{
 "cells": [
  {
   "cell_type": "markdown",
   "id": "8ea54fcd-ef4a-42cb-ae26-cbdc6f6ffc64",
   "metadata": {
    "tags": []
   },
   "source": [
    "# Duct Tape Pipeline\n",
    "To explore how users may interact with interactive visualizations of counterfactuals for evolving the Interactive Model Card, we will need to first find a way to generate counterfactuals based on a given input. We want the user to be able to provide their input and direct the system to generate counterfactuals based on a part of speech that is significant to the model. The system should then provide a data frame of counterfactuals to be used in an interactive visualization. Below is an example wireframe of the experience based on previous research.\n",
    "\n",
    "![wireframe](Assets/VizNLC-Wireframe-example.png)\n",
    "\n",
    "## Goals of this notebook\n",
    "* Test which libraries (Ex. [spaCy](https://spacy.io/) and [NLTK](https://www.nltk.org/)) will work\n",
    "* Identify defaults to use\n",
    "* Build a rudimentary script for generating counterfactuals from user input\n",
    "* Ensure the counterfactuals are in a useable format for visualization"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "736e6375-dd6d-4188-b8b1-92bded2bcd02",
   "metadata": {},
   "source": [
    "## Loading the libraries and models"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "id": "7f581785-e642-4f74-9f67-06a63820eaf2",
   "metadata": {},
   "outputs": [],
   "source": [
    "#Import the libraries we know we'll need for the Generator.\n",
    "import pandas as pd, spacy, nltk, numpy as np\n",
    "from spacy import displacy\n",
    "from spacy.matcher import Matcher\n",
    "#!python -m spacy download en_core_web_sm\n",
    "nlp = spacy.load(\"en_core_web_sm\")\n",
    "lemmatizer = nlp.get_pipe(\"lemmatizer\")\n",
    "\n",
    "#Import the libraries to support the model, predictions, and LIME.\n",
    "from transformers import AutoTokenizer, AutoModelForSequenceClassification, TextClassificationPipeline\n",
    "import lime\n",
    "import torch\n",
    "import torch.nn.functional as F\n",
    "from lime.lime_text import LimeTextExplainer\n",
    "\n",
    "#Import the libraries for generating interactive visualizations.\n",
    "import altair as alt"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "cbe2b292-e33e-4915-8e61-bba5327fb643",
   "metadata": {},
   "outputs": [],
   "source": [
    "#Defining all necessary variables and instances.\n",
    "tokenizer = AutoTokenizer.from_pretrained(\"distilbert-base-uncased-finetuned-sst-2-english\")\n",
    "model = AutoModelForSequenceClassification.from_pretrained(\"distilbert-base-uncased-finetuned-sst-2-english\")\n",
    "class_names = ['negative', 'positive']\n",
    "explainer = LimeTextExplainer(class_names=class_names)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "197c3e26-0fdf-49c6-9135-57f1fd55d3e3",
   "metadata": {},
   "outputs": [],
   "source": [
    "#Defining a Predictor required for LIME to function.\n",
    "def predictor(texts):\n",
    "    outputs = model(**tokenizer(texts, return_tensors=\"pt\", padding=True))\n",
    "    probas = F.softmax(outputs.logits, dim=1).detach().numpy()\n",
    "    return probas"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "e731dcbb-4fcf-41c6-9493-edef02fdb1b6",
   "metadata": {},
   "source": [
    "## Exploring concepts to see what might work\n",
    "To begin building the pipeline I started by identifying whether or not I needed to build my own matcher or if spaCy has something built in that would allow us to make it easier. Having to build our own matcher, to account for each of the possible patterns, would be exceptionally cumbersome with all of the variations we need to look out for. Instead, I found that using the built in `noun_chunks` attribute allows for a simplification to the parts of speech we most care about. \n",
    "* I built a few helper functions from tutorials to explore the parts-of-speech within given sentences and the way `noun_chunks` work\n",
