notebooks/1.2_Probing.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/gsarti/lcl23-xnlm-lab/blob/main/notebooks/1.2_Probing.ipynb)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# **Probing Transformer Internal Representations**\n",
    "\n",
    "In this notebook, we will see how to train a probing model to assess the amount of linguistic knoweldge encoded in a pre-trained version of a Transformer model. More specifically, we will test whether and how sentence-level morpho-syntactic and syntatic properites are implicitly encoded in the internal representations of a BERT model.\n",
    "\n",
    "This notebook is adapted from the experiments devised in \"*Linguistic Profiling of a Neural Language Model*\" (Miaschi et al., 2020, https://aclanthology.org/2020.coling-main.65.pdf) and is structured as follows:\n",
    "\n",
    "*  **Part 1**: Loading of the model and the dataset we will use to perform the probing experiments;\n",
    "* **Part 2**: Extraction of the internal representations from the model;\n",
    "* **Part 3**: Training of the probing model;\n",
    "* **Part 4**: Evaluation and visualization of the results."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## **Probing Tasks**\n",
    "\n",
    "“*We take a model that was trained on some task, such as a language model. We generate representations using the model, and train another classifier that takes the representations and predicts some property. If the classifier performs well, we say that the model has learned information relevant for the property.*” (Probing Classifiers: Promises, Shortcomings, and Advances, Belinkov Y., 2022, https://direct.mit.edu/coli/article/48/1/207/107571/Probing-Classifiers-Promises-Shortcomings-and).\n",
    "\n",
    "<div>\n",
    "  <img src=\"https://alemiaschi.github.io/files/probing_tasks.png\" width=\"1000\">\n",
    "</div>\n",
    "\n",
    "In this tutorial, we will be focusing on training a probing model that utilizes a linear regressor. The model leverages BERT sentence-level representations to predict the value of a set of linguistic features automatically extracted from gold annotated sentences."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## **1. Installation and imports**"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Defaulting to user installation because normal site-packages is not writeable\n",
      "Requirement already satisfied: transformers in /home/alessio/.local/lib/python3.10/site-packages (4.23.1)\n",
      "Requirement already satisfied: regex!=2019.12.17 in /home/alessio/.local/lib/python3.10/site-packages (from transformers) (2022.9.13)\n",
      "Requirement already satisfied: tqdm>=4.27 in /home/alessio/.local/lib/python3.10/site-packages (from transformers) (4.64.1)\n",
      "Requirement already satisfied: requests in /usr/lib/python3/dist-packages (from transformers) (2.25.1)\n",
      "Requirement already satisfied: filelock in /home/alessio/.local/lib/python3.10/site-packages (from transformers) (3.8.0)\n",
      "Requirement already satisfied: huggingface-hub<1.0,>=0.10.0 in /home/alessio/.local/lib/python3.10/site-packages (from transformers) (0.10.0)\n",
      "Requirement already satisfied: pyyaml>=5.1 in /usr/lib/python3/dist-packages (from transformers) (5.4.1)\n",
      "Requirement already satisfied: numpy>=1.17 in /home/alessio/.local/lib/python3.10/site-packages (from transformers) (1.23.2)\n",
      "Requirement already satisfied: tokenizers!=0.11.3,<0.14,>=0.11.1 in /home/alessio/.local/lib/python3.10/site-packages (from transformers) (0.13.1)\n",
      "Requirement already satisfied: packaging>=20.0 in /home/alessio/.local/lib/python3.10/site-packages (from transformers) (21.3)\n",
      "Requirement already satisfied: typing-extensions>=3.7.4.3 in /home/alessio/.local/lib/python3.10/site-packages (from huggingface-hub<1.0,>=0.10.0->transformers) (4.3.0)\n",
      "Requirement already satisfied: pyparsing!=3.0.5,>=2.0.2 in /usr/lib/python3/dist-packages (from packaging>=20.0->transformers) (2.4.7)\n"
     ]
    }
   ],
   "source": [
    "!pip install transformers"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "import os\n",
    "#from google.colab import drive\n",
    "\n",
    "import pickle\n",
    "\n",
    "import torch\n",
    "\n",
    "import numpy as np\n",
    "import pandas as pd\n",
    "\n",
    "import matplotlib.pyplot as plt\n",
    "import seaborn as sns\n",
    "\n",
    "# AutoModel and AutoTokenizer for loading the Transformer model (and the corresponding tokenizer) via Huggingface's Transformers library\n",
    "from transformers import AutoModel, AutoTokenizer\n",
    "\n",
    "# Imports for the probing model\n",
    "from sklearn.model_selection import train_test_split\n",
    "from sklearn.svm import LinearSVR\n",
    "from sklearn.dummy import DummyRegressor\n",
    "\n",
    "from sklearn.metrics import r2_score, mean_squared_error\n",
    "from scipy.stats import spearmanr"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## **1. Loading the model**\n",
    "\n",
    "We will use a pre-trained version of the BERT model (12 layers, 12 attention heads, 768 hidden units), available through the [Transformers](https://huggingface.co/docs/transformers/index) library. However, it is possible to use any other model from those available from the Huggingface Model: [Huggingface Models](https://huggingface.co/models).\n",
    "\n",
    "Along with the model, we also load the corresponding tokenizer.\n",
    "\n",
    "### **1.1 BERT architecture**\n",
    "\n",
    "<div>\n",
    "  <img src=\"https://media.geeksforgeeks.org/wp-content/uploads/20200420231335/elmo-eemmbeddings.jpg\" width=\"800\">\n",
    "</div>\n",
    "\n",
    "(Source: https://www.geeksforgeeks.org/explanation-of-bert-model-nlp/)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Some weights of the model checkpoint at bert-base-uncased were not used when initializing BertModel: ['cls.predictions.decoder.weight', 'cls.predictions.transform.dense.bias', 'cls.predictions.transform.dense.weight', 'cls.seq_relationship.bias', 'cls.predictions.transform.LayerNorm.weight', 'cls.predictions.bias', 'cls.seq_relationship.weight', 'cls.predictions.transform.LayerNorm.bias']\n",
      "- This IS expected if you are initializing BertModel from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).\n",
      "- This IS NOT expected if you are initializing BertModel from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).\n"
     ]
    }
   ],
   "source": [
    "model_name = \"bert-base-uncased\"\n",
    "\n",
    "device = \"cuda:0\" if torch.cuda.is_available() else \"cpu\"\n",
    "\n",
    "# Loading the tokenizer and the model via Huggingface's Transformers library\n",
    "tokenizer = AutoTokenizer.from_pretrained(model_name)\n",
    "\n",
    "# We set 'output_hidden_states' equal to 'True' because we want to extract the representations from all the layers of the model\n",
    "model = AutoModel.from_pretrained(model_name, output_hidden_states=True)\n",
    "model = model.to(device)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## **2. Dataset**\n",
    "\n",
    "In this notebook, we employ a dataset comprising sentences extracted from the GUM Corpus (Zeldes A., 2017, https://gucorpling.org/amir/pdf/GUM_paper_prepub.pdf). GUM, the Georgetown University Multilayer corpus, is an open source collection of richly annotated texts from multiple text types. \n",
    "\n",
    "We chose this dataset since it has been annotated according to the [Universal Dependencies](https://universaldependencies.org/) formalism. Therefore, it allows the extraction of gold linguistc properites from the annotated sentences.\n",
    "\n",
    "The UD English GUM is available at the following link: https://universaldependencies.org/treebanks/en_gum/index.html."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "drive.mount('/content/drive')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### **2.1. Data Investigation**\n",
    "\n",
    "Before proceeding with the data processing, let's examine the structure of the dataset more closely."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>0</th>\n",
       "      <th>1</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>GUM_academic_exposure-1</td>\n",
       "      <td>Introduction</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>GUM_academic_exposure-2</td>\n",
       "      <td>Research on adult-learned second language (L2)...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>GUM_academic_exposure-3</td>\n",
       "      <td>Of interest here, studies suggest that, despit...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>GUM_academic_exposure-4</td>\n",
       "      <td>However, it is not enough to have attained suc...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>GUM_academic_exposure-5</td>\n",
       "      <td>Crucially, it is also desirable to retain them...</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                         0                                                  1\n",
       "0  GUM_academic_exposure-1                                       Introduction\n",
       "1  GUM_academic_exposure-2  Research on adult-learned second language (L2)...\n",
       "2  GUM_academic_exposure-3  Of interest here, studies suggest that, despit...\n",
       "3  GUM_academic_exposure-4  However, it is not enough to have attained suc...\n",
       "4  GUM_academic_exposure-5  Crucially, it is also desirable to retain them..."
