DEEP LEARNING SENTIMENT ANALYSIS OF AMAZON.COM REVIEWS AND RATINGS

International Journal on Soft Computing, Artificial Intelligence and Applications (IJSCAI), Vol.8, No.1, February 2019
DOI :10.5121/ijscai.2019.8101 1
DEEP LEARNING SENTIMENT ANALYSIS OF
AMAZON.COM REVIEWS AND RATINGS
Nishit Shrestha and Fatma Nasoz
Department of Computer Science, University of Nevada Las Vegas, Las Vegas, Nevada,
USA
ABSTRACT
Our study employs sentiment analysis to evaluate the compatibility of Amazon.com reviews with their
corresponding ratings. Sentiment analysis is the task of identifying and classifying the sentiment expressed
in a piece of text as being positive or negative. On e-commerce websites such as Amazon.com, consumers
can submit their reviews along with a specific polarity rating. In some instances, there is a mismatch between
the review and the rating. To identify the reviews with mismatched ratings we performed sentiment analysis
using deep learning on Amazon.com product review data. Product reviews were converted to vectors using
paragraph vector, which then was used to train a recurrent neural network with gated recurrent unit. Our
model incorporated both semantic relationship of review text and product information. We also developed a
web service application that predicts the rating score for a submitted review using the trained model and if
there is a mismatch between predicted rating score and submitted rating score, it provides feedback to the
reviewer.
KEYWORDS
Deep Learning, Sentiment Analysis, User Reviews and Ratings
1. INTRODUCTION
Sentiment analysis is the task of computationally identifying and categorizing the sentiment
expressed by an author in a piece of text. It has a wide range of applications in industry from
forecasting market movements based on sentiment expressed in news and blogs, to identifying
customer satisfaction and dissatisfaction from their reviews and social media posts. It also forms
the basis for other applications like recommender systems.
Today, most e-commerce websites have a separate section where their customers can post reviews
for products or service. Important information like customers' opinion on products, reasons for
negative reviews, suggestions, etc., can be extracted from the posted reviews by performing
sentiment analysis on them. Consumers can also assign a numerical value (i.e., rating) to the product
or service they are reviewing. On Amazon.com the rating can be between 1 and 5 where 1 is the
worst and 5 is the best. In some instances, there is a mismatch between a customer’s review and
rating. It is important to identify the reviews with mismatched ratings since individual ratings are
used to compute the average rating.
Conventionally, a review text was converted to fixed-length feature vector using bag-of-words or
bag-of-n-grams and these feature vectors were later used to train a shallow classifier such as naïve
Bayes or support vector machine [1]. Although bag-of-words performed surprisingly well and was
popular for many years, it has two major flaws. It loses ordering of words and doesn't consider the
semantic relationship between words. For example, words such as “bad", “worst", and “Las Vegas"
are equally distant despite the fact that “bad" should be semantically closer to “worst" than “Las