    "* I explore dusing `displacy` as a means of visualizing sentences to call out what the pre-trained models already understand"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "1f2eca3c-525c-4e29-8cc1-c87e89a3fadf",
   "metadata": {},
   "outputs": [],
   "source": [
    "#A quick test of Noun Chunks\n",
    "text = \"The movie was filmed in New Zealand.\"\n",
    "doc = nlp(text)\n",
    "def n_chunk(doc):\n",
    "    for chunk in doc.noun_chunks:\n",
    "        print(f\"Text: {chunk.text:<12}| Root:{chunk.root.text:<12}| Root Dependency: {chunk.root.dep_:<12}| Root Head: {chunk.root.head.text:<12}\")\n",
    "n_chunk(doc)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "98978c29-a39c-48e3-bdbb-b74388ded6bc",
   "metadata": {},
   "outputs": [],
   "source": [
    "#The user will need to enter text. For now, we're going to provide a series of sentences generated to have things we care about. For clarity \"upt\" means \"user provide text\".\n",
    "upt1 = \"I like movies starring black actors.\"\n",
    "upt2 = \"I am a black trans-woman.\"\n",
    "upt3 = \"Native Americans deserve to have their land back.\"\n",
    "upt4 = \"This movie was filmed in Iraq.\"\n",
    "\n",
    "#Here I provide a larger text with mixed messages one sentence per line.\n",
    "text1 = (\n",
    "\"I like movies starring black actors.\"\n",
    "\"I am a black trans-woman.\"\n",
    "\"Native Americans deserve to have their land back.\"\n",
    "\"This movie was filmed in Iraq.\"\n",
    "\"The Chinese cat and the African bat walked into a Jamaican bar.\"\n",
    "\"There once was a flexible pole that met an imovable object.\"\n",
    "\"A Catholic nun, a Buddhist monk, a satanic cultist, and a Wiccan walk into your garage.\")\n",
    "\n",
    "doc1 = nlp(upt1)\n",
    "doc2 = nlp(upt2)\n",
    "doc3 = nlp(upt3)\n",
    "doc4 = nlp(upt4)\n",
    "doct = nlp(text1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "38023eca-b224-412d-aa71-02bd694530e0",
   "metadata": {},
   "outputs": [],
   "source": [
    "#Using displacy to explore how the NLP model views sentences.\n",
    "displacy.render(doc, style=\"ent\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "c28edec8-dc30-4ef9-8c1e-131b0e1b1a45",
   "metadata": {},
   "outputs": [],
   "source": [
    "#Another visual for understanding how the model views sentences.\n",
    "displacy.render(doc, style=\"dep\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "id": "dd0d5f8e-ee80-48f7-be92-effa5f84c723",
   "metadata": {},
   "outputs": [],
   "source": [
    "#A simple token to print out the \n",
    "def text_pos(doc):\n",
    "    for token in doc:\n",
    "        # Get the token text, part-of-speech tag and dependency label\n",
    "        token_text = token.text\n",
    "        token_pos = token.pos_\n",
    "        token_dep = token.dep_\n",
    "        token_ent = token.ent_type_\n",
    "        token_morph = token.morph\n",
    "        # This is for formatting only\n",
    "        print(f\"Text: {token_text:<12}| Part of Speech: {token_pos:<10}| Dependency: {token_dep:<10}| Entity: {token_ent:<10} | Morph: {token_morph}\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "id": "5dfee095-3852-4dba-a7dc-5519e8ec6eaa",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Text: Who         | Part of Speech: PRON      | Dependency: nsubj     | Entity:            | Morph: \n",
      "Text: put         | Part of Speech: VERB      | Dependency: ROOT      | Entity:            | Morph: Tense=Past|VerbForm=Fin\n",
      "Text: a           | Part of Speech: DET       | Dependency: det       | Entity:            | Morph: Definite=Ind|PronType=Art\n",
      "Text: tiny        | Part of Speech: ADJ       | Dependency: amod      | Entity:            | Morph: Degree=Pos\n",
      "Text: pickle      | Part of Speech: NOUN      | Dependency: dobj      | Entity:            | Morph: Number=Sing\n",
      "Text: in          | Part of Speech: ADP       | Dependency: prep      | Entity:            | Morph: \n",
      "Text: the         | Part of Speech: DET       | Dependency: det       | Entity:            | Morph: Definite=Def|PronType=Art\n",