      ]
     },
     "execution_count": 6,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "sentences_file = \"../data/GUM_Treebank_Sentences.tsv\"\n",
    "\n",
    "df = pd.read_csv(sentences_file, delimiter='\\t', header=None)\n",
    "\n",
    "df.head()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Total number of sentences: 5153\n"
     ]
    }
   ],
   "source": [
    "tot_sentences = len(df)\n",
    "print(\"Total number of sentences:\", tot_sentences)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## **3. Extraction of the internal representations**\n",
    "\n",
    "After loading the model and the dataset, we can extract BERT's internal representations. First, we need to tokenize each individual sequence (i.e. sentence) using the tokenizer:\n",
    "\n",
    "```python\n",
    "input_ids = tokenizer.encode(<sentence>, add_special_tokens=True)\n",
    "```\n",
    "We set the parameter '*add_special_tokens*' equal to '*True*' because we want to add the '*\\[CLS]*' and '*\\[SEP]*' tokens defined in the model's vocabulary.\n",
    "\n",
    "<div>\n",
    "  <img src=\"https://miro.medium.com/v2/resize:fit:4800/format:webp/1*ktkqnCsFx1VPmCNYUOUAyw.png\" width=\"800\">\n",
    "</div>\n",
    "\n",
    "(Source: https://towardsdatascience.com/sentence-correctness-classifier-using-transfer-learning-with-huggingface-bert-8884795ba5ca)\n",
    "\n",
    "\n",
    "Then, we can extract the representations as follows:\n",
    "\n",
    "```python\n",
    "outputs = model(input_ids)\n",
    "```\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [],
   "source": [
    "embeddings = []\n",
    "ids = []\n",
    "sentences = []\n",
    "\n",
    "# Extracting the representations from each sentence of the probing dataset, with model(*input_ids*), where *input_ids* corresponds to the tokenized sentence\n",
    "with open(sentences_file, 'r') as f:\n",
    "  for line in f:\n",
    "    elements = line.rstrip('\\n').split('\\t')\n",
    "    sent_id = elements[0]\n",
    "    sentence = elements[1]\n",
    "\n",
    "    # Tokenization of the sentence\n",
    "    input_ids = tokenizer.encode(sentence, add_special_tokens=True)\n",
    "    input_ids = torch.tensor([input_ids]).to(device)\n",
    "\n",
    "    # Extraction of the representations\n",
    "    with torch.no_grad():\n",
    "      outputs = model(input_ids)\n",
    "      embeddings.append(outputs[\"hidden_states\"])\n",
    "      ids.append(sent_id)\n",
    "      sentences.append(sentence)\n",
    "\n",
    "# Saving the representations \n",
    "#with open(\"../data/GUM_Treebank_Representations.pkl\", \"wb\") as fOut:\n",
    "#  pickle.dump({'ids': ids, 'sentences': sentences, 'embeddings': embeddings}, fOut)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### **3.1. Exploring Model output**\n",
    "\n",
    "In the following, we examine more closely the output of the extraction process (e.g. number of layers, struture of the output).\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Tokens: ['[CLS]', 'the', 'reason', 'for', 'it', 'is', 'because', 'people', 'were', 'fighting', 'for', 'our', 'land', ',', 'either', 'defending', 'it', 'or', 'extending', 'it', ',', 'because', 'that', 'was', 'the', 'main', 'source', '.', '[SEP]']\n",
      "\n",
      "Numer of layers: 13\n",
      "\n",
      "tensor([[[ 0.0757,  0.0164, -0.1568,  ..., -0.3808,  0.2195,  0.4642],\n",
      "         [ 0.0765, -0.1931, -0.1814,  ..., -0.1002,  0.5683,  0.4332],\n",
      "         [-0.8869,  0.4616, -0.1631,  ..., -0.0199, -0.5137,  0.3742],\n",
      "         ...,\n",
      "         [ 0.3685,  0.4208, -0.4885,  ..., -0.2029, -0.4166, -0.1863],\n",
      "         [ 0.6282,  0.2763, -0.3411,  ...,  0.0072, -0.4181, -0.4368],\n",
      "         [ 0.0877,  0.6951,  0.1139,  ..., -0.0176, -0.3817, -0.1705]]])\n",
      "\n",
      "Tensor dimension: torch.Size([1, 29, 768])\n",
      "\n",
      "tensor([-8.8689e-01,  4.6162e-01, -1.6312e-01, -3.7560e-01,  6.2536e-01,\n",
      "         3.0484e-01,  6.2627e-01,  1.1932e+00, -6.8599e-01, -3.9991e-01,\n",
      "         3.6593e-01, -6.5737e-01, -7.5233e-01,  7.6856e-01, -1.6948e-01,\n",
      "         2.5790e-01,  7.2615e-02,  2.4460e-01,  1.8574e-01,  1.0515e+00,\n",
      "         4.7979e-01,  1.7964e-01,  5.6881e-01,  1.0912e+00,  8.2771e-01,\n",
      "        -2.2045e-01, -6.0624e-01, -5.5395e-01, -5.9157e-01, -5.5416e-01,\n",
      "         1.2422e+00,  4.2939e-01, -5.2975e-01, -2.8975e-01, -7.6108e-01,\n",
      "        -4.3600e-01, -6.5851e-01,  4.0318e-01, -5.6188e-02,  1.1439e+00,\n",
      "        -5.7002e-01, -8.1634e-01, -7.5649e-01,  9.0025e-01,  4.5174e-01,\n",
      "        -2.8057e-01, -2.4821e-01,  3.5793e-01,  4.6069e-01, -1.2844e+00,\n",
      "        -1.0662e-01,  7.2555e-02,  1.9712e-02, -1.5112e-01,  4.2486e-01,\n",
      "         6.4505e-01, -7.2785e-01, -2.3597e-01, -9.6867e-01, -7.3171e-01,\n",
      "        -2.5401e-01, -6.5173e-01, -2.3090e-01, -6.7419e-01, -1.4728e-01,\n",
      "         1.5158e-01,  1.5260e-05,  3.5710e-01, -4.8172e-01, -2.3353e-02,\n",
      "        -7.2866e-01, -1.9507e-01, -6.9101e-01, -4.7274e-02, -1.3086e+00,\n",
      "         1.0223e+00,  3.0634e-01,  9.4845e-01, -2.1485e-02, -6.0983e-01,\n",
      "         9.9601e-01,  6.0455e-01, -6.0105e-01,  4.3987e-01, -1.2039e-01,\n",
      "         3.7162e-01, -3.1497e-01, -7.9345e-02, -5.6276e-01,  9.5663e-01,\n",
      "        -1.5731e-01,  3.1029e-02, -2.2261e-01,  8.4542e-01, -4.4144e-01,\n",
      "        -3.0649e-01,  1.2388e+00, -4.9794e-02, -4.7938e-01, -4.4441e-01,\n",
      "         8.7742e-01, -2.9527e-01, -3.7444e-01, -2.1991e-01,  5.5906e-01,\n",
      "        -6.8097e-01,  6.5606e-01, -1.0658e-01, -2.8296e-01,  2.2709e-01,\n",
      "        -4.7963e-03, -1.8436e-01, -9.9518e-01, -4.0586e-01, -2.1806e-01,\n",
      "         2.3755e-01, -6.8554e-01, -6.5495e-01,  4.4421e-01,  9.9641e-01,\n",
      "         4.8590e-01,  2.2299e-01, -7.3065e-01,  7.1811e-01, -9.9088e-01,\n",
      "        -4.5623e-01, -4.1097e-01,  4.6399e-01, -8.4742e-01,  8.7657e-03,\n",
      "         6.3189e-01, -1.0988e-01,  9.2061e-01, -9.0661e-01,  3.5616e-01,\n",
      "        -1.1775e-02, -2.1773e-01, -6.1928e-02, -8.2727e-01, -3.9356e-01,\n",