2
Vegas". Although bag-of-n-grams somewhat considers word order in short context, it suffers from
data sparsity and high dimensionality [2].
Recently, deep learning has shown promising results in the field of sentiment analysis. They have
been shown to outperform bag-of-words approach when it comes to feature generation. Mikolov et
al. [3] introduced an efficient way of representing words in vector space by predicting the current
word given context and by predicting surrounding words given current word. Their proposed
unsupervised algorithm was able to embed words in a continuous vector space where semantically
similar words are mapped to nearby points.
Mikolov et al. also [4] introduced word vectors, which is an unsupervised algorithm to efficiently
capture semantics of words. Word vectors represent words as vectors and semantically similar
words are closer to each other in vector space. In other words, similar words such as “good" and
“great" are closer to each other and are far apart from words like “quick", etc. Interestingly, this
vector representation of text also captures linguistic patterns. For example, the result of vector
calculation vec("King") - vec("Queen") + vec("Woman") will output a vector that is closer to the
vector representation of word “Man" than to any other word. Their work was made available as an
open-source project titled word2vec and can be found on Google Code [5].
Socher et al. [6] proposed a recursive neural tensor network for semantic compositionality over a
sentiment tree bank. Their proposed model achieved state-of-the-art result for binary sentiment
classification of Stanford sentiment tree bank.
Kim [7] designed a simple yet powerful convolutional neural network (CNN) that achieved a very
good performance across different datasets for the task of sentiment analysis. The input layer of
this network comprised of concatenated word2vec word embeddings, which is then followed by a
convolutional layer with multiple filters, then a max pooling layer and finally a softmax layer.
Johnson et al. [8] used a variation of bag-of-words model to create feature vectors instead of using
low-dimensional word vectors as input to the convolutional neural network (CNN). They also
experimented with parallel CNN that had two or more convolutional layers in parallel to learn
multiple types of embeddings from one hot input vector.
Ouyang et al. [9] used a seven-layer convolutional neural network architecture to perform sentiment
analysis on movie reviews collected from rottentomatoes.com. Their architecture consisted of three
convolutional layers, three max-pooling layers and a fully-connected layer with softmax activation.
Following on previous work of Mikolov et al. [3][4][5], Le et al. [2] developed paragraph vectors,
which is also an unsupervised algorithm and learns fixed-length feature vectors for variable length
text such as sentences, paragraphs, or an entire document. Paragraph vectors has been shown to
outperform bag-of-words models and other form of text representation and has achieved state-of-
the-art results in several classification tasks including sentiment analysis. Since they are faster to
train and have high efficiency in capturing syntactic and semantic relationship of a text, our model
used paragraph vectors to learn low-dimensional vector of review text.
In the past, most researchers were interested in only the syntactic and semantic relationship between
review texts. Currently, there is a growing interest in adding user and product information to
strengthen the feature vector. Tang et al. [10] proposed a neural network model that not only
captures the semantics of the review text but also user information that expresses that sentiment. A
user was represented as a continuous matrix and the product of word matrix and user matrix was
used for sentiment classification using a neural network.

3
Tang et al. [11] further used a neural network model to embed both user and product information
into a vector representation of a document. They first converted a document text to continuous
vector representation using a convolutional neural network (CNN). Each user and product were
represented as a continuous matrix and the concatenation of user-text and product-text was fed to
a CNN for sentiment analysis. IMDB, Yelp 2014, and Yelp 2013 datasets were used for sentiment
analysis using their model. They were able to achieve state-of-the-art results with all three datasets.
Chen et al. [12] used a similar one-layer convolutional neural network (CNN) to learn review
embeddings of IMDB and Yelp datasets. Their convolutional neural network took reviews of
varying length and produced 300-dimensional vectors. To handle varying lengths of reviews, they
padded shorter reviews with zero vectors. Filters of width 3 and 5 moved on the word embeddings
to perform one-dimensional convolution and produced multiple feature maps. Only useful features
were captured by using max-overtime pooling in pooling layer. The output of multiple filters was
then concatenated to create a 300-dimensional vector. Softmax function was used as an activation
function to train the network over K-classes. They then used a Recurrent Neural Network (RNN)
to learn the temporal information as well as to capture both product and user information and
reported state-of-the-art results on IMDB and Yelp datasets. Their work is based on the argument
that a product that receives positive reviews initially is likely to get more positive reviews in the
future and vice-versa.
Motivated by the results of Chen et al.’s model [13], along with syntactic and semantic
relationships, our model also learned product and temporal relations of reviews. We believe that
these are important features and can significantly improve sentiment classification accuracy. Our
work is based on the argument that a popular product might receive more number of higher ratings
and a less popular product might receive higher number of lower ratings. We used recurrent neural
networks (RNN) to capture such information as they are great in representing sequences.
2. METHODOLOGY
In order to analyse the sentiment of Amazon.com reviews we built a model using recurrent neural
networks (RNN) with gated recurrent unit (GRU) that learned low-dimensional vector
representation of reviews using paragraph vectors and product embeddings.
We first converted Amazon.com product reviews to fixed-length feature vectors using paragraph
vectors. These feature vectors were then grouped by product and sorted in temporal order. Each
group was used to train an RNN with GRU. The vectors generated in the penultimate layer of the
RNN are called product embeddings. These embeddings capture important information like product
qualities and temporal relations among reviews. We then concatenated product embeddings with
fixed-length vectors generated by paragraph vectors and trained a support vector machine. We also
developed a user interface to tackle the review-rating mismatches. There are situations where a user
may write a highly positive review but give it 1 or 2 stars or write a highly negative review but give
it 4 or 5 stars [13]. Although such cases are rare, they create confusion among users who read these
reviews. To address this issue, we labeled reviews with rating 1 and 2 are as ‘negative’, reviews
with rating 3 as ‘neutral’, and reviews with rating 4 and 5 as ‘positive’. We then developed a web
service application that uses our classifier to predict a class from the user review. If the predicted
class and the class that the submitted rating belongs to are different a warning is submitted to the
user so that they can review and correct their rating. Figure 1 illustrates our overall approach.

