NextWordGlove.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Some tests of glove dependency!"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 84,
   "metadata": {},
   "outputs": [],
   "source": [
    "%matplotlib inline\n",
    "%config InlineBackend.figure_format = 'retina'\n",
    "\n",
    "import matplotlib.pyplot as plt"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 69,
   "metadata": {},
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "import npp\n",
    "\n",
    "def get_glove_embedding(dim):\n",
    "    with open('glove.6B/glove.6B.%dd.txt'%dim, 'r') as f:\n",
    "        lines = f.readlines()\n",
    "    glove = {}\n",
    "    for line in lines:\n",
    "        parts = line.split(' ')\n",
    "        parts[-1] = parts[-1].replace('\\n', '')\n",
    "        flag = True\n",
    "        for word in parts[0]:\n",
    "            for c in word:\n",
    "                if c not in 'abcdefghijklmnopqrstuvwxyz' and c != '-' and c!= \"'\":\n",
    "                    flag = False\n",
    "                    break\n",
    "        if flag:\n",
    "            glove[parts[0]] = np.array(list(map(float, parts[1:])))\n",
    "            assert glove[parts[0]].shape[0] == dim\n",
    "    glove = {word: glove[word]/np.linalg.norm(glove[word]) for word in glove}\n",
    "    glove['oov'] = np.zeros(dim)\n",
    "    return glove"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "glove = get_glove_embedding(50)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "336674"
      ]
     },
     "execution_count": 3,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "len(glove)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "transcript = open('monkey.txt').read().split(\" \")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "4790"
      ]
     },
     "execution_count": 5,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "len(transcript)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Average words per second: 2.661\n"
     ]
    }
   ],
   "source": [
    "print(\"Average words per second: %0.3f\" % (len(transcript) / (30 * 60)))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [],
   "source": [
    "wordvectors = np.array([glove.get(w.lower().strip(\",.\"), np.zeros(50)) for w in transcript])\n",
    "prev_words = wordvectors[:-1]\n",
    "next_words = wordvectors[1:]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(4790, 50)"
      ]
     },
     "execution_count": 10,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "wordvectors.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 59,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "789"
      ]
     },
     "execution_count": 59,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "len(prev_words[4000:])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Fit model that predicts glove(word_{i-1}) from glove(word_i)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/Users/huth/anaconda3/anaconda3/lib/python3.6/site-packages/ipykernel_launcher.py:1: FutureWarning: `rcond` parameter will change to the default of machine precision times ``max(M, N)`` where M and N are the input matrix dimensions.\n",
      "To use the future default and silence this warning we advise to pass `rcond=None`, to keep using the old, explicitly pass `rcond=-1`.\n",
      "  \"\"\"Entry point for launching an IPython kernel.\n"
     ]
    }
   ],
   "source": [
    "# fit using first 4000 words\n",
    "wt,rank,res,sing = np.linalg.lstsq(next_words[:4000], prev_words[:4000])\n",
    "\n",
    "# test on last 789 words\n",
    "pred_prev = next_words[4000:].dot(wt)\n",
    "\n",