      "        -2.4149e-01, -3.8317e-01,  3.6991e-01, -3.9914e-01, -6.2335e-01,\n",
      "        -2.8854e-01, -3.0410e-01, -2.2176e-01,  8.4785e-01, -4.9842e-01,\n",
      "        -1.9041e-01,  9.9188e-01, -6.4972e-04, -4.6294e-01, -6.4075e-01,\n",
      "        -1.3691e-01, -1.2056e+00, -1.9095e-01,  5.8702e-01,  2.7563e-01,\n",
      "         5.3265e-02, -5.3518e-01,  4.7180e-02,  1.6059e-01, -2.4778e-01,\n",
      "         3.5395e-01, -1.4133e-01,  2.3929e-01, -1.3973e+00,  5.8570e-01,\n",
      "        -8.2640e-01,  1.5627e-01,  1.3354e+00, -1.9540e-01,  2.5914e-01,\n",
      "        -1.4908e-01, -4.2956e-01,  3.1177e-01,  4.9603e-01,  9.4637e-01,\n",
      "        -3.6022e-01,  1.3880e+00,  7.4887e-02,  1.2753e-01,  7.1762e-01,\n",
      "         6.2483e-03, -4.6670e-02,  2.8182e-01, -8.6589e-02, -4.4108e-01,\n",
      "        -1.0606e+00, -4.2504e-01, -2.6398e-01,  8.9281e-02,  1.5395e-01,\n",
      "        -2.9603e-01, -1.4397e-01, -2.9412e-01,  2.4177e-01,  1.7527e-01,\n",
      "        -2.8904e-01, -4.5109e-01,  6.2378e-01, -1.7771e-01,  1.1817e-01,\n",
      "         2.5951e-01,  1.2017e-01, -4.0744e-01,  6.3800e-01,  2.7594e-02,\n",
      "         2.5581e-02,  7.0599e-01,  7.2401e-01, -5.2112e-01,  1.7535e-02,\n",
      "        -1.5259e-01,  7.6868e-01, -7.5319e-02,  3.2853e-01, -6.8469e-01,\n",
      "         7.5285e-02, -1.2509e-01, -8.3072e-01,  1.2909e+00, -2.5451e-01,\n",
      "         7.0597e-01,  2.3599e-01, -2.6441e-01,  2.5542e-01,  4.1602e-01,\n",
      "         7.8051e-01,  3.8789e-01, -2.3593e-01, -1.0508e-01,  1.6241e-01,\n",
      "        -8.6722e-01, -1.4379e-01,  4.3389e-02, -1.5605e-01, -4.7384e-01,\n",
      "         1.6784e-01,  2.4427e-01, -2.6608e-01, -2.5202e-01, -4.1172e-01,\n",
      "        -1.9139e-01,  4.7579e-01, -6.6900e-02, -6.4621e-01, -9.5139e-01,\n",
      "        -3.1347e-01, -1.1461e-01,  3.5665e-01, -1.8074e-01, -4.6301e-01,\n",
      "         5.2472e-01, -3.4603e-01, -2.3420e-01,  2.4082e-01, -8.3063e-02,\n",
      "         1.1963e-01, -4.4053e-01, -5.8176e-01,  9.6474e-01, -9.3839e-01,\n",
      "        -1.1759e+00,  9.1406e-01,  8.0471e-01, -1.7505e-01, -2.8044e-01,\n",
      "         9.0666e-01,  9.9142e-03,  1.7171e-01,  1.0924e+00, -5.3539e-01,\n",
      "        -7.1461e-01,  7.8809e-01,  3.5852e-01, -6.6015e-01, -3.6229e-01,\n",
      "         2.7044e-01,  7.9231e-01, -1.0728e+00,  3.1018e-01, -2.7117e-01,\n",
      "         3.4018e-02,  2.1491e-01,  2.1021e-01, -1.4574e+00,  4.3419e-01,\n",
      "        -2.2977e-01, -7.4124e-01, -5.7790e-01, -4.6595e-01,  1.3811e+00,\n",
      "         8.0354e-02, -5.8226e-01,  1.8795e-02,  8.5542e-01, -3.6165e-02,\n",
      "         1.7711e-01, -8.7344e-02, -5.4506e-01,  7.5691e-01, -7.4697e-01,\n",
      "        -1.6645e-01,  3.2336e-01,  4.0403e-01, -2.9824e+00, -1.8571e-01,\n",
      "         1.6002e-01, -1.8324e-01,  4.5940e-01, -3.1621e-01,  4.8810e-01,\n",
      "         9.3929e-02, -6.9615e-01, -1.8258e-01, -8.7667e-01, -3.9650e-01,\n",
      "         1.5732e-01,  2.9638e-01,  4.5179e-01,  5.9018e-02,  4.9135e-01,\n",
      "        -9.0469e-02, -2.5207e-01,  7.2817e-01, -7.4595e-02, -5.4192e-01,\n",
      "        -4.6692e-01, -7.9291e-01,  1.4317e-01,  1.0829e+00, -1.0636e+00,\n",
      "        -3.9125e-01,  1.0472e-01, -1.1853e-01,  1.2594e-01, -1.1997e+00,\n",
      "         5.7903e-01,  1.0330e+00,  2.9357e-01, -4.7107e-01, -4.6589e-01,\n",
      "        -6.2332e-01, -2.7425e-01,  5.1200e-02, -6.4073e-01, -1.2007e+00,\n",
      "         2.4160e-01,  7.5847e-01,  1.2595e+00,  2.0445e-01, -1.5124e-01,\n",
      "        -6.4600e-01, -1.8826e-01,  5.4897e-01,  4.8759e-01,  1.1577e+00,\n",
      "        -1.2373e-01,  3.8587e-01,  4.8412e-02, -3.7691e-01,  1.2528e+00,\n",
      "         3.2897e-01,  1.0335e-01,  4.7511e-02, -4.0748e-01,  1.6839e-01,\n",
      "        -4.9066e-01, -2.0847e-02, -1.1327e-01, -6.0732e-01, -1.5008e+00,\n",
      "        -9.0335e-01,  5.6204e-02,  5.8984e-02,  4.6811e-01,  1.1824e+00,\n",
      "        -1.3121e+00, -9.7844e-01, -1.4787e-01, -7.7212e-01,  4.9594e-01,\n",
      "        -6.3983e-01,  3.0577e-01,  3.4643e-01,  3.1641e-01, -1.4972e-01,\n",
      "        -7.9557e-01, -2.2401e-01, -3.3397e-01, -8.1038e-01,  1.7444e-01,\n",
      "        -6.6365e-01,  2.5340e-01,  5.0280e-02,  2.6711e-01, -7.2260e-01,\n",
      "        -6.4208e-01,  1.4567e-01,  3.2944e-01, -3.9091e-01,  2.2873e-01,\n",
      "        -2.6489e-01,  1.2160e+00,  5.0004e-01,  1.7131e-01,  1.4718e-01,\n",
      "         3.0265e-01,  3.3074e-01,  5.5674e-01, -5.7220e-01,  3.5820e-01,\n",
      "         3.7831e-01,  3.3191e-01, -4.2188e-01,  1.4464e+00, -4.9862e-02,\n",
      "         9.3532e-01, -4.0217e-01, -5.5264e-01,  6.6094e-01,  1.1537e+00,\n",
      "         5.6713e-01,  4.9742e-01, -8.4295e-01,  8.2769e-01,  1.0226e+00,\n",
      "        -7.1785e-01,  1.0653e-02, -1.1078e+00,  4.7600e-01,  1.3361e-01,\n",
      "         2.0431e-01, -2.1441e-01,  7.8424e-01,  2.5054e-01, -6.4707e-01,\n",
      "        -6.2654e-01,  9.4697e-01,  4.4451e-01,  4.4797e-01,  8.7630e-02,\n",
      "        -3.4245e-01,  6.7209e-02, -5.3551e-02,  1.6958e+00,  3.5177e-01,\n",
      "         3.4269e-01, -2.4533e-01, -4.6743e-01,  4.6221e-01, -2.7190e-01,\n",
      "        -1.9692e-01, -1.4494e-01, -2.1775e-01, -1.0184e+00, -7.0530e-01,\n",
      "        -1.0078e-01,  5.1885e-01, -6.9753e-01,  9.4294e-02,  1.1290e+00,\n",
      "         4.7087e-01, -3.3207e-01, -1.4075e+00, -4.5346e-01,  2.8070e-01,\n",
      "        -5.9079e-01,  3.3931e-01, -1.1578e-01,  1.6162e+00, -5.2354e-01,\n",
      "        -7.5400e-01,  8.6542e-01,  1.8928e-02, -6.9786e-01,  2.8165e-01,\n",
      "        -2.9998e-01, -3.0007e-01,  2.5624e-01, -3.0052e-01, -7.4053e-02,\n",
      "        -6.7802e-01, -8.1716e-01,  8.7106e-01, -2.0580e-01,  1.7915e-01,\n",
      "        -8.8339e-01, -8.7272e-01,  4.1625e-01,  5.0437e-01,  7.7468e-02,\n",
      "         2.9441e-01,  1.6952e-01, -2.7101e-01,  1.0166e+00,  4.1739e-02,\n",
      "        -4.9443e-01,  3.7585e-01, -4.3510e-01,  2.3324e-01,  1.9105e-01,\n",
      "         2.0195e-01,  5.0202e-01,  8.8878e-01,  2.0703e-01,  4.8689e-01,\n",
      "        -6.1389e-02, -1.8884e-01,  1.8858e-01, -4.4768e-01,  2.6839e-01,\n",
