Merge pull request #49526 from PhilKang0704/broken-link-fix-jeffprosise

Broken links fixing - chnoring

Author: prmerger-automator[bot]

learn-pr/student-evangelism/analyze-review-sentiment-with-keras/includes/2-build-and-train-a-neural-network.md

@@ -104,7 +104,7 @@ In this unit, you'll use Keras to build and train a neural network that analyzes
 
     This code is the essence of how you construct a neural network with Keras. It first instantiates a `Sequential` object representing a "sequential" model — one that is composed of an end-to-end stack of layers in which the output from one layer provides input to the next.
 
-    The next several statements add layers to the model. First is an [embedding layer](https://keras.io/layers/embeddings/), which is crucial to neural networks that process words. The embedding layer essentially maps many-dimensional arrays containing integer word indexes into floating-point arrays containing fewer dimensions. It also allows words with similar meanings to be treated alike. A full treatment of word embeddings is beyond the scope of this lab, but you can learn more by reading [Why You Need to Start Using Embedding Layers](https://towardsdatascience.com/deep-learning-4-embedding-layers-f9a02d55ac12). If you prefer a more scholarly explanation, refer to [Efficient Estimation of Word Representations in Vector Space](https://arxiv.org/pdf/1301.3781.pdf). The call to [Flatten](https://keras.io/layers/core/#flatten) following the addition of the embedding layer reshapes the output for input to the next layer.
+    The next several statements add layers to the model. First is an [embedding layer](https://keras.io/layers/embeddings/), which is crucial to neural networks that process words. The embedding layer essentially maps many-dimensional arrays containing integer word indexes into floating-point arrays containing fewer dimensions. It also allows words with similar meanings to be treated alike. A full treatment of word embeddings is beyond the scope of this lab. If you prefer a more scholarly explanation, refer to [Efficient Estimation of Word Representations in Vector Space](https://arxiv.org/pdf/1301.3781.pdf). The call to [Flatten](https://keras.io/layers/core/#flatten) following the addition of the embedding layer reshapes the output for input to the next layer.
 
     The next three layers added to the model are [dense](https://keras.io/layers/core/#dense) layers, also known as *fully connected* layers. These are the traditional layers that are common in neural networks. Each layer contains *n* nodes or [neurons](https://en.wikipedia.org/wiki/Artificial_neuron), and each neuron receives input from every neuron in the previous layer, hence the term "fully connected." It is these layers that permit a neural network to "learn" from input data by iteratively guessing at the output, checking the results, and fine-tuning the connections to produce better results. The first two dense layers in this network contain 16 neurons each. This number was arbitrarily chosen; you might be able to improve the accuracy of the model by experimenting with different sizes. The final dense layer contains just one neuron because the ultimate goal of the network is to predict one output — namely, a sentiment score from 0.0 to 1.0.
 
@@ -172,4 +172,4 @@ In this unit, you'll use Keras to build and train a neural network that analyzes
 
     What is the computed accuracy of your model?
 
-You probably achieved an accuracy in the 85% to 90% range. That's acceptable considering you built the model from scratch (as opposed to using a pretrained neural network) and the training time was short even without a GPU. It *is* possible to achieve accuracies of 95% or higher with alternate neural network architectures, particularly [recurrent neural networks](https://en.wikipedia.org/wiki/Recurrent_neural_network) (RNNs) that utilize [Long Short-Term Memory](https://en.wikipedia.org/wiki/Long_short-term_memory) (LSTM) layers. Keras makes it easy to build such networks, but training time can increase exponentially. The model that you built strikes a reasonable balance between accuracy and training time. However, if you would like to learn more about building RNNs with Keras, see [Understanding LSTM and its Quick Implementation in Keras for Sentiment Analysis](https://towardsdatascience.com/understanding-lstm-and-its-quick-implementation-in-keras-for-sentiment-analysis-af410fd85b47).
+You probably achieved an accuracy in the 85% to 90% range. That's acceptable considering you built the model from scratch (as opposed to using a pretrained neural network) and the training time was short even without a GPU. It *is* possible to achieve accuracies of 95% or higher with alternate neural network architectures, particularly [recurrent neural networks](https://en.wikipedia.org/wiki/Recurrent_neural_network) (RNNs) that utilize [Long Short-Term Memory](https://en.wikipedia.org/wiki/Long_short-term_memory) (LSTM) layers. Keras makes it easy to build such networks, but training time can increase exponentially. The model that you built strikes a reasonable balance between accuracy and training time. However, if you would like to learn more about building RNNs with Keras, see [Understanding LSTM and its Quick Implementation in Keras for Sentiment Analysis](https://medium.com/data-science/long-short-term-memory-lstm-in-keras-2b5749e953ac).