private/NlpInPracticeCourse/InformationExtraction: IA161_Hypernym_Extraction.ipynb

{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "provenance": []
    },
    "kernelspec": {
      "name": "python3",
      "display_name": "Python 3"
    },
    "language_info": {
      "name": "python"
    }
  },
  "cells": [
    {
      "cell_type": "markdown",
      "source": [
        "# Hypernym Extraction\n",
        "Hypernym extraction is crucial in lexicography (a good start of a dictionary definition is the hypernym) and ontology engineering (hypernym-hyponym or subclass-superclass is the backbone relation is ontologies).\n",
        "\n",
        "## Get the data\n",
        "\n",
        "Download input data using wget.\n",
        "Check the nature of the data by listing first n lines (by default n=10)."
      ],
      "metadata": {
        "id": "LCYhjKC9TosE"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "!wget http://nlp.fi.muni.cz/trac/research/raw-attachment/wiki/private/NlpInPracticeCourse/RelationExtraction/input.txt"
      ],
      "metadata": {
        "id": "_zIYBky4TzlW"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "!head input.txt"
      ],
      "metadata": {
        "id": "thqNaB3zEcq6"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "## Tools\n",
        "\n",
        "### Tokenizer and PoS tagger\n",
        "\n",
        "We use NLTK standard tokenizer to split the texts into words. And then NLTK Part-of-Speech tagger to get PoS information."
      ],
      "metadata": {
        "id": "7U_pqaF6UOe_"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "import nltk\n",
        "from nltk.tokenize import word_tokenize\n",
        "\n",
        "nltk.download('punkt')\n",
        "nltk.download('averaged_perceptron_tagger')"
      ],
      "metadata": {
        "id": "sHUnUjjZUoCt"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "## Extract hypernyms\n",
        "\n",
        "Our script reads the lines from input.txt (in the form: word|definition) and calls *find_hyper* method to detect hypernym from the definition. The detection is quite naive, it takes first word tagged as noun (NN). You should extend and modify *find_hyper* method to produce better results."
      ],
      "metadata": {
        "id": "Bcfjj8zjUt99"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "def find_hyper(word, text):\n",
        "    tokenized = word_tokenize(text)\n",
        "    pos_tagged = nltk.pos_tag(tokenized)\n",
        "    # find first noun\n",
        "    nouns = [pos_tagged[i] for i in range(len(pos_tagged)) if pos_tagged[i][1] == 'NN' and (i>0 and pos_tagged[i-1][1] in ['JJ', 'DT', 'NNP','NN']) and (i<len(pos_tagged)-1 and pos_tagged[i+1][1]!='NN')]\n",
        "    if len(nouns) > 0:\n",
        "        return (word, nouns[0][0])\n",
        "    return (word, u'')\n",
        "\n",
        "\n",
        "if __name__ == \"__main__\":\n",
        "\tinput_data = 'input.txt'\n",
        "\tfor line in open(input_data, 'r'):\n",
        "\t\tif line.strip() != '':\n",
        "\t\t\tline_split = line.strip().split('|')\n",
        "\t\t\thyper = find_hyper(line_split[0], line_split[1])\n",
        "\t\t\tprint(\"%s = %s\" % hyper)\n"
      ],
      "metadata": {
        "id": "teuFO7p_VE0y"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "for line in open(input_data, 'r'):\n",
        "  if line.strip() != '':\n",
        "    line_split = line.strip().split('|')\n",
        "    sentence = line_split[1]\n",
        "    print(sentence)\n",
        "    tokenized = word_tokenize(sentence)\n",
        "    print(\"tokenized\")\n",
        "    print(tokenized)\n",
        "    pos_tagged = nltk.pos_tag(tokenized)\n",
        "    print(\"POS-tagged\")\n",
        "    print(pos_tagged)\n",
        "    tags = [y for x,y in pos_tagged]\n",
        "    print('-'*30)"
      ],
      "metadata": {
        "id": "l06TvsEFxzBk"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "## Evaluate\n",
        "For evaluation, select some straighforward metric such as accuracy. Calculate and publish the performance of the naive approach, then re-calculate and publish performance for improved `find_hyper` function. For evaluation, use the `gold_en.txt`."
      ],
      "metadata": {
        "id": "jcKqLv7mY0X-"
      }
    },
    {
      "cell_type": "markdown",
      "source": [
        "# Generative Models"
      ],
      "metadata": {
        "id": "V9o6gHFyJ8Z_"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "!pip install openai"
      ],
      "metadata": {
        "id": "dyVOSQZVKIjg"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "API_KEY=\"\" # your key"
      ],
      "metadata": {
        "id": "kaPQR8PKJ-qh"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "import openai\n",
        "openai.__version__"
      ],
      "metadata": {
        "id": "kzhHwHdJKGYa"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "\n",
        "# Specify the model\n",
        "model = \"gpt-4o-mini\"\n",
        "#model = \"gpt-3.5-turbo-1106\"\n",
        "\n",
        "# Define the prompt for the completion\n",
        "prompt = \"\"\"\n",
        "I give you definitions of words in the format: word|definition. Extract the nearest hypernym of the word from the definition.\n",
        "{}\n",
        "\"\"\"\n",
        "\n",
        "data = \"\"\"\n",
        "chair|a piece of furniture for one person to sit on, with a back, legs, and sometimes two arms\n",
        "table|a piece of furniture that consists of a flat surface held above the floor, usually by legs\n",
        "rose|a flower that has a sweet smell and thorns (=sharp pieces) on its stem\n",
        "herb|a plant used for adding flavour to food or as a medicine\n",
        "dog|an animal kept as a pet, for guarding buildings, or for hunting\n",
        "tiger|a large Asian wild animal that has yellowish fur with black lines and is a member of the cat family\n",
        "hippopotamus|a large African animal with a wide head and mouth and thick grey skin\n",
        "sunflower|a very tall plant that has large yellow flowers with a round brown centre\n",
        "horse|a large animal that people ride\n",
        "human|a person\n",
        "\"\"\"\n",
        "\n",
        "import os\n",
        "from openai import OpenAI\n",
        "\n",
        "client = OpenAI(\n",
        "    # This is the default and can be omitted\n",
        "    api_key=API_KEY\n",
        ")\n",
        "\n",
        "chat_completion = client.chat.completions.create(\n",
        "    messages=[\n",
        "        {\n",
        "            \"role\": \"user\",\n",
        "            \"content\": prompt.format(data),\n",
        "        }\n",
        "    ],\n",
        "    model=model, #\"gpt-3.5-turbo\",\n",
        ")"
      ],
      "metadata": {
        "id": "QURXL17EKLr-"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "print(chat_completion.choices[0].message.content)"
      ],
      "metadata": {
        "id": "YNogJwXXKtkh"
      },
      "execution_count": null,
      "outputs": []
    }
  ]
}