      "         1.0257e+00,  2.9537e-01,  2.3258e-01, -1.2980e+00, -2.4698e-01,\n",
      "        -5.4368e-01,  5.3721e-01,  7.2024e-01, -9.7138e-01,  9.0183e-01,\n",
      "        -8.2023e-01,  4.2706e-02,  3.6325e-01, -3.7400e-01, -1.8998e-01,\n",
      "        -6.6465e-01,  2.1249e-01, -6.0318e-01, -5.9308e-01, -1.1469e+00,\n",
      "         3.5020e-01, -9.3164e-01, -8.6689e-01, -2.5582e-02, -2.9916e-01,\n",
      "         8.0844e-01, -1.5710e-01, -8.2289e-02,  8.8280e-01, -4.5202e-02,\n",
      "        -9.3130e-02,  2.9727e-01, -2.2996e-01, -5.5478e-01, -8.1902e-02,\n",
      "         1.2100e-01,  2.9092e-01,  2.8683e-01, -1.1117e-01,  7.4796e-02,\n",
      "         8.9186e-01, -8.7591e-01,  6.9814e-01,  2.2784e-02, -2.5837e-01,\n",
      "        -5.6330e-01,  1.8747e-01,  2.0714e-01, -1.0203e+00, -6.8160e-01,\n",
      "        -6.8584e-01, -1.1952e-01, -4.4903e-01, -1.6636e-02,  4.6399e-01,\n",
      "         3.6084e-01,  3.0111e-01, -4.4724e-01, -2.0467e-01,  1.6023e-01,\n",
      "         9.8974e-01,  1.6133e-02, -9.2056e-01,  2.8329e-01,  1.0734e-01,\n",
      "        -8.0274e-02, -2.1328e-01,  7.6150e-01, -6.1439e-01,  3.1615e-02,\n",
      "        -9.0405e-02,  7.1440e-01, -1.0017e+00, -3.6563e-02, -1.5175e-01,\n",
      "        -5.5372e-01, -2.4234e-01,  4.3537e-01,  1.7306e-01, -4.6495e-01,\n",
      "         8.4414e-01, -5.2540e-01,  1.6868e-01, -6.6339e-01,  5.3519e-01,\n",
      "         3.8812e-01, -8.7393e-01,  1.2671e-01, -5.0444e-01,  2.0631e-01,\n",
      "         1.8576e-01, -2.6289e-01,  4.7080e-02,  1.8985e-02, -1.2163e-01,\n",
      "         2.9266e-01,  2.9242e-01,  1.3010e-01, -1.3686e-01,  5.1016e-01,\n",
      "        -3.4924e-01,  5.7062e-01,  2.8064e-02,  8.8244e-01, -8.6630e-01,\n",
      "         1.1532e+00,  8.0274e-01,  3.6251e-01, -4.6264e-01,  3.4409e-01,\n",
      "        -3.1904e-01, -5.5277e-01,  1.4292e-01,  5.4333e-02,  1.7952e-01,\n",
      "         9.0363e-01,  5.0702e-03,  1.3006e-01,  3.2163e-01, -6.9863e-01,\n",
      "        -7.0834e-01, -7.2493e-01, -3.2035e-01, -4.3060e-01,  2.7226e-01,\n",
      "        -2.5552e-01,  9.9147e-03, -1.8004e-01, -8.6135e-01,  7.6104e-02,\n",
      "         5.8861e-01,  6.6073e-01,  9.5726e-02,  3.0870e-01,  3.9524e-02,\n",
      "         1.1581e+00,  1.5917e-01,  1.0326e+00, -1.8883e-01,  9.7593e-02,\n",
      "        -3.6733e-01,  9.1084e-01,  1.3527e-01, -4.0886e-01,  7.8937e-01,\n",
      "        -8.7728e-01, -5.5689e-01,  1.1760e+00, -3.1969e-01,  6.3176e-01,\n",
      "         2.2646e-02,  6.1935e-01,  9.8500e-01, -1.1401e-01, -3.1806e-02,\n",
      "         5.5669e-01, -8.8649e-01,  3.6074e-01,  8.9117e-01, -5.2979e-01,\n",
      "         2.0645e-01,  1.8548e-01, -4.8730e-01,  6.7940e-01,  2.9054e-01,\n",
      "         3.6395e-01, -5.4437e-01, -4.4531e-01, -1.0691e-01,  3.1313e-01,\n",
      "        -9.8515e-01, -4.1047e-01,  6.2792e-01,  2.5729e-01, -1.8080e-02,\n",
      "        -4.9675e-01,  2.3457e-01, -3.8436e-01, -1.2190e+00, -2.4886e-01,\n",
      "         3.7046e-02, -8.1894e-02, -2.1963e-02,  6.7541e-02, -2.7778e-01,\n",
      "        -3.9511e-01, -2.6203e-02,  7.3240e-01,  3.2022e-01,  1.1633e+00,\n",
      "         3.4343e-01,  2.8330e-02,  3.4283e-02,  7.4158e-01, -2.2219e-01,\n",
      "        -2.5401e-01, -3.8609e-02,  1.4499e-02,  4.2961e-01,  7.7811e-01,\n",
      "        -1.1189e-01, -4.2716e-01, -4.7100e-01,  5.6098e-01,  9.9822e-01,\n",
      "        -5.6586e-01,  3.4152e-01, -4.0030e-01,  3.7160e-01,  1.2657e+00,\n",
      "        -7.4294e-01, -2.5856e-01,  1.0511e+00, -5.1448e-01, -2.8305e-01,\n",
      "        -1.2752e+00,  3.9430e-01,  1.7865e-01, -1.1391e+00, -2.5375e-01,\n",
      "         2.9433e-01,  3.3161e-02, -5.9093e-01, -4.9724e-01, -4.7100e-01,\n",
      "         5.6664e-01, -3.3220e-03, -8.7955e-02, -9.2679e-03, -6.0545e-01,\n",
      "        -4.4211e-02, -1.0034e+00, -6.1620e-01,  1.5603e+00, -3.5445e-01,\n",
      "         2.3883e-01, -1.1375e+00,  1.7564e+00, -6.4688e-02, -1.1283e-01,\n",
      "        -2.6293e-01,  4.8747e-02, -1.9597e-01,  2.5200e-02, -1.0024e+00,\n",
      "         7.5098e-02, -1.2834e-01, -2.0527e-01,  1.0887e-01,  4.6331e-02,\n",
      "        -1.9926e-02, -5.1366e-01,  3.7418e-01])\n"
     ]
    }
   ],
   "source": [
    "sample_sentence = sentences[300]\n",
    "sample_tokens = tokenizer.tokenize(sample_sentence, add_special_tokens=True)\n",
    "sample_embedding = embeddings[300]\n",
    "\n",
    "# Visualize the output of the tokenization process on the sample sentence extracted from the dataset\n",
    "print(\"Tokens:\", sample_tokens)\n",
    "print()\n",
    "\n",
    "# Number of layers of the model\n",
    "n_layers = len(sample_embedding)\n",
    "print(\"Numer of layers:\", n_layers)\n",
    "print()\n",
    "\n",
    "# Word representations of a specific selected layer\n",
    "layer = 12\n",
    "hidden_state = sample_embedding[layer]\n",
    "print(hidden_state)\n",
    "print()\n",
    "\n",
    "# Visualize the dimension of the 'hidden_state' object (i.e. number of sentences * number of tokens * number of hidden units)\n",
    "dim_hidden_state = hidden_state.shape\n",
    "print(\"Tensor dimension:\", dim_hidden_state)\n",
    "print()\n",
    "\n",
    "# Visualize the embedding of a specific token of the selected sample sentence\n",
    "word_embeddings = {k:v for k, v in zip(sample_tokens, hidden_state[0])}\n",
    "print(word_embeddings[\"reason\"])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## **4. Extraction of the linguistic (probing) features**\n",
    "\n",
    "As already mentioned, each probing task will consist in predicting the value of a set of linguistic features automatically extracted from the GUM annotated sentences.\n",
    "\n",
    "The set of features used in the notebook was extracted using ***Profiling-UD*** (Brunato et al., 2020, http://www.lrec-conf.org/proceedings/lrec2020/pdf/2020.lrec-1.883.pdf), a text analysis tool inspired to the principles of linguistic profiling that allows the extraction of more than 130 features, spanning across different levels of linguistic annotation. \n",
    "\n",