      "Text: jar         | Part of Speech: NOUN      | Dependency: pobj      | Entity:            | Morph: Number=Sing\n"
     ]
    }
   ],
   "source": [
    "x = nlp(\"Who put a tiny pickle in the jar\")\n",
    "text_pos(x)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "id": "2485d88d-2dd4-4fa3-9d62-4dcbec4e9138",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0"
      ]
     },
     "execution_count": 11,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "len(x[0].morph)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "013af6ac-f7d1-41d2-a601-b0f9a4870815",
   "metadata": {},
   "outputs": [],
   "source": [
    "#Instantiate a matcher and use it to test some patterns.\n",
    "matcher = Matcher(nlp.vocab)\n",
    "pattern = [{\"ENT_TYPE\": {\"IN\":[\"NORP\",\"GPE\"]}}]\n",
    "matcher.add(\"proper_noun\", [pattern])\n",
    "pattern_test = [{\"DEP\": \"amod\"},{\"DEP\":\"attr\"},{\"TEXT\":\"-\"},{\"DEP\":\"attr\",\"OP\":\"+\"}]\n",
    "matcher.add(\"amod_attr\",[pattern_test])\n",
    "pattern_an = [{\"DEP\": \"amod\"},{\"POS\":{\"IN\":[\"NOUN\",\"PROPN\"]}},{\"DEP\":{\"NOT_IN\":[\"attr\"]}}]\n",
    "matcher.add(\"amod_noun\", [pattern_an])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "f6ac821d-7b56-446e-b9ca-42a5f5afd198",
   "metadata": {},
   "outputs": [],
   "source": [
    "def match_this(matcher, doc):\n",
    "    matches = matcher(doc)\n",
    "    for match_id, start, end in matches:\n",
    "        matched_span = doc[start:end]\n",
    "        print(f\"Mached {matched_span.text} by the rule {nlp.vocab.strings[match_id]}.\")\n",
    "    return matches"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "958e4dc8-6652-4f32-b7ae-6aa5ee287cf7",
   "metadata": {},
   "outputs": [],
   "source": [
    "match_this(matcher, doct)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "5bf40fa5-b636-47f7-98b2-e872c78e7114",
   "metadata": {},
   "outputs": [],
   "source": [
    "text_pos(doc3)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "c5365304-5edb-428d-abf5-d579dcfbc269",
   "metadata": {},
   "outputs": [],
   "source": [
    "n_chunk(doc3)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "b7f3d3c8-65a1-433f-a47c-adcaaa2353e2",
   "metadata": {},
   "outputs": [],
   "source": [
    "displacy.render(doct, style=\"ent\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "84df8e30-d142-4e5b-b3a9-02e3133ceba9",
   "metadata": {},
   "outputs": [],
   "source": [
    "txt = \"Savannah is a city in Georgia, in the United States\"\n",
    "doc = nlp(txt)\n",
    "displacy.render(doc, style=\"ent\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "4a85f713-92bc-48ba-851e-de627d7e8c77",
   "metadata": {},
   "outputs": [],
   "source": [
    "displacy.render(doc2, style='dep')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "032f1134-7560-400b-824b-bc0196058b66",
   "metadata": {},
   "outputs": [],
   "source": [
    "n_chunk(doct)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "188044a1-4cf4-4141-a520-c5f11198aed8",
   "metadata": {},
   "source": [
    "* The Model does not recognize `wiccan` as a NORP but it will recognize `Wiccan` as NORP\n",
    "* The Model does not know what to do with `-` and makes a mess of `trans-woman` because of this"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "2dc82250-e26e-49d5-a7f2-d4eeda170e4e",
   "metadata": {},
   "outputs": [],
   "source": [
    "chunks = list(doc1.noun_chunks)\n",
    "print(chunks[-1][-2].pos_)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "c23d48c4-f5ab-4428-9244-0786e9903a8e",
   "metadata": {},
   "source": [
    "## Building the Duct-Tape Pipeline cell-by-cell"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "7ed22421-4401-482e-b54a-ee70d3187037",
   "metadata": {},
   "outputs": [],
   "source": [
    "#Lists of important words\n",
    "gender = [\"man\", \"woman\",\"girl\",\"boy\",\"male\",\"female\",\"husband\",\"wife\",\"girlfriend\",\"boyfriend\",\"brother\",\"sister\",\"aunt\",\"uncle\",\"grandma\",\"grandpa\",\"granny\",\"granps\",\"grandmother\",\"grandfather\",\"mama\",\"dada\",\"Ma\",\"Pa\",\"lady\",\"gentleman\"]\n",