4
Figure 1 Sentiment analysis of Amazon.com reviews and ratings
2.1. DATA AND DATA PRE-PROCESSING
The data used in this study is a set of approximately 3.5 million product reviews collected from
Amazon.com by Fang et al. [14]. Each review includes information on rating, product id,
helpfulness, reviewer id, review title, review time, and review text. The rating is based on a 5-star
scale. Figure 2 displays a review in the dataset.

5
Figure 2 Example Amazon.com review
As shown in Figure 3, there are 2.2 million 5-star reviews; 622,308 4-star reviews; 265,684 3-star
reviews; 171,153 2-star reviews; and 288,789 1-star reviews.
Figure 3 Distribution of reviews by rating
Online reviews contain a lot of noise like hyperlinks, html tags, informal words, etc. and many
words don’t have any significant impact on the sentiment of the review. Keeping such words in the
review text will increase the dimensionality of the problem. To address this issue, we pre-processed
each review before converting it to its corresponding vector. Our step-by-step approach included:

6
1. Removing hyperlinks.
2. Removing unwanted spaces between words.
3. Converting informal words such as ‘I’ll’, ‘I’ve’ to its formal form ‘I will’, ‘I have’, etc.
4. Adding spaces between punctuation. For example, ‘This is great!It works.’ is converted to
‘This is great ! It works .’. Punctuations are treated as separate tokens to try to improve the
accuracy of the classifier.
2.2. MODELING OF REVIEWS USING PARAGRAPH VECTORS
Paragraph vectors (PV) [2] are highly inspired from the framework of word vectors [4]. Word
vectors framework learns semantic relationships by predicting the next word from words in a given
context. Similarly, PV framework learns vectors by predicting next word given many contexts that
are sampled from a paragraph. There are two flavors of paragraph vector, a) Distributed memory
model (PV-DM) b) Distributed bag of words model (PV-DBOW). The difference between the two
models is that PV-DM captures word order while PV-DBOW ignores it.
Figure 4 illustrates the distributed memory model of the paragraph vector [2]. Every paragraph is
mapped to a column of matrix D and similarly every word is mapped to a column of matrix W.
Given a context sampled from a paragraph or document (for example xi, xi+1 and xi+2), the model
predicts the next word xi+3 by concatenating the paragraph vector with vectors of words in that
context. The model updates both the paragraph matrix and the word matrix while training to
minimize error.
Figure 4 Distributed Memory Model of Paragraph Vector
Figure 5 illustrates the distributed bag of words model of paragraph vector [2]. This is different
from PV-DM as it doesn’t consider word ordering. PV-DBOW samples a random context from a
paragraph and then a random word from that context, then based on that word it tries to predict the
context. It doesn’t use a word matrix and hence requires less data to store. According to the

7
experiments carried out in [2], PV-DM is shown to be more efficient than PV-DBOW, therefore
we have applied PV-DM to convert the reviews in our dataset to fixed length vectors.
Figure 5 Distributed Bag of Words Model of Paragraph Vector
2.3. LEARNING PRODUCT EMBEDDINGS USING RECURRENT NEURAL NETWORK WITH
GATED RECURRENT UNIT
By using paragraph vectors, we converted 3.5 million product reviews to 300 dimensional fixed-
length vectors. To compute product embeddings, we grouped the reviews by ‘product id’ and then
ordered them by their ‘review time’. Table 1 shows how a particular review is ordered for
computing product embedding for that review. Review text is substituted by their respective
paragraph vector. This temporally sorted vectors are the input sequence and their corresponding
ratings are the targeted output and used to train a gated recurrent unit (GRU) to learn embeddings
for that particular product.
Table 1 Review sequence for a product
Product Id Review Review Time
0060245867 This is the first of several of this type and clearly
the best of the group…
2002-08-21
0060245867 I loved the movies, however, I wasn’t interested
in merely “reading the movies” again. There was
no need...
2006-07-01
0060245867 “If You Give a Mouse a Cookie” really has been
the “It” book for some time. Parents love this
story, …
2006-09-05