    "In particular, these features model linguistic phenomena ranging from raw text ones, to morpho–syntactic information and inflectional properties of verbs, to more complex aspects of sentence structure modeling global and local properties of the whole parsed tree and of specific subtrees, such as the order of subjects and objects with respect to the verb, the distribution of UD syntactic relations, and more.\n",
    "\n",
    "<div>\n",
    "  <img src=\"https://alemiaschi.github.io/files/features.png\" width=\"1000\">\n",
    "</div>\n",
    "\n",
    "\n",
    "Profiling-UD is available at the following link: http://linguistic-profiling.italianlp.it/. "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### **4.1. Loading the dataset with the linguistic features**\n",
    "\n",
    "In the following, we load and explore our dataset, which has already processed with the Profiling-UD tool. However, it is possible to replicate these experiments by using a different set of sentences and uploading them to Profiling-UD in order to extract the linguistic features to probe the Transformer model."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>n_sentences</th>\n",
       "      <th>n_tokens</th>\n",
       "      <th>tokens_per_sent</th>\n",
       "      <th>char_per_tok</th>\n",
       "      <th>upos_dist_DET</th>\n",
       "      <th>upos_dist_ADV</th>\n",
       "      <th>upos_dist_VERB</th>\n",
       "      <th>upos_dist_NOUN</th>\n",
       "      <th>upos_dist_PART</th>\n",
       "      <th>upos_dist_NUM</th>\n",
       "      <th>...</th>\n",
       "      <th>subordinate_pre</th>\n",
       "      <th>subordinate_post</th>\n",
       "      <th>verb_edges_dist_1</th>\n",
       "      <th>verb_edges_dist_3</th>\n",
       "      <th>verb_edges_dist_4</th>\n",
       "      <th>verb_edges_dist_2</th>\n",
       "      <th>verb_edges_dist_0</th>\n",
       "      <th>verb_edges_dist_6</th>\n",
       "      <th>verb_edges_dist_5</th>\n",
       "      <th>avg_verb_edges</th>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>identifier</th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>GUM_academic_exposure-1</th>\n",
       "      <td>1</td>\n",
       "      <td>1</td>\n",
       "      <td>1.0</td>\n",
       "      <td>12.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>100.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>...</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>GUM_academic_exposure-2</th>\n",
       "      <td>1</td>\n",
       "      <td>32</td>\n",
       "      <td>32.0</td>\n",
       "      <td>5.960000</td>\n",
       "      <td>6.250000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>6.250000</td>\n",
       "      <td>28.125000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>6.250000</td>\n",
       "      <td>...</td>\n",
       "      <td>0.0</td>\n",
       "      <td>100.0</td>\n",
       "      <td>50.0</td>\n",
       "      <td>50.0</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>2.000000</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>GUM_academic_exposure-3</th>\n",
       "      <td>1</td>\n",
       "      <td>35</td>\n",
       "      <td>35.0</td>\n",
       "      <td>5.740741</td>\n",
       "      <td>2.857143</td>\n",
       "      <td>2.857143</td>\n",
       "      <td>8.571429</td>\n",
       "      <td>25.714286</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>5.714286</td>\n",
       "      <td>...</td>\n",
       "      <td>0.0</td>\n",
       "      <td>100.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>33.333333</td>\n",
       "      <td>33.333333</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>33.333333</td>\n",
       "      <td>3.666667</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>GUM_academic_exposure-4</th>\n",
       "      <td>1</td>\n",
       "      <td>13</td>\n",
       "      <td>13.0</td>\n",
       "      <td>5.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>7.692308</td>\n",
       "      <td>7.692308</td>\n",
       "      <td>7.692308</td>\n",
       "      <td>15.384615</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>...</td>\n",
       "      <td>0.0</td>\n",
       "      <td>100.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>100.000000</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>2.000000</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>GUM_academic_exposure-5</th>\n",
       "      <td>1</td>\n",
       "      <td>23</td>\n",
       "      <td>23.0</td>\n",
       "      <td>4.500000</td>\n",
       "      <td>8.695652</td>\n",
       "      <td>13.043478</td>\n",
       "      <td>4.347826</td>\n",
       "      <td>17.391304</td>\n",
       "      <td>4.347826</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>...</td>\n",
       "      <td>0.0</td>\n",
       "      <td>100.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>100.000000</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>2.000000</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "<p>5 rows × 144 columns</p>\n",
       "</div>"
      ],
      "text/plain": [
       "                         n_sentences  n_tokens  tokens_per_sent  char_per_tok  \\\n",
       "identifier                                                                      \n",
       "GUM_academic_exposure-1            1         1              1.0     12.000000   \n",
       "GUM_academic_exposure-2            1        32             32.0      5.960000   \n",
       "GUM_academic_exposure-3            1        35             35.0      5.740741   \n",
       "GUM_academic_exposure-4            1        13             13.0      5.000000   \n",
       "GUM_academic_exposure-5            1        23             23.0      4.500000   \n",
       "\n",
       "                         upos_dist_DET  upos_dist_ADV  upos_dist_VERB  \\\n",
       "identifier                                                              \n",
       "GUM_academic_exposure-1       0.000000       0.000000        0.000000   \n",
       "GUM_academic_exposure-2       6.250000       0.000000        6.250000   \n",
       "GUM_academic_exposure-3       2.857143       2.857143        8.571429   \n",
       "GUM_academic_exposure-4       0.000000       7.692308        7.692308   \n",
       "GUM_academic_exposure-5       8.695652      13.043478        4.347826   \n",
       "\n",
       "                         upos_dist_NOUN  upos_dist_PART  upos_dist_NUM  ...  \\\n",