    "#consider pulling ethnicities from https://github.com/cgio/global-ethnicities"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "8b02a5d4-8a6b-4e5e-8f15-4f9182fe341f",
   "metadata": {},
   "outputs": [],
   "source": [
    "def select_crit(document, options=False, limelist=False):\n",
    "    '''This function is meant to select the critical part of a sentence. Critical, in this context means\n",
    "    the part of the sentence that is either: A) a PROPN from the correct entity group; B) an ADJ associated with a NOUN;\n",
    "    C) a NOUN that represents gender. It also checks this against what the model thinks is important if the user defines \"options\" as \"LIME\" or True.'''\n",
    "    chunks = list(document.noun_chunks)\n",
    "    pos_options = []\n",
    "    lime_options = []\n",
    "    \n",
    "    #Identify what the model cares about.\n",
    "    if options:\n",
    "        exp = explainer.explain_instance(document.text, predictor, num_features=20, num_samples=2000)\n",
    "        results = exp.as_list()[:10]\n",
    "        #prints the results from lime for QA.\n",
    "        if limelist == True:\n",
    "            print(results)\n",
    "        for feature in results:\n",
    "            lime_options.append(feature[0])\n",
    "    \n",
    "    #Identify what we care about \"parts of speech\"\n",
    "    for chunk in chunks:\n",
    "        #The use of chunk[-1] is due to testing that it appears to always match the root\n",
    "        root = chunk[-1]\n",
    "        #This currently matches to a list I've created. I don't know the best way to deal with this so I'm leaving it as is for the moment.\n",
    "        if root.text.lower() in gender:\n",
    "            cur_values = [token.text for token in chunk if token.pos_ in [\"NOUN\",\"ADJ\"]]\n",
    "            if (all(elem in lime_options for elem in cur_values) and ((options == \"LIME\") or (options == True))) or ((options != \"LIME\") and (options != True)):\n",
    "                pos_options.extend(cur_values)\n",
    "                #print(f\"From {chunk.text}, {cur_values} added to pos_options due to gender.\") #for QA\n",
    "        #This is currently set to pick up entities in a particular set of groups (which I recently expanded). Should it just pick up all named entities?\n",
    "        elif root.ent_type_ in [\"GPE\",\"NORP\",\"DATE\",\"EVENT\"]:\n",
    "            cur_values = []\n",
    "            if (len(chunk) > 1) and (chunk[-2].dep_ == \"compound\"):\n",
    "                #creates the compound element of the noun\n",
    "                compound = [x.text for x in chunk if x.dep_ == \"compound\"]\n",
    "                print(f\"This is the contents of {compound} and it is {all(elem in lime_options for elem in compound)} that all elements are present in {lime_options}.\") #for QA\n",
    "                #checks to see all elements in the compound are important to the model or use the compound if not checking importance.\n",
    "                if (all(elem in lime_options for elem in compound) and ((options == \"LIME\") or (options == True))) or ((options != \"LIME\") and (options != True)):\n",
    "                    #creates a span for the entirety of the compound noun and adds it to the list.\n",
    "                    span = -1 * (1 + len(compound))\n",
    "                    pos_options.append(chunk[span:].text)\n",
    "                    cur_values + [token.text for token in chunk if token.pos_ == \"ADJ\"]\n",
    "            else: \n",
    "                cur_values = [token.text for token in chunk if (token.ent_type_ in [\"GPE\",\"NORP\",\"DATE\",\"EVENT\"]) or (token.pos_ == \"ADJ\")]\n",
    "            if (all(elem in lime_options for elem in cur_values) and ((options == \"LIME\") or (options == True))) or ((options != \"LIME\") and (options != True)):\n",
    "                pos_options.extend(cur_values)\n",
    "                print(f\"From {chunk.text}, {cur_values} and {pos_options} added to pos_options due to entity recognition.\") #for QA\n",
