nlp | report | Network | 神经网络 | Machine learning | project作业 – Text Classification

Text Classification

nlp | report | Network | network作业 | Machine learning | project作业 – 本题是一个利用Network进行练习的代做, 对Network的流程进行训练解析, 涵盖了nlp/report/Network/network/Machine learning等程序代做方面, 这个项目是project代写的代写题目

project代写 代写project

  1. Text Classification It is highly recommended that you complete this project using Keras^1 and Python.
(a) In this problem, we are trying to build a classifier to analyze the sentiment of
reviews. You are provided with text data in two folders: one folder involves
positive reviews, and one folder involves negative reviews.
(b)Data Exploration and Pre-processing
i. You can use binary encoding for the sentiments , i.ey= 1 for positive senti-
ments andy=1 for negative sentiments.
ii. The data are pretty clean. Remove the punctuation and numbers from the
iii. The name of each text file starts withcvnumber. Use text files 0-699 in each
class for training and 700-999 for testing.
iv. Count the number of unique words in the whole dataset (train + test) and
print it out.
v. Calculate the average review length and the standard deviation of review
lengths.  report the results.
vi. Plot the histogram of review lengths.
vii. To represent each text (= data point), there are many ways. In NLP/Deep
Learning terminology, this task is called tokenization. It is common to rep-
resent text using popularity/ rank of words in text. The most common word
in the text will be represented as 1, the second most common word will be
represented as 2, etc. Tokenize each text document using this method.^2
viii. Select a review lengthLthat 70% of the reviews have a length below it. If
you feel more adventurous, set the threshold to 90%.
ix. Truncate reviews longer thanLwords and zero-pad reviews shorter thanL
so that all texts (= data points) are of lengthL.^3
(c) Word Embeddings
i. One can use tokenized text as inputs to a deep neural network. However, a re-
cent breakthrough in  Machine learning 人工智能"> nlp suggests that more sophisticated representations of
text yield better results. These sophisticated representations are calledword
embeddings. Word embedding is a term used for representation of words
for text analysis, typically in the form of a real-valued vector that encodes
the meaning of the word such that the words that are closer in the vector
space are expected to be similar in meaning.^4. Most deep learning modules
(including Keras) provide a convenient way to convert positive integer rep-
resentations of words into a word embedding by an Embedding layer. The
layer accepts arguments that define the mapping of words into embeddings,

(^1) (^2) Keras has an API called Tokenizer. It can yield bag of words, one-hot encoded features, etc. (^3) Keras has padsequences for doing this. (^4)

including the maximum number of expected words also called the vocabulary
size (e.g. the largest integer value). The layer also allows you to specify the
dimension for each word vector, called the output dimension. We would like
to use a word embedding layer for this project. Assume that we are inter-
ested in the top 5,000 words. This means that in each integer sequence that
represents each document, we set to zero those integers that represent words
that are not among the top 5,000 words in the document.^5 If you feel more
adventurous, use all the words that appear in this corpus. Choose the length
of the embedding vector for each word to be 32. Hence, each document is
represented as a 32Lmatrix.
ii. Flatten the matrix of each document to a vector.
(d)Multi-Layer Perceptron
i. Train a MLP with three (dense) hidden layers each of which has 50 ReLUs
and one output layer with a single sigmoid neuron. Use a dropout rate of
20% for the first layer and 50% for the other layers. Use ADAM optimizer
and binary cross entropy loss (which is equivalent to having a softmax in the
output). To avoid overfitting, just set the number of epochs as 2. Use a batch
size of 10.
ii. Report the train and test accuracies of this model.
(e) One-Dimensional Convolutional Neural Network:
Although CNNs are mainly used for image data, they can also be applied to text
data, as text also has adjacency information. Keras supports one-dimensional
convolutions and pooling by the Conv1D and MaxPooling1D classes respectively.
i. After the embedding layer, insert a Conv1D layer. This convolutional layer
has 32 feature maps , and each of the 32 kernels has size 3, i.e. reads embedded
word representations 3 vector elements of the word embedding at a time. The
convolutional layer is followed by a 1D max pooling layer with a length and
stride of 2 that halves the size of the feature maps from the convolutional
layer. The rest of the   network is the same as the neural network above.
ii. Report the train and test accuracies of this model.
(f)Long Short-Term Memory Recurrent Neural Network:
The structure of the LSTM we are going to use is shown in the following figure.
i. Each word is represented to LSTM as a vector of 32 elements and the LSTM
is followed by a dense layer of 256 ReLUs. Use a dropout rate of 0.2 for both
LSTM and the dense layer. Train the model using 10-50 epochs and batch
size of 10.
ii. Report the train and test accuracies of this model.

(^5) This is done by setting an argument in the embedding layer provided by Keras. Exam- ple: model.add(Embedding(topwords, 32, inputlength=maxwords)), where topwords=5,000 and maxwords=L.