       "identifier                                                              ...   \n",
       "GUM_academic_exposure-1      100.000000        0.000000       0.000000  ...   \n",
       "GUM_academic_exposure-2       28.125000        0.000000       6.250000  ...   \n",
       "GUM_academic_exposure-3       25.714286        0.000000       5.714286  ...   \n",
       "GUM_academic_exposure-4        7.692308       15.384615       0.000000  ...   \n",
       "GUM_academic_exposure-5       17.391304        4.347826       0.000000  ...   \n",
       "\n",
       "                         subordinate_pre  subordinate_post  verb_edges_dist_1  \\\n",
       "identifier                                                                      \n",
       "GUM_academic_exposure-1              0.0               0.0                0.0   \n",
       "GUM_academic_exposure-2              0.0             100.0               50.0   \n",
       "GUM_academic_exposure-3              0.0             100.0                0.0   \n",
       "GUM_academic_exposure-4              0.0             100.0                0.0   \n",
       "GUM_academic_exposure-5              0.0             100.0                0.0   \n",
       "\n",
       "                         verb_edges_dist_3  verb_edges_dist_4  \\\n",
       "identifier                                                      \n",
       "GUM_academic_exposure-1                0.0           0.000000   \n",
       "GUM_academic_exposure-2               50.0           0.000000   \n",
       "GUM_academic_exposure-3                0.0          33.333333   \n",
       "GUM_academic_exposure-4                0.0           0.000000   \n",
       "GUM_academic_exposure-5                0.0           0.000000   \n",
       "\n",
       "                         verb_edges_dist_2  verb_edges_dist_0  \\\n",
       "identifier                                                      \n",
       "GUM_academic_exposure-1           0.000000                0.0   \n",
       "GUM_academic_exposure-2           0.000000                0.0   \n",
       "GUM_academic_exposure-3          33.333333                0.0   \n",
       "GUM_academic_exposure-4         100.000000                0.0   \n",
       "GUM_academic_exposure-5         100.000000                0.0   \n",
       "\n",
       "                         verb_edges_dist_6  verb_edges_dist_5  avg_verb_edges  \n",
       "identifier                                                                     \n",
       "GUM_academic_exposure-1                0.0           0.000000        0.000000  \n",
       "GUM_academic_exposure-2                0.0           0.000000        2.000000  \n",
       "GUM_academic_exposure-3                0.0          33.333333        3.666667  \n",
       "GUM_academic_exposure-4                0.0           0.000000        2.000000  \n",
       "GUM_academic_exposure-5                0.0           0.000000        2.000000  \n",
       "\n",
       "[5 rows x 144 columns]"
      ]
     },
     "execution_count": 10,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Loading the dataset with the linguistic features extracted with ProfilingUD\n",
    "df = pd.read_csv(\"../data/GUM_Treebank_Features.tsv\", sep='\\t')\n",
    "df = df.set_index(\"identifier\")\n",
    "df.head()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## **5. Train the probing model**\n",
    "\n",
    "Below is reported the function for training the probing model. The function takes as input the following arguments:\n",
    "* BERT internal representations;\n",
    "* The dataframe with the linguistic features extracted with Profiling-UD;\n",
    "* The ids of the sentences in the original dataset;\n",
    "* The BERT layer from which extract the representation as input features of the probing model;\n",
    "* The linguistic feature to be used as prediction label of the probing model.\n",
    "\n",
    "**Preprocessing**: Before starting the actual training process, it is necessary to obtain an aggregated representation of each sentence in our dataset using the BERT token-level representations. For instance, we can apply a mean pooling strategy to compute the average of all token representations in each sentence. In our case, to obtain the embeddings representations for our sentence-level tasks we used for each of its 12 layers the activation of the first input token ('*\\[CLS]*'), which is generally used as classification token and somehow summarizes the information from the actual tokens, as suggested in \"*What does BERT learn about the structure of language?*\" (Jawahar et al., 2019, https://aclanthology.org/P19-1356.pdf).\n",
    "\n",
    "**Train/test split**: Next, we split the dataset into training and test sets, using an 80% split for training and a 20% split for the test set (cross-fold validation could also be performed if desired).\n",
    "\n",
    "**Model definition and training**: Finally, we train the probing model. The chosen model for these experiments is a linear Support Vector Regressor (LinearSVR). We opt for a linear model to keep it as simple as possible and avoid interference with the linguistic properties implicitly encoded in the model representations. Since the features to predict are continuous, the model is set up as a regressor.\n",
    "\n",
    "**Evaluation**: Upon completion of the training, we can evaluate the predictions using commonly used regression metrics (e.g., mean squared error (MSE), R-squared (R2) or Spearman correlation). In our case, we use Spearman correlation, but other metrics can also be tested for performance evaluation."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [],
   "source": [
    "def probing(embeddings, df, ids, feature, layer):\n",
    "\n",
    "  # Dataset containing only the selected linguistic feature\n",
    "  df_feature = df[feature]\n",
    "\n",
    "  X = []\n",
    "  y = []\n",
    "  \n",
    "  for id, sentence in zip(ids, embeddings):\n",
    "    embedding_layer = sentence[layer][0]\n",
    "\n",
    "    # We extract the representation from the [CLS] token, that is the first token in each input sequence, having set 'add_special_tokens' equal to 'True' during tokenization\n",
    "    cls_embedding = embedding_layer[0].tolist()\n",