    "        elif len(chunk) > 1:\n",
    "            cur_values = [token.text for token in chunk if token.pos_ in [\"NOUN\",\"ADJ\"]]\n",
    "            if (all(elem in lime_options for elem in cur_values) and ((options == \"LIME\") or (options == True))) or ((options != \"LIME\") and (options != True)):\n",
    "                pos_options.extend(cur_values)\n",
    "                print(f\"From {chunk.text}, {cur_values} added to pos_options due to wildcard.\") #for QA\n",
    "        else:\n",
    "            print(f\"No options added for \\'{chunk.text}\\' \")\n",
    "    \n",
    "    #Return the correct set of options based on user input, defaults to POS for simplicity.\n",
    "    if options == \"LIME\":\n",
    "        return lime_options\n",
    "    else:\n",
    "        return pos_options"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "fa95e9fe-36ea-4b95-ab51-6bb82f745c23",
   "metadata": {},
   "outputs": [],
   "source": [
    "#Testing a method to make sure I had the ability to match one list inside the other. Now incorporated in the above function's logic.\n",
    "one = ['a','b','c']\n",
    "two = ['a','c']\n",
    "all(elem in one for elem in two)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "d43e202e-64b9-4cea-b117-82492c9ee5f4",
   "metadata": {},
   "outputs": [],
   "source": [
    "#Test to make sure all three options work\n",
    "pos4 = select_crit(doc4)\n",
    "lime4 = select_crit(doc4,options=\"LIME\")\n",
    "final4 = select_crit(doc4,options=True,limelist=True)\n",
    "print(pos4, lime4, final4)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "5623015e-fdb2-44f0-b5ac-812203b639b3",
   "metadata": {},
   "outputs": [],
   "source": [
    "#This is a test to make sure compounds of any length are captured. \n",
    "txt = \"I went to Papua New Guinea for Christmas Eve and New Years.\"\n",
    "doc_t = nlp(txt)\n",
    "select_crit(doc_t)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "58be22eb-a5c3-4a01-820b-45d190fce52d",
   "metadata": {},
   "outputs": [],
   "source": [
    "#Test to make sure all three options work. A known issue is that if we combine the compounds then they will not end up in the final_options...\n",
    "pos_t = select_crit(doc_t)\n",
    "lime_t = select_crit(doc_t,options=\"LIME\")\n",
    "final_t = select_crit(doc_t,options=True,limelist=True)\n",
    "print(pos_t, lime_t, final_t)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "1158de94-1472-4001-b3a1-42a488bcb20f",
   "metadata": {},
   "outputs": [],
   "source": [
    "select_crit(doc_t,options=True)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "05063ede-422f-4536-8408-ceb5441adbe8",
   "metadata": {},
   "source": [
    "> Note `Papua` and `Eve` have such low impact on the model that they do not always appear... so there will always be limitations to matching."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "2c7c1ca9-4962-4fbe-b18b-1e20a223aff9",
   "metadata": {},
   "outputs": [],
   "source": [
    "select_crit(doc_t,options=\"LIME\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "c70387a5-c431-43a5-a3b8-7533268a94e3",
   "metadata": {},
   "outputs": [],
   "source": [
    "displacy.render(doc_t, style=\"ent\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "4b92d276-7d67-4c1c-940b-d3b2dcc756b9",
   "metadata": {},
   "outputs": [],
   "source": [
    "#This run clearly indicates that this pipeline from spaCy does not know what to do with hyphens(\"-\") and that we need to be aware of that.\n",
    "choices = select_crit(doct)\n",
    "choices"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "ea6b29d0-d0fa-4eb3-af9c-970759124145",
   "metadata": {},
   "outputs": [],
   "source": [
    "user_choice = choices[2]\n",
    "matcher2 = Matcher(nlp.vocab)\n",
    "pattern = [{\"TEXT\": user_choice}]\n",
    "matcher2.add(\"user choice\", [pattern])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "d32754b8-f1fa-4781-a6b0-829ad7ec2e50",
   "metadata": {},
   "outputs": [],
   "source": [
    "#consider using https://github.com/writerai/replaCy instead\n",