8
2.3.1. RECURRENT NEURAL NETWORKS
Traditional neural networks consider inputs to be independent of each other. But this approach is
flawed for tasks such as predicting the next word in a sentence. Here, the next word depends upon
the previous word or sequence of previous words. In such cases, RNNs are useful as they are great
at capturing sequential information. RNNs have loops in their architecture that allow them to pass
information they have collected from previous inputs while processing new input. Figure 6 shows
a recurrent neural network unrolled through all time steps.
Figure 6 Unrolled RNN
In this architecture Xt is input at time step t and Yt is output at the same time step. U, V and W are
weight matrices that need to be learned and St is the hidden state that is computed at every time
step t. St is also known as the memory of the network as it stores features about the input sequences
based on its observation of previous inputs. For a given input sequence x1, x2, x3, …, xT, RNN
computes sequence of hidden states and outputs using the Algorithm 1 shown below.
bh and bo are bias vectors applied at hidden and output layers respectively. f and g are nonlinear
functions.
Recurrent neural networks are trained by gradient descent of error using an extension of back-
propagation called back propagation through time (BPTT). Algorithm 2 shown below is used for
computing gradients using BPTT.

9
2.3.2. RECURRENT NEURAL NETWORK WITH GATED FEEDBACK UNIT
When recurrent neural networks (RNN) are trained, the weight matrices U, V, and W are updated
using backpropagation through time (BPTT). Update to each weight matrix is proportional to the
gradient of the error with respect to that matrix and BPTT computes the gradients using the chain
rule. When RNN is learning long-term context, depending on the activation functions, the gradient
tends to get smaller (vanishing gradient problem) or bigger (exploding gradient problem) towards
the earlier layers, which makes it difficult to train the network [15]. Gated Recurrent Units (GRU)
were used to combat vanishing gradient problem through gating mechanism.
As shown in Table 1, for every product, the review vectors are sorted based on the review’s posted
time. These sorted review vectors along with their corresponding ratings were used as the training
sequence for that particular product. These sequences were fed to a GRU to learn the 128-
dimensional feature vector representation of product information. Training of GRU is performed
using Algorithm 3 shown below.
As shown in Algorithm 3, the hidden states are reset after the start of a new product sequence since
the objective is to capture information that exists between reviews belonging to a unique product
sequence.
Output vector at the last time step of every product sequence is considered to be the embedding
sequence for that particular product. We trained our GRU using 0.25 as dropout rate; Adam as
stochastic optimization method; categorical cross entropy as loss function; time distributed dense
layer; and 128 hidden units. During the training phase, product embedding sequence for each
unique product was retrieved and stored in a file for efficient retrieval later on.

10
2.3.3. SENTIMENT CLASSIFICATION USING SUPPORT VECTOR MACHINES
Support vector machine (SVM) is a traditional machine learning algorithm to classify both linear
and nonlinear data. Given a training data with binary outputs, support vector machine tries to find
a hyper-plane as the decision surface such that the separation between positive and negative
samples is maximized. If the data is non-linear, SVM transforms the data into a higher dimension
and then solves the problem by finding a linear hyper-plane. The kernel used for such nonlinear
data is called Gaussian Radial Basis function (RBF). SVM was used to classify each review to a
label of “positive”, “neutral”, or “negative”.
2.3.4. WEB SERVICE TO DETECT USER RATING MISMATCH
Figure 7 illustrates the web service we developed to prevent inconsistent review and rating pairs. It
uses the support vector machine model trained with 3.5 million reviews and their product
embeddings to predict the sentiment class for a given review.
Figure 7 Web service architecture
When a user writes a review for a product and assigns a rating, if the predicted class and the rating
class doesn’t match, a feedback is provided to user so that if they wish, they can update their rating.
The web service takes three inputs: ‘review text’, ‘assigned rating’, and ‘product id.’ The 300-
dimensional embedding for the ‘review text’ is predicted by the previously-trained paragraph vector
model. As discussed earlier, product embeddings for all products are saved in a file. Embedding
for the given ‘product id’ is retrieved by the web service from the file and concatenated with review
embedding to form the final 428-dimensional vector. This vector is then used by our trained support
vector machine model to predict a sentiment class. If there is mismatch between the predicted class
and the sentiment class the assigned rating falls under, a feedback is given to the user.