    "# compute correlation for each word\n",
    "corrs = np.array([np.corrcoef(p,pp)[0,1] for p,pp in zip(prev_words[4000:], pred_prev)])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 61,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "Text(0.5,0,'Corr(predicted glove(word i) from glove(word i+1), glove(word i))')"
      ]
     },
     "execution_count": 61,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.hist(corrs[~np.isnan(corrs)], 20)\n",
    "plt.ylabel(\"Num. words\")\n",
    "plt.xlabel(\"Corr(predicted glove(word i) from glove(word i+1), glove(word i))\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "These values are clearly WAY above chance!"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 40,
   "metadata": {},
   "outputs": [],
   "source": [
    "# what if we test by looking at the correlation _across_ words for each glove dimension?\n",
    "corrs2 = np.array([np.corrcoef(p,pp)[0,1] for p,pp in zip(prev_words[4000:].T, pred_prev.T)])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 42,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAW4AAAD8CAYAAABXe05zAAAABHNCSVQICAgIfAhkiAAAAAlwSFlzAAALEgAACxIB0t1+/AAAADl0RVh0U29mdHdhcmUAbWF0cGxvdGxpYiB2ZXJzaW9uIDIuMi4yLCBodHRwOi8vbWF0cGxvdGxpYi5vcmcvhp/UCwAAC0dJREFUeJzt3WuM5Xddx/HP166A9VrsPKIM0yZoUowJuhqFiAokXKrUxD6oCQaRZBMviJfELCGGxEc1MUYemJANXgMRpBJDLKJVwMQHVNtSKaVWSlmhFJWLgqgBK18f7AGn29md/3bnzJnv9vVKTvbMmf8553t+c/a9//zP/LPV3QFgjq/a9AAAXBjhBhhGuAGGEW6AYYQbYBjhBhhGuAGGEW6AYYQbYJhj63jQK6+8snd2dtbx0ACXpDvuuONT3b21ZNu1hHtnZye33377Oh4a4JJUVf+0dFuHSgCGEW6AYYQbYBjhBhhGuAGGEW6AYYQbYBjhBhhGuAGGWcuZk7CfnZO3bOR5T9903UaeFw6SPW6AYYQbYBjhBhhGuAGGEW6AYYQbYBjhBhhGuAGGEW6AYYQbYBjhBhhGuAGGEW6AYYQbYBjhBhhGuAGGEW6AYYQbYJhF4a6qX6iqe6rqA1X1h1X1pHUPBsDe9g13VT0lyc8lOd7d35bksiQ3rnswAPa29FDJsSRfU1XHklye5KH1jQTA+ewb7u7+eJJfT/LRJJ9I8tnu/ot1DwbA3o7tt0FVXZHk+iRXJ/n3JG+tqpd29xvP2u5EkhNJsr29vYZRYa6dk7ds7LlP33Tdxp6b9VhyqOT5ST7S3Z/s7v9J8rYkzzp7o+4+1d3Hu/v41tbWQc8JwMqScH80yfdU1eVVVUmel+Te9Y4FwLksOcZ9W5Kbk9yZ5O7VfU6teS4AzmHfY9xJ0t2vTfLaNc8CwALOnAQYRrgBhhFugGGEG2AY4QYYRrgBhhFugGGEG2AY4QYYRrgBhhFugGGEG2AY4QYYRrgBhhFugGGEG2AY4QYYRrgBhhFugGGEG2AY4QYYRrgBhhFugGGEG2AY4QYYRrgBhhFugGGEG2AY4QYYRrgBhhFugGGEG2AY4QYYRrgBhhFugGGEG2AY4QYYZlG4q+qbqurmqvqHqrq3qr533YMBsLdjC7d7XZJ3dvcNVfWEJJevcSYAzmPfcFfVNyR5TpKfSJLu/mKSL653LADOZcmhkmuSfDLJ71bV+6rqDVX1tWueC4BzWHKo5FiS70jyyu6+rapel+Rkkl/ZvVFVnUhyIkm2t7cPes612zl5y8ae+/RN123keTf5mjfl8fiaufQs2eN+MMmD3X3b6uubcybkj9Ddp7r7eHcf39raOsgZAdhl33B39z8n+VhVfevqpucl+eBapwLgnJb+Vskrk7xp9RslDyR5+fpGAuB8FoW7u+9KcnzNswCwgDMnAYYRboBhhBtgGOEGGEa4AYYRboBhhBtgGOEGGEa4AYYRboBhhBtgGOEGGEa4AYYRboBhhBtgGOEGGEa4AYYRboBhhBtgGOEGGEa4AYYRboBhhBtgGOEGGEa4AYYRboBhhBtgGOEGGEa4AYYRboBhhBtgGOEGGEa4AYYRboBhhBtgGOEGGEa4AYZZHO6quqyq3ldVf7rOgQA4vwvZ435VknvXNQgAyywKd1VdleS6JG9Y7zgA7GfpHvdvJvnlJF9a4ywALHBsvw2q6oeS/Gt331FVP3Ce7U4kOZEk29vbBzYgcHF2Tt6y6REO3embrtv0CGu1ZI/72UleUlWnk7w5yXOr6o1nb9Tdp7r7eHcf39raOuAxAfiyfcPd3a/u7qu6eyfJjUne1d0vXftkAOzJ73EDDLPvMe7duvs9Sd6zlkkAWMQeN8Awwg0wjHADDCPcAMMIN8Awwg0wjHADDCPcAMMIN8Awwg0wjHADDCPcAMMIN8Awwg0wjHADDCPcAMMIN8Awwg0wzAX912Wsx87JWzY9AjCIPW6AYYQbYBjhBhhGuAGGEW6AYYQbYBjhBhhGuAGGEW6AYYQbYBjhBhhGuAGGEW6AYYQbYBjhBhhGuAGGEW6AYYQbYJh9w11VT62qd1fVvVV1T1W96jAGA2BvS/7PyYeT/FJ331lVX5/kjqq6tbs/uObZANjDvnvc3f2J7r5zdf0/ktyb5CnrHgyAvV3QMe6q2knyzCS3rWMYAPa35FBJkqSqvi7JHyf5+e7+3B7fP5HkRJJsb28/5oF2Tt7ymO8LkGyuI6dvuu5QnmfRHndVfXXORPtN3f22vbbp7lPdfby7j29tbR3kjADssuS3SirJbye5t7t/Y/0jAXA+S/a4n53kx5M8t6ruWl1evOa5ADiHfY9xd/ffJKlDmAWABZw5CTCMcAMMI9wAwwg3wDDCDTCMcAMMI9wAwwg3wDDCDTCMcAMMI9wAwwg3wDDCDTCMcAMMI9wAwwg3wDDCDTCMcAMMI9wAwwg3wDDCDTCMcAMMI9wAwwg3wDDCDTCMcAMMI9wAwwg3wDDCDTCMcAMMI9wAwwg3wDDCDTCMcAMMI9wAwwg3wDDCDTDMonBX1Qur6r6qur+qTq57KADObd9wV9VlSX4ryYuSXJvkx6rq2nUPBsDeluxxf3eS+7v7ge7+YpI3J7l+vWMBcC5Lwv2UJB/b9fWDq9sA2IBjC7apPW7rR21UdSLJidWXn6+q+y5msHO4Msmn1vC401mXvVmXR7MmezuQdalfu6i7P23phkvC/WCSp+76+qokD529UXefSnJq6RM/FlV1e3cfX+dzTGRd9mZdHs2a7G3auiw5VPJ3SZ5eVVdX1ROS3Jjk7esdC4Bz2XePu7sfrqqfTfLnSS5L8jvdfc/aJwNgT0sOlaS735HkHWueZYm1HooZzLrszbo8mjXZ26h1qe5Hfc4IwBHmlHeAYY5MuPc7rb6qnlhVb1l9/7aq2tn1vVevbr+vql5wmHOv22Ndl6raqar/rqq7VpfXH/bs67JgTZ5TVXdW1cNVdcNZ33tZVX1odXnZ4U29fhe5Lv+7671ySf3ywYJ1+cWq+mBVvb+q/qqqnrbre0fz/dLdG7/kzIeeH05yTZInJPn7JNeetc1PJ3n96vqNSd6yun7tavsnJrl69TiXbfo1HYF12UnygU2/hg2tyU6Sb0/yB0lu2HX7k5M8sPrzitX1Kzb9mja9LqvvfX7Tr2GD6/KDSS5fXf+pXX+Hjuz75ajscS85rf76JL+/un5zkudVVa1uf3N3f6G7P5Lk/tXjXQouZl0uVfuuSXef7u73J/nSWfd9QZJbu/sz3f1vSW5N8sLDGPoQXMy6XMqWrMu7u/u/Vl++N2fOVUmO8PvlqIR7yWn1X9mmux9O8tkk37zwvlNdzLokydVV9b6q+uuq+r51D3tILubn/Xh/r5zPk6rq9qp6b1X9yMGOtlEXui6vSPJnj/G+h2bRrwMegiWn1Z9rm0Wn5A91MevyiSTb3f3pqvrOJH9SVc/o7s8d9JCH7GJ+3o/398r5bHf3Q1V1TZJ3VdXd3f3hA5ptkxavS1W9NMnxJN9/ofc9bEdlj3vJafVf2aaqjiX5xiSfWXjfqR7zuqwOHX06Sbr7jpw5zvcta594/S7m5/14f6+cU3c/tPrzgSTvSfLMgxxugxatS1U9P8lrkryku79wIffdiE0fZF99CHAsZw78X53//wDhGWdt8zN55Idwf7S6/ow88sPJB3LpfDh5Meuy9eV1yJkPZj6e5Mmbfk2HsSa7tv29PPrDyY/kzAdNV6yuj1+TA1iXK5I8cXX9yiQfylkf4E29LPw79Myc2bF5+lm3H9n3y8YH2LVIL07yj6sFfM3qtl/NmX8Bk+RJSd6aMx8+/m2Sa3bd9zWr+92X5EWbfi1HYV2S/GiSe1Zv1DuT/PCmX8shrsl35cze0n8m+XSSe3bd9ydXa3V/kpdv+rUchXVJ8qwkd6/eK3cnecWmX8shr8tfJvmXJHetLm8/6u8XZ04CDHNUjnEDsJBwAwwj3ADDCDfAMMINMIxwAwwj3ADDCDfAMP8Hc1frNguSom4AAAAASUVORK5CYII=\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# also clearly above chance\n",
    "plt.hist(corrs2, 10);"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# What if we used random vectors instead of glove?"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 34,
   "metadata": {},
   "outputs": [],
   "source": [
    "rand_vectors = {w: np.random.randn(50) for w in glove.keys()}"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 35,
   "metadata": {},
   "outputs": [],
   "source": [
    "randwordvectors = np.array([rand_vectors.get(w.lower().strip(\",.\"), np.zeros(50)) for w in transcript])\n",
    "rand_prev_words = randwordvectors[:-1]\n",
    "rand_next_words = randwordvectors[1:]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 62,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/Users/huth/anaconda3/anaconda3/lib/python3.6/site-packages/ipykernel_launcher.py:1: FutureWarning: `rcond` parameter will change to the default of machine precision times ``max(M, N)`` where M and N are the input matrix dimensions.\n",
      "To use the future default and silence this warning we advise to pass `rcond=None`, to keep using the old, explicitly pass `rcond=-1`.\n",
      "  \"\"\"Entry point for launching an IPython kernel.\n",
      "/Users/huth/anaconda3/anaconda3/lib/python3.6/site-packages/numpy/lib/function_base.py:2534: RuntimeWarning: invalid value encountered in true_divide\n",
      "  c /= stddev[:, None]\n",
      "/Users/huth/anaconda3/anaconda3/lib/python3.6/site-packages/numpy/lib/function_base.py:2535: RuntimeWarning: invalid value encountered in true_divide\n",
      "  c /= stddev[None, :]\n"
     ]
    }
   ],
   "source": [
    "# fit model that predicts rand_embedding(word i-1) from rand_embedding(word i)\n",
    "randwt,randrank,randres,randsing = np.linalg.lstsq(rand_next_words[:4000], rand_prev_words[:4000])\n",
    "rand_pred_prev = rand_next_words[4000:].dot(randwt)\n",
    "randcorrs = np.array([np.corrcoef(p,pp)[0,1] for p,pp in zip(rand_prev_words[4000:], rand_pred_prev)])\n",
    "randcorrs2 = np.array([np.corrcoef(p,pp)[0,1] for p,pp in zip(rand_prev_words[4000:].T, rand_pred_prev.T)])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 86,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "image/png": {