    "match_id, start, end = match_this(matcher2,doc2)[0]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "a0362734-020b-49ad-b566-fdc7196e705c",
   "metadata": {},
   "outputs": [],
   "source": [
    "docx = doc2.text.replace(user_choice,\"man\")\n",
    "docx"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "bf0512b6-336e-4842-9bde-34e03a1ca7c6",
   "metadata": {},
   "source": [
    "### Testing predictions and visualization\n",
    "Here I will attempt to import the model from huggingface, generate predictions for each of the sentences, and then visualize those predictions into a dot plot. If I can get this to work then I will move on to testing a full pipeline for letting the user pick which part of the sentence they wish to generate counterfactuals for."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "e0bd4134-3b22-4ae8-870c-3a66c1cf8b23",
   "metadata": {},
   "outputs": [],
   "source": [
    "#Testing to see how to get predictions from the model. Ultimately, this did not work.\n",
    "token = tokenizer(upt4, return_tensors=\"pt\")\n",
    "labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1\n",
    "outputs = model(**token, labels=labels)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "74c639bb-e74a-4a46-8047-3552265ae6a4",
   "metadata": {},
   "outputs": [],
   "source": [
    "#Discovering that there's a pipeline specifically to provide scores. \n",
    "#I used it to get a list of lists of dictionaries that I can then manipulate to calculate the proper prediction score.\n",
    "pipe = TextClassificationPipeline(model=model, tokenizer=tokenizer, return_all_scores=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "8e1ff15d-0fb9-475b-bd24-4548c0782343",
   "metadata": {},
   "outputs": [],
   "source": [
    "preds = pipe(upt4)\n",
    "print(preds[0][0])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "d8abb9ca-36cf-441a-9236-1f7e44331b53",
   "metadata": {},
   "outputs": [],
   "source": [
    "score_1 = preds[0][0]['score']\n",
    "score_2 = (score_1 - .5) * 2\n",
    "print(score_1, score_2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "8726a284-99bd-47f1-9756-1c3ae603db10",
   "metadata": {},
   "outputs": [],
   "source": [
    "def eval_pred(text):\n",
    "    '''A basic function for evaluating the prediction from the model and turning it into a visualization friendly number.'''\n",
    "    preds = pipe(text)\n",
    "    neg_score = preds[0][0]['score']\n",
    "    pos_score = preds[0][1]['score']\n",
    "    if pos_score >= neg_score:\n",
    "        return pos_score\n",
    "    if neg_score >= pos_score:\n",
    "        return -1 * neg_score"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "f38f5061-f30a-4c81-9465-37951c3ad9f4",
   "metadata": {},
   "outputs": [],
   "source": [
    "def eval_pred_test(text, return_all = False):\n",
    "    '''A basic function for evaluating the prediction from the model and turning it into a visualization friendly number.'''\n",
    "    preds = pipe(text)\n",
    "    neg_score = -1 * preds[0][0]['score']\n",
    "    sent_neg = preds[0][0]['label']\n",
    "    pos_score = preds[0][1]['score']\n",
    "    sent_pos = preds[0][1]['label']\n",
    "    prediction = 0\n",
    "    sentiment = ''\n",
    "    if pos_score > abs(neg_score):\n",
    "        prediction = pos_score\n",
    "        sentiment = sent_pos\n",
    "    elif abs(neg_score) > pos_score:\n",
    "        prediction = neg_score\n",
    "        sentiment = sent_neg\n",
    "        \n",
    "    if return_all:\n",
    "        return prediction, sentiment\n",
    "    else:\n",
    "        return prediction"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "abd5dd8c-8cff-4865-abf1-f5a744f2203b",
   "metadata": {},
   "outputs": [],
   "source": [
    "score = eval_pred(upt4)\n",
    "og_data = {'Country': ['Iraq'], 'Continent': ['Asia'], 'text':[upt4], 'pred':[score]}\n",
    "og_df = pd.DataFrame(og_data)\n",
    "og_df"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8b349a87-fe83-4045-a63a-d054489bb461",
   "metadata": {},