11
Figure 8 shows a review with a positive sentiment but with a negative rating. There is a review-
rating mismatch, therefore our system returns a message as shown in Figure 9. The user can then
choose to either update the rating or keep it.
Figure 8 User interface for reviews
Figure 9 Feedback to the user

12
3. EXPERIMENTS AND RESULTS
3.1. SENTIMENT CLASSIFICATION USING REVIEW EMBEDDING
We first designed a paragraph vector only model using Genism framework. 3.5 million
Amazon.com product reviews were converted to 300-dimensional fixed length vector. We trained
this model for 15 epochs with learning rate of 0.025, window size of 10, and 35 worker threads.
The learning rate was decreased by 0.002 after each epoch. This vector with its corresponding rating
was used to train a support vector machine (SVM). 10-fold cross-validation method was used to
evaluate the performance of this SVM. To evaluate the performance of our model, we computed
precision, recall, and classification accuracy. Table 2 lists these metrics at each iteration as well as
their average over 10 iterations. With only review embedding, the model had an average precision
score of 0.5861, recall score of 0.4085, and classification accuracy of 81.29%.
Table 2 Precision, recall, and accuracy of the model with paragraph vectors only
Iteration Precision Recall Accuracy
1 0.5890 0.4051 0.8125
2 0.5693 0.4072 0.8115
3 0.5817 0.4092 0.8128
4 0.6004 0.4091 0.8127
5 0.6028 0.4086 0.8135
6 0.5819 0.4088 0.8138
7 0.5731 0.4077 0.8139
8 0.5859 0.4097 0.8120
9 0.5841 0.4081 0.8123
10 0.5926 0.4115 0.8144
Average 0.5861 0.4085 0.8129
3.1. SENTIMENT CLASSIFICATION USING REVIEW EMBEDDING AND PRODUCT
EMBEDDING
After generating the 300-dimensional embeddings for the product reviews, we grouped them by
their ‘product id’ and sorted them by their ‘review time.’ Each group was an input to the gated
recurrent unit. The embedding generated at the last time step of the sequence was treated as the
embedding for that product. This product embedding was concatenated with the review embedding
and used to train a support vector machine (SVM). 10-fold cross-validation was used to evaluate
this model. We computed precision, recall and classification accuracy as we did with paragraph
vector only approach. Table 3 lists these metrics at each iteration as well as their average over 10
iterations. With review embedding and product embedding, the model had an average precision
score of 0.5945, recall score of 0.4252, and classification accuracy of 81.82%.

13
Table 3 Precision, recall, and accuracy of the model with paragraph vectors and GRU
Iteration Precision Recall Accuracy
1 0.5987 0.4183 0.8168
2 0.5940 0.4314 0.8189
3 0.5915 0.4265 0.8185
4 0.5962 0.4252 0.8186
5 0.5844 0.4310 0.8179
6 0.5853 0.4231 0.8186
7 0.5951 0.4249 0.8180
8 0.6016 0.4301 0.8182
9 0.5925 0.4206 0.8179
10 0.6064 0.4211 0.8182
Average 0.5945 0.4252 0.8182
Classification using both review and product embeddings gave us an increase in classification
accuracy from 81.29% to 81.82%, increase in precision from 0.5860 to 0.5945 and increase in recall
from 0.408 to 0.4252. The performance of both the approach is compared in Figure 10.
Figure 10 Comparison of metrics computed by review embedding only approach and both review
embedding and product embedding approach