       "height": 250,
       "width": 377
      }
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "# this looks much closer to zero, but is there still significant information?\n",
    "plt.hist(randcorrs[~np.isnan(randcorrs)], 20);"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 77,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/Users/huth/code/notebook-sharing/npp.py:11: RuntimeWarning: invalid value encountered in true_divide\n",
      "  zscore = lambda v: (v-v.mean(0))/v.std(0)\n"
     ]
    }
   ],
   "source": [
    "# significance test.....\n",
    "nperm = 1000\n",
    "perm_ccs = np.zeros(nperm)\n",
    "for ii in range(nperm):\n",
    "    permwt = np.random.permutation(randwt)\n",
    "    perm_rand_pred_prev = rand_next_words[4000:].dot(permwt)\n",
    "    #perm_randcorrs = np.array([np.corrcoef(p,pp)[0,1] for p,pp in zip(rand_prev_words[4000:], perm_rand_pred_prev)])\n",
    "    perm_randcorrs = np.nanmean(npp.mcorr(rand_prev_words[4000:].T, perm_rand_pred_prev.T))\n",
    "    perm_ccs[ii] = np.nanmean(perm_randcorrs)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 85,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<matplotlib.legend.Legend at 0x1155315c0>"
      ]
     },
     "execution_count": 85,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "image/png": {
       "height": 263,
       "width": 470
      }
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.figure(facecolor=\"white\")\n",
    "plt.hist(perm_ccs, 50, label=\"Null dist.\", density=True);\n",
    "plt.vlines(np.nanmean(randcorrs), 0, plt.ylim()[1], label=\"True value\", color=\"red\")\n",
    "plt.xlabel(\"Corr(predicted word i random vector from word i+1 random vector, word i random vector)\")\n",
    "plt.ylabel(\"Density\")\n",
    "plt.legend()\n",
    "\n",
    "# Yeah this looks really significant!"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Use random vector to predict glove vectors?"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 46,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/Users/huth/anaconda3/anaconda3/lib/python3.6/site-packages/ipykernel_launcher.py:1: FutureWarning: `rcond` parameter will change to the default of machine precision times ``max(M, N)`` where M and N are the input matrix dimensions.\n",
      "To use the future default and silence this warning we advise to pass `rcond=None`, to keep using the old, explicitly pass `rcond=-1`.\n",
      "  \"\"\"Entry point for launching an IPython kernel.\n",
      "/Users/huth/anaconda3/anaconda3/lib/python3.6/site-packages/numpy/lib/function_base.py:2534: RuntimeWarning: invalid value encountered in true_divide\n",
      "  c /= stddev[:, None]\n",
      "/Users/huth/anaconda3/anaconda3/lib/python3.6/site-packages/numpy/lib/function_base.py:2535: RuntimeWarning: invalid value encountered in true_divide\n",
      "  c /= stddev[None, :]\n"
     ]
    }
   ],
   "source": [
    "# fit model that predicts glove(word i-1) from random_emb(word i)\n",
    "rand2wt,rand2rank,rand2res,rand2sing = np.linalg.lstsq(rand_next_words[:4000], prev_words[:4000])\n",
    "rand2_pred_prev = rand_next_words[4000:].dot(rand2wt)\n",
    "rand2corrs = np.array([np.corrcoef(p,pp)[0,1] for p,pp in zip(prev_words[4000:], rand2_pred_prev)])\n",
    "rand2corrs2 = np.array([np.corrcoef(p,pp)[0,1] for p,pp in zip(prev_words[4000:].T, rand2_pred_prev.T)])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 50,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "Text(0.5,0,'Corr(predicted word i glove vector from word i+1 random vector, word i glove vector)')"