   "source": [
    "## Load the dummy countries I created to test generating counterfactuals\n",
    "I decided to test the pipeline with a known problem space. Taking the text from Aurélien Géron's observations in twitter, I built a built a small scale test using the learnings I had to prove that we can identify a particular part of speech, use it to generate counterfactuals, and then build a visualization off it."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "46ab3332-964c-449f-8cef-a9ff7df397a4",
   "metadata": {},
   "outputs": [],
   "source": [
    "#load my test data from https://github.com/dbouquin/IS_608/blob/master/NanosatDB_munging/Countries-Continents.csv\n",
    "df = pd.read_csv(\"Assets/Countries/countries.csv\")\n",
    "df.head()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "51c75894-80af-4625-8ce8-660e500b496b",
   "metadata": {},
   "outputs": [],
   "source": [
    "#Note: we will need to build the function that lets the user choose from the options available. For now I have hard coded it as \"selection\", from \"user_options\".\n",
    "user_options = select_crit(doc4)\n",
    "print(user_options)\n",
    "selection = user_options[1]\n",
    "selection"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "3d6419f1-bf7d-44bc-afb8-ac26ef9002df",
   "metadata": {},
   "outputs": [],
   "source": [
    "#Create a function that generates the counterfactuals within a data frame.\n",
    "def gen_cf_country(df,document,selection):\n",
    "    df['text'] = df.Country.apply(lambda x: document.text.replace(selection,x))\n",
    "    df['prediction'] = df.text.apply(eval_pred_test)\n",
    "    #added this because I think it will make the end results better if we ensure the seed is in the data we generate counterfactuals from.\n",
    "    df['seed'] = df.Country.apply(lambda x: 'seed' if x == selection else 'alternative')\n",
    "    return df\n",
    "\n",
    "df = gen_cf_country(df,doc4,selection)\n",
    "df.head()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "aec241a6-48c3-48c6-9e7f-d22612eaedff",
   "metadata": {},
   "outputs": [],
   "source": [
    "#Display Counterfactuals and Original in a layered chart. I couldn't get this to provide a legend.\n",
    "og = alt.Chart(og_df).encode(\n",
    "    x='Continent:N',\n",
    "    y='pred:Q'\n",
    ").mark_square(color='green', size = 200, opacity=.5)\n",
    "\n",
    "cf = alt.Chart(df).encode(\n",
    "    x='Continent:N',  # specify nominal data\n",
    "    y='prediction:Q',  # specify quantitative data\n",
    ").mark_circle(color='blue', size=50, opacity =.25)\n",
    "\n",
    "alt_plot = alt.LayerChart(layer=[cf,og], width = 300)\n",
    "alt_plot"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "ecb9dd41-2fab-49bd-bae5-30300ce39e41",
   "metadata": {},
   "outputs": [],
   "source": [
    "single_nearest = alt.selection_single(on='mouseover', nearest=True)\n",
    "full = alt.Chart(df).encode(\n",
    "    alt.X('Continent:N'),  # specify nominal data\n",
    "    alt.Y('prediction:Q'),  # specify quantitative data\n",
    "    color=alt.Color('seed:N', legend=alt.Legend(title=\"Seed or Alternative\")),\n",
    "    size=alt.Size('seed:N', alt.scale(domain=[50,100])),\n",
    "    tooltip=('Country','prediction')\n",
    ").mark_circle(opacity=.5).properties(width=300).add_selection(single_nearest)\n",
    "\n",
    "full"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "56bc30d7-03a5-43ff-9dfe-878197628305",
   "metadata": {},
   "outputs": [],
   "source": [
    "df2 = df.nlargest(5, 'prediction')\n",
    "df3 = df.nsmallest(5, 'prediction')\n",
    "frames = [df2,df3]\n",
    "results = pd.concat(frames)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "1610bb48-c9b9-4bee-bcb5-999886acb9e3",
   "metadata": {},
   "outputs": [],
   "source": [
    "bar = alt.Chart(results).encode(  \n",
    "    alt.X('prediction:Q'), \n",
    "    alt.Y('Country:N', sort=\"-x\"),\n",
    "    color=alt.Color('seed:N', legend=alt.Legend(title=\"Seed or Alternative\")),\n",
    "    size='seed:N',\n",
    "    tooltip=('Country','prediction')\n",