    "\n",
    "    # Access to the value of the linguistic feature 'feature' for the given sentence, based on the id\n",
    "    feat = df_feature.loc[id]\n",
    "\n",
    "    X.append(cls_embedding)\n",
    "    y.append(feat)\n",
    "\n",
    "  # Diving the dataset into train and test files\n",
    "  X_train, X_test, y_train, y_test, ids_train, ids_test = train_test_split(X, y, ids, test_size=0.20, random_state=42)\n",
    "\n",
    "  # Defining the probing model\n",
    "  probing_model = LinearSVR(dual=False, loss='squared_epsilon_insensitive')\n",
    "  \n",
    "  # Fit the model on the training data and then predict on X_test\n",
    "  probing_model.fit(X_train, y_train)\n",
    "  y_pred = probing_model.predict(X_test)\n",
    "\n",
    "  # Evaluation of the predictions\n",
    "  corr, p_val = spearmanr(y_test, y_pred)\n",
    "\n",
    "  # Save the predictions for further analysis\n",
    "  df_preds = pd.DataFrame(columns=[\"sent_id\", \n",
    "                                   \"y_true\", \n",
    "                                   \"y_pred\"])\n",
    "  df_preds[\"sent_id\"] = ids_test\n",
    "  df_preds[\"y_true\"] = y_test\n",
    "  df_preds[\"y_pred\"] = y_pred\n",
    "  df_preds.to_csv(f'../data/results_{layer}_{feature}.tsv', \n",
    "                  sep='\\t', index=True)\n",
    "  \n",
    "  return corr, p_val"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### **5.1. Baseline model**\n",
    "\n",
    "To ensure a meaningul comparison of the results, we define a baseline model (LinearSVR) that exclusively relies on the length of the input sentences to predict the set of linguistic features. This baseline serves as a reference point for evaluating the performance of the probing models trained with BERT internal representations."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [],
   "source": [
    "def baseline(df, feature):\n",
    "  # Select \"n_tokens\" as X for the baseline probing model\n",
    "  X = df[\"n_tokens\"].to_numpy()\n",
    "  X = X.reshape(-1, 1)\n",
    "  y = df[feature].to_numpy()\n",
    "\n",
    "  # Diving the dataset into train and test files\n",
    "  X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.20, random_state=42)\n",
    "  \n",
    "  # Defining the probing model\n",
    "  probing_model = LinearSVR(dual=False, loss='squared_epsilon_insensitive')\n",
    "  \n",
    "  # Fit the model on the training data and then predict on X_test\n",
    "  probing_model.fit(X_train, y_train)\n",
    "  y_pred = probing_model.predict(X_test)\n",
    "\n",
    "  # Evaluation of the predictions\n",
    "  corr, p_val = spearmanr(y_test, y_pred)\n",
    "  \n",
    "  return corr, p_val"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Once the two functions for training the probing model and the baseline have been defined, we can execute the probing tasks, one for each layer of the model and for each lingusitic feature extracted with Profiling-UD.\n",
    "\n",
    "Below is the code to perform the probing tasks on all 12 layers of BERT and on a set of 6 linguistic features extracted with Profiling-UD:\n",
    "* *n_tokens*: total number of tokens;\n",
    "* *char_per_tok*: average number of characters per word (excluded punctuation);\n",
    "* *upos_dist_NOUN*: distribution of nouns in the sentence;\n",
    "* *upos_dist_VERB*: distribution of verbs in the sentence;\n",
    "* *avg_token_per_clause*: average clause length, calculated in terms of the average number of tokens per clause, where a clause is defined as the ratio between the number of tokens in a sentence and the number of either verbal or copular head; \n",
    "* *avg_subordinate_chain_len*: average length of subordinate chains, where a subordinate 'chain' is calculated as the number of subordinate clauses embedded on a first subordinate clause.\n",
    "\n",
    "Feel free to customize the code according to your requirements, such as executing it selectively on specific layers, distinct groups of features, or even on all of them."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [],
   "source": [
    "selected_features = [\"n_tokens\", \n",
    "                     \"char_per_tok\", \n",
    "                     \"upos_dist_NOUN\", \n",
    "                     \"upos_dist_VERB\", \n",
    "                     \"avg_token_per_clause\", \n",
    "                     \"avg_subordinate_chain_len\"\n",
    "                     ]\n",
    "\n",
    "layers = list(range(n_layers)[1:])\n",
    "\n",
    "# Create an empty pandas DataFrame to store the probing results\n",
    "results = pd.DataFrame(index=selected_features, columns=layers + [\"Baseline\"])\n",
    "\n",
    "# Iterate over the selected features\n",
    "for feature in selected_features:\n",
    "  # Iterate over the 12 layers of the model\n",
    "  for layer in layers:\n",
    "    corr, p_val = probing(embeddings, df, ids, feature, layer)\n",
    "\n",
    "    # Save the score only if the correlation is statistically significant (p-value < 0.05)\n",
    "    if p_val < 0.05:\n",
    "      results.loc[feature][layer] = corr\n",
    "\n",
    "  # Simple baseline model trained to predict the selected linguistic feature using only \"n_tokens\" as input feature\n",
    "  corr_baseline, p_val_baseline = baseline(df, feature)\n",
    "  if p_val < 0.05:\n",
    "    results.loc[feature][\"Baseline\"] = corr_baseline"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## **6. Results visualization**\n",
    "\n",
    "We can now look at the results that we previously stored in a DataFrame and compare with respect to the '*n_tokens*' baseline.\n",
    "\n",
    "Since we have performed the probing tasks using representations extracted from all 12 layers of the model, we can also explore the progression across the layers, from the input (1) to the output one (12)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>1</th>\n",
       "      <th>2</th>\n",
       "      <th>3</th>\n",
       "      <th>4</th>\n",
       "      <th>5</th>\n",
       "      <th>6</th>\n",
       "      <th>7</th>\n",
       "      <th>8</th>\n",
       "      <th>9</th>\n",
       "      <th>10</th>\n",
       "      <th>11</th>\n",
       "      <th>12</th>\n",