      ]
     },
     "execution_count": 50,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.hist(rand2corrs[~np.isnan(rand2corrs)], 20);\n",
    "plt.ylabel(\"Num. words\")\n",
    "plt.xlabel(\"Corr(predicted word i glove vector from word i+1 random vector, word i glove vector)\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Definitely positive on average!"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Regress previous word vector out of next word vector, then try the same?"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 56,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/Users/huth/anaconda3/anaconda3/lib/python3.6/site-packages/ipykernel_launcher.py:1: FutureWarning: `rcond` parameter will change to the default of machine precision times ``max(M, N)`` where M and N are the input matrix dimensions.\n",
      "To use the future default and silence this warning we advise to pass `rcond=None`, to keep using the old, explicitly pass `rcond=-1`.\n",
      "  \"\"\"Entry point for launching an IPython kernel.\n",
      "/Users/huth/anaconda3/anaconda3/lib/python3.6/site-packages/ipykernel_launcher.py:5: FutureWarning: `rcond` parameter will change to the default of machine precision times ``max(M, N)`` where M and N are the input matrix dimensions.\n",
      "To use the future default and silence this warning we advise to pass `rcond=None`, to keep using the old, explicitly pass `rcond=-1`.\n",
      "  \"\"\"\n",
      "/Users/huth/anaconda3/anaconda3/lib/python3.6/site-packages/numpy/lib/function_base.py:2534: RuntimeWarning: invalid value encountered in true_divide\n",
      "  c /= stddev[:, None]\n",
      "/Users/huth/anaconda3/anaconda3/lib/python3.6/site-packages/numpy/lib/function_base.py:2535: RuntimeWarning: invalid value encountered in true_divide\n",
      "  c /= stddev[None, :]\n"
     ]
    }
   ],
   "source": [
    "# predict glove(word i) from glove(word i-1)\n",
    "nextwt,_,_,_ = np.linalg.lstsq(prev_words[:4000], next_words[:4000])\n",
    "\n",
    "# subtract prediction to get glove(word i) \"independent of\" glove(word i-1)\n",
    "next_words_sub = next_words - prev_words.dot(nextwt)\n",
    "\n",
    "# fit model that predicts glove(word i-1) using the modified vectors from above\n",
    "subwt,subrank,subres,subsing = np.linalg.lstsq(next_words_sub[:4000], prev_words[:4000])\n",
    "subpred_prev = next_words_sub[4000:].dot(subwt)\n",
    "subcorrs = np.array([np.corrcoef(p,pp)[0,1] for p,pp in zip(prev_words[4000:], subpred_prev)])\n",
    "subcorrs2 = np.array([np.corrcoef(p,pp)[0,1] for p,pp in zip(prev_words[4000:].T, subpred_prev.T)])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 60,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "Text(0.5,0,'Corr(pred. glove(word i) from glove(word i+1) - pred. glove(word i+1) from word i, word i glove vector)')"
      ]
     },
     "execution_count": 60,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.hist(subcorrs[~np.isnan(subcorrs)], 20);\n",
    "plt.ylabel(\"Num. words\")\n",
    "plt.xlabel(\"Corr(pred. glove(word i) from glove(word i+1) - pred. glove(word i+1) from word i, word i glove vector)\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 58,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.033224859033786525"
      ]
     },
     "execution_count": 58,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "subcorrs[~np.isnan(subcorrs)].mean()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "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": "3.6.8"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}