    ").mark_circle().properties(width=300).add_selection(single_nearest)\n",
    "\n",
    "bar"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "84c40b74-95be-4c19-bd57-74e6004b950c",
   "metadata": {},
   "source": [
    "#### QA"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "7d15c7d8-9fdb-4c5b-84fa-599839cbceac",
   "metadata": {},
   "outputs": [],
   "source": [
    "qa_txt = \"They serve halal food in Iraq and Egypt.\"\n",
    "qa_doc = nlp(qa_txt)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "d6956ddf-9287-419a-bb08-a3618f77700a",
   "metadata": {},
   "outputs": [],
   "source": [
    "displacy.render(qa_doc, style=\"dep\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "88768d68-fe44-49ab-ac12-d41e6716b3b3",
   "metadata": {},
   "outputs": [],
   "source": [
    "select_crit(qa_doc)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "7bbc6c2e-df5d-4076-8532-8648fd818be4",
   "metadata": {},
   "source": [
    "# NLC-Gen\n",
    "### A Natural Language Counterfactual Generator for Exploring Bias in Sentiment Analysis Algorithms\n",
    "\n",
    "##### Overview\n",
    "This project is an extension of [Interactive Model Cards](https://github.com/amcrisan/interactive-model-cards). It focuses on providing a person more ways to explore the bias of a model through the generation of alternatives (technically [counterfactuals](https://plato.stanford.edu/entries/counterfactuals/#WhatCoun)). We believe the use of alternatives people can better understand the limitations of a model and develop productive skepticism around its usage and trustworthiness.\n",
    "\n",
    "##### Set up\n",
    "\n",
    "Download the files from Github then perform the commands below in \n",
    "```sh\n",
    "cd NLC-Gen\n",
    "pipenv install\n",
    "pipenv shell\n",
    "python -m spacy download en_core_web_lg\n",
    "streamlit run NLC-app.py\n",
    "```\n",
    "\n",
    "##### Known Limitations\n",
    "* Words not in the spaCy vocab for `en_core_web_lg` won't have vectors and so won't have the ability to create similarity scores.\n",
    "* WordNet provides many limitations due to its age and lack of funding for ongoing maintenance. It provides access to a large variety of the English language but certain words simply do not exist.\n",
    "* There are currently only 2 lists (Countries and Professions). We would like to find community curated lists for: Race, Sexual Orientation and Gender Identity (SOGI), Religion, age, and protected status.\n",
    "\n",
    "\n",
    "##### Key Dependencies and Packages\n",
    "\n",
    "1. [Hugging Face Transformers](https://huggingface.co/) - the model we've designed this iteration for is hosted on hugging face. It is: [distilbert-base-uncased-finetuned-sst-2-english](https://huggingface.co/distilbert-base-uncased-finetuned-sst-2-english).\n",
    "2. [Streamlit](https://streamlit.io) - This is the library we're using to build the prototype app because it is easy to stand up and quick to fix.\n",
    "3. [spaCy](https://spacy.io) - This is the main NLP Library we're using and it runs most of the text manipulation we're doing as part of the project.\n",
    "4. [NLTK + WordNet](https://www.nltk.org/howto/wordnet.html) - This is the initial lexical database we're using because it is accessible directly through Python and it is free. We will be considering a move to [ConceptNet](https://conceptnet.io/) for future iterations based on better lateral movement across edges.\n",
    "5. [Lime](https://github.com/marcotcr/lime) - We chose Lime over Shap because Lime has more of the functionality we need. Shap appears to provide greater performance but is not as easily suited to our original designs.\n",
    "6. [Altair](https://altair-viz.github.io/user_guide/encoding.html) - We're using Altair because it's well integrated into Streamlit.\n",
    "\n",
    "\n",
    "\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "fa224bed-3630-4485-8dbc-670aaf5e6b0a",
   "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": "3.8.8"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}