       "      <th>Baseline</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>n_tokens</th>\n",
       "      <td>0.990812</td>\n",
       "      <td>0.989576</td>\n",
       "      <td>0.988145</td>\n",
       "      <td>0.982168</td>\n",
       "      <td>0.971336</td>\n",
       "      <td>0.961379</td>\n",
       "      <td>0.945926</td>\n",
       "      <td>0.937147</td>\n",
       "      <td>0.939375</td>\n",
       "      <td>0.934813</td>\n",
       "      <td>0.925843</td>\n",
       "      <td>0.919626</td>\n",
       "      <td>1.0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>char_per_tok</th>\n",
       "      <td>0.739853</td>\n",
       "      <td>0.748602</td>\n",
       "      <td>0.746031</td>\n",
       "      <td>0.729927</td>\n",
       "      <td>0.66928</td>\n",
       "      <td>0.640196</td>\n",
       "      <td>0.583172</td>\n",
       "      <td>0.582061</td>\n",
       "      <td>0.550919</td>\n",
       "      <td>0.564811</td>\n",
       "      <td>0.546866</td>\n",
       "      <td>0.558774</td>\n",
       "      <td>-0.08138</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>upos_dist_NOUN</th>\n",
       "      <td>0.697179</td>\n",
       "      <td>0.71566</td>\n",
       "      <td>0.74901</td>\n",
       "      <td>0.766281</td>\n",
       "      <td>0.756462</td>\n",
       "      <td>0.770728</td>\n",
       "      <td>0.748331</td>\n",
       "      <td>0.713599</td>\n",
       "      <td>0.732016</td>\n",
       "      <td>0.722192</td>\n",
       "      <td>0.68072</td>\n",
       "      <td>0.645216</td>\n",
       "      <td>-0.038892</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>upos_dist_VERB</th>\n",
       "      <td>0.642125</td>\n",
       "      <td>0.668414</td>\n",
       "      <td>0.724502</td>\n",
       "      <td>0.749218</td>\n",
       "      <td>0.737134</td>\n",
       "      <td>0.741821</td>\n",
       "      <td>0.720013</td>\n",
       "      <td>0.704305</td>\n",
       "      <td>0.7032</td>\n",
       "      <td>0.674628</td>\n",
       "      <td>0.681722</td>\n",
       "      <td>0.550223</td>\n",
       "      <td>-0.045091</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>avg_token_per_clause</th>\n",
       "      <td>0.724917</td>\n",
       "      <td>0.731751</td>\n",
       "      <td>0.732388</td>\n",
       "      <td>0.719989</td>\n",
       "      <td>0.71966</td>\n",
       "      <td>0.712989</td>\n",
       "      <td>0.726106</td>\n",
       "      <td>0.686361</td>\n",
       "      <td>0.71692</td>\n",
       "      <td>0.665632</td>\n",
       "      <td>0.650959</td>\n",
       "      <td>0.650067</td>\n",
       "      <td>0.553152</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>avg_subordinate_chain_len</th>\n",
       "      <td>0.743092</td>\n",
       "      <td>0.760478</td>\n",
       "      <td>0.770249</td>\n",
       "      <td>0.756293</td>\n",
       "      <td>0.742003</td>\n",
       "      <td>0.740605</td>\n",
       "      <td>0.733927</td>\n",
       "      <td>0.713321</td>\n",
       "      <td>0.718466</td>\n",
       "      <td>0.701982</td>\n",
       "      <td>0.673835</td>\n",
       "      <td>0.655855</td>\n",
       "      <td>0.602162</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                                  1         2         3         4         5  \\\n",
       "n_tokens                   0.990812  0.989576  0.988145  0.982168  0.971336   \n",
       "char_per_tok               0.739853  0.748602  0.746031  0.729927   0.66928   \n",
       "upos_dist_NOUN             0.697179   0.71566   0.74901  0.766281  0.756462   \n",
       "upos_dist_VERB             0.642125  0.668414  0.724502  0.749218  0.737134   \n",
       "avg_token_per_clause       0.724917  0.731751  0.732388  0.719989   0.71966   \n",
       "avg_subordinate_chain_len  0.743092  0.760478  0.770249  0.756293  0.742003   \n",
       "\n",
       "                                  6         7         8         9        10  \\\n",
       "n_tokens                   0.961379  0.945926  0.937147  0.939375  0.934813   \n",
       "char_per_tok               0.640196  0.583172  0.582061  0.550919  0.564811   \n",
       "upos_dist_NOUN             0.770728  0.748331  0.713599  0.732016  0.722192   \n",
       "upos_dist_VERB             0.741821  0.720013  0.704305    0.7032  0.674628   \n",
       "avg_token_per_clause       0.712989  0.726106  0.686361   0.71692  0.665632   \n",
       "avg_subordinate_chain_len  0.740605  0.733927  0.713321  0.718466  0.701982   \n",
       "\n",
       "                                 11        12  Baseline  \n",
       "n_tokens                   0.925843  0.919626       1.0  \n",
       "char_per_tok               0.546866  0.558774  -0.08138  \n",
       "upos_dist_NOUN              0.68072  0.645216 -0.038892  \n",
       "upos_dist_VERB             0.681722  0.550223 -0.045091  \n",
       "avg_token_per_clause       0.650959  0.650067  0.553152  \n",
       "avg_subordinate_chain_len  0.673835  0.655855  0.602162  "
      ]
     },
     "execution_count": 14,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Visualize the results memorized in the DataFrame\n",
    "results"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<matplotlib.legend.Legend at 0x7f885637e020>"
      ]
     },
     "execution_count": 15,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 800x600 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# Formatting the dataframe in order to be visualized as a multiline plot\n",
    "results = results[layers]\n",
    "results = results.T\n",
    "results[\"Layers\"] = results.index\n",
    "plot_results = pd.melt(results, [\"Layers\"])\n",
    "\n",
    "# Visualization with seaborn\n",
    "fig, ax = plt.subplots(figsize=(8, 6))\n",
    "g = sns.lineplot(data=plot_results, x=\"Layers\", y=\"value\", hue=\"variable\")\n",
    "handles, labels = g.get_legend_handles_labels()\n",
    "g.legend(handles=handles, labels=labels, bbox_to_anchor=(0., 1.02, 1., .102),\n",
    "              loc='lower left', ncol=3, mode=\"expand\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3 (ipykernel)",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.10.6"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}