diff --git a/chapter09_natural-language-processing/tree-lstm.ipynb b/chapter09_natural-language-processing/tree-lstm.ipynb
index a78c821..8483eda 100644
--- a/chapter09_natural-language-processing/tree-lstm.ipynb
+++ b/chapter09_natural-language-processing/tree-lstm.ipynb
@@ -6,17 +6,50 @@
    "source": [
     "# Tree LSTM modeling for semantic relatedness\n",
     "\n",
+    "Just five years ago, many of the most successful models for doing supervised learning with text\n",
+    "ignored word order altogether. \n",
+    "Some of the most successful models represented documents or sentences \n",
+    "with the order-invariant *bag-of-words* representation.\n",
+    "Anyone thinking hard should probably have realized that these models couldn't dominate forever.\n",
+    "That's because we all know that word order actually does matter.\n",
+    "Bag-of-words models, which ignored word order, left some information on the table.\n",
+    "\n",
+    "The recurrent neural networks that\n",
+    "[we introduced in chapter 5](../chapter05_recurrent-neural-networks/simple-rnn.ipynb)\n",
+    "model word order, by passing over the sequence of words in order,\n",
+    "updating the models representation of the sentence after each word. \n",
+    "And, with LSTM recurrent cells and training on GPUS, \n",
+    "even the straightforward LSTM far outpaces classical approaches,\n",
+    "on a number of tasks, including language modeling,\n",
+    "named entity recognition and more. \n",
+    "\n",
+    "But while those models are impressive, they still may be leaving some knowledge on the table.\n",
+    "To begin with, we know a priori that sentence have a grammatical structure. \n",
+    "And we already have some tools that are very good at recovering parse trees that reflect grammatical structure of the sentences.\n",
+    "While it may be possible for an LSTM to learn this informatino implicitly,\n",
+    "it's often a good idea to build known information into the structure of a neural network.\n",
+    "Take for example convolutional neural networks.\n",
+    "They build in the prior knowledge that low level feature should be translation-invariant.\n",
+    "It's possible to come up with a fully connected net that does the same thing,\n",
+    "but it would require many more nodes and would be much more susceptible to overfitting. \n",
+    "In this case, we would like to build the grammatical tree structure of the sentences \n",
+    "into the architecture of an LSTM recurrent neural network.\n",
+    "This tutorial walks through *tree LSTMs*,\n",
+    "an approach that does precisely that.\n",
+    "The models here are based on the [tree-structured LSTM](https://nlp.stanford.edu/pubs/tai-socher-manning-acl2015.pdf)\n",
+    "by Kai Sheng Tai, Richard Socher, and Chris Manning.\n",
+    "Our implementation borrows from [this Pytorch example](https://github.com/dasguptar/treelstm.pytorch).\n",
+    "\n",
+    "\n",
     "### Sentences involving Compositional Knowledge\n",
     "This tutorial walks through training a child-sum Tree LSTM model for analyzing semantic relatedness of sentence pairs given their dependency parse trees.\n",
     "\n",
     "### Preliminaries\n",
-    "Requires the latest MXNet with the new `gluon` interface. One can either build from source or install the pre-release package through `pip install --pre mxnet`. Use of GPUs is preferred if one wants to run the complete training to match the state-of-the-art results.\n",
-    "\n",
-    "Besides, to show a progress meter, one should install the `tqdm` (\"progress\" in Arabic) through  `pip install tqdm`. One should also install the HTTP library through `pip install requests`.\n",
-    "\n",
+    "Before getting going, you'll probably want to note a couple preliminary details:\n",
     "\n",
-    "### Inspiration\n",
-    "This tutorial borrows heavily from [Pytorch](https://github.com/dasguptar/treelstm.pytorch) example."
+    "* Use of GPUs is preferred if one wants to run the complete training to match the state-of-the-art results.\n",
+    "* To show a progress meter, one should install the `tqdm` (\"progress\" in Arabic) through  `pip install tqdm`. One should also install the HTTP library through `pip install requests`.\n",
+    "\n"
    ]
   },
   {
@@ -654,7 +687,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.5.3"
+   "version": "3.4.3"
   }
  },
  "nbformat": 4,