14
4. CONCLUSIONS
In this research, we propose a deep learning approach to solve a prevailing problem in e-commerce
websites where the review submitted by the user doesn’t match its rating. Reviews are a very
important source of information for a potential customer before deciding to purchase a product [11].
We believe that such reviews with mismatched ratings can create negative user experience and
hence the motivation for this research.
To construct our model, we first employed paragraph vectors to learn the syntactic and semantic
relationship of a ‘review text’. We further grouped and sorted review embedding to form a product
sequence which is fed to a gated recurrent unit (GRU) to learn product embedding. The
concatenation of review embedding generated from paragraph vectors and product embedding
generated from GRU is used to train a support vector machine (SVM) for sentiment classification.
With only review embedding our classifier performs at an accuracy of 81.29%. Inclusion of product
embedding increases the accuracy to 81.82%. This shows that product information is a powerful
feature that can be employed in sentiment analysis. We later use this classifier through a web service
to predict rating of a review and compare it against given rating. This web service takes ‘review
text’ and ‘review rating’ and provides a warning to the reviewer if there is an inconsistency between
the given rating and review.
In summary, we showed how review text and product information together can be used to build a
robust deep learning model that can be employed in sentiment analysis. We believe that a similar
technique can be used to learn user information. Some users are more lenient and can give much
higher ratings for the same review text compared to critical users [11]. We are motivated to exploit
this information in future.
REFERENCES
[1] Kumar, J. A. & Abirami, S. (2015) “An experimental study of feature extraction Techniques in opinion
mining”, International Journal on Soft Computing, Artificial Intelligence and Applications (IJSCAI),
4(1): 15-21.
[2] Le, Q. V. & Mikolov, T. (2014) “Distributed representations of sentences and documents”, Proceedings
of the 31st International Conference on Machine Learning, Beijing, China, 2014.
[3] Mikolov, T., Chen, K., Corrado, G., & Dean, J. (2013) “Efficient estimation of word representations in
vector space”, Proceedings of International Conference on Learning Representations Workshop, 2013.
[4] Mikolov, T., Sutskever, I., Chen, K., Corrado, G. S., & Dean, J. (2013) “Distributed representations of
words and phrases and their compositionality”, Advances in Neural Information Processing Systems,
2013.
[5] word2vec https://code.google.com/archive/p/word2vec/
[6] Socher, R., Perelygin, A., Wu, J. Y., Chuang, J., Manning, C. D., Ng, A. Y., Potts, C. (2013) “Recursive
deep models for semantic compositionality over a sentiment treebank”, Proceedings of the Conference
on Empirical Methods in Natural Language Processing, 2013.
[7] Kim, Y. (2014) “Convolutional neural networks for sentence classification”, Proceedings of the 2014
Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 1746–1751, October
25-29, 2014, Doha, Qatar.

15
[8] Johnson, R. & Zhang, T. (2015) “Effective use of word order for text categorization with convolutional
neural networks”, Proceedings of Human Language Technologies: The 2015 Annual Conference of the
North American Chapter of the ACL, pp 103–112, Denver, Colorado, May 31 – June 5, 2015.
[9] Ouyang, X. Zhou, P., Li, C. H., & Liu, L. (2015) “Sentiment analysis using convolutional neural
network. Proceedings of IEEE International Conference on Computer and Information Technology;
Ubiquitous Computing and Communications; Dependable, Autonomic and Secure Computing;
Pervasive Intelligence and Computing, pp. 2359-2364, October 2015.
[10] Tang, D., Qin, B., Liu, T. & Yang, Y. (2015) “User modeling with neural network for review rating
prediction”, Proceedings of International Joint Conferences on Artificial Intelligence, pp. 1340-1346,
2015.
[11] Tang, D., Qin, B., & Liu, T. (2015) Learning semantic representations of users and products for
document level sentiment classification. Proceedings of the 53rd Annual Meeting of the Association for
Computational Linguistics and the 7th International Joint Conference on Natural Language Processing,
pp. 1014–1023, Beijing, China, July 26-31, 2015.
[12] Chen, T., Xu, R., He, Y., Xia, Y., & Wang. X. (2016) “Learning user and product distributed
representations using a sequence model for sentiment analysis”, IEEE Computational Intelligence
Magazine, 11(3):34-44, 2016.
[13] Sharma, K. & Lin, K. (2013) “Review spam detector with rating consistency check”, Proceedings of
the 51st ACM Southeast Conference, ACM, 2013.
[14] Fang, X. & Zhan. J (2015) “Sentiment analysis using product review data”, Journal of Big Data, 2(1):5,
2015.
[15] Bengio, Y., Simard, P. & Frasconi. P. (1994) “Learning long-term dependencies with gradient descent
is difficult”. IEEE transactions on neural networks, 5(2):157-166, 1994.

DEEP LEARNING SENTIMENT ANALYSIS OF AMAZON.COM REVIEWS AND RATINGS

More Related Content

What's hot (17)

Similar to DEEP LEARNING SENTIMENT ANALYSIS OF AMAZON.COM REVIEWS AND RATINGS (20)

Recently uploaded (20)

DEEP LEARNING SENTIMENT ANALYSIS OF AMAZON.COM REVIEWS AND RATINGS