity_group': 'Diagnosis_or_Pathology',\n",
       "  'score': 0.8651293,\n",
       "  'word': '2 - Mutation',\n",
       "  'start': 6,\n",
       "  'end': 16}]"
      ]
     },
     "execution_count": 46,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "pipe_long('BRCA1/2-Mutation')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 47,
   "id": "d8816cfd-0eba-4e3e-bcb0-6e1e467ebf1f",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[{'entity_group': 'Therapeutic',\n",
       "  'score': 0.92089695,\n",
       "  'word': 'Zweitlinien',\n",
       "  'start': 0,\n",
       "  'end': 11},\n",
       " {'entity_group': 'Therapeutic',\n",
       "  'score': 0.9220769,\n",
       "  'word': 'Drittliniensystem',\n",
       "  'start': 18,\n",
       "  'end': 35},\n",
       " {'entity_group': 'Therapeutic',\n",
       "  'score': 0.92307276,\n",
       "  'word': 'chemotherapie',\n",
       "  'start': 43,\n",
       "  'end': 56}]"
      ]
     },
     "execution_count": 47,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "pipe_long('Zweitlinien- oder Drittliniensystem- bzw. -chemotherapie')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "50569302-5387-4f8a-8763-08a21936e32a",
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python [conda env:ggponc]",
   "language": "python",
   "name": "conda-env-ggponc-py"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.8.0"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}