Automatic Star Rating for Movie Reviews Part 3
In the previous part of this series, I created a Bag Of Words model to classify the reviews. In this part, I will present two other models I have implemented, a Recurrent and Convolutional model.
The data pre-processing used in the previous part of this series was applying tokenization to the movie reviews, removing the common Portuguese stop words and turning each token into a word-id that references a row in the word embedding matrix (This row represents the original token). This was to feed the movie reviews into the Bag Of Words model.
For both new models, I adopted the same preprocessing approach as the Bag Of Words model, except that the stop words are not removed from the reviews.
The Recurrent Model
A Recurrent Neural Network is an specialized model used, mostly, to handle text data. That is because this model uses as input not only a word to process, but also the last state of the model, that was produced by all the other words before the current word that is being fed to the model. This can be better visualized in the following picture:
Figure 1: Representation of a Recurrent Neural Network. Take a look the second timestep. We can see that the network will take as input not only x(2), but also h(1), which was generated based on all the past x inputs in the network
Therefore, this model will read the whole movie review, token by token. At the end, it will have generated a final vector representation of the text. Afterwards, this vector is then used to classify the review into the 5 possible star ratings.
I employed the LSTM cell with 128 units to process the hidden step of this network (the vector represetation of the review). Since I have a small dataset, I applied dropout to the embeddings matrix and the hidden states of the LSTM. The dropout mask must be the same along all the timesteps of the LSTM calculation.
Finally, I have also added L2 regularization to the weights of my model, as another way to prevent overfitting. However, this was not enough, since the best accuracy for the recurrent model was 30% using the Full dataset. The overfitting result can be observed in the following confusion matrix:
We can see that the features the model learned to classify 3 and 4-star reviews do not generalize well, the other classes (1, 2 and 5) are also mostly classified as 3 and 4-star reviews.
We can notice that the model’s learned features do not generalise well. The other classes (1, 2 and 5) are also mostly ranked as 3 and 4-star reviews.
Another problem with this model is the average size of the reviews, as seen in the first part of this series. The review size creates a heavy burden on the LSTM memory cell since it must now handle long term dependencies in the text.
The Convolutional Model
The convolutional model I have created is based on the work of Kim, Y., which can be best visualized in the following picture:
Representation of the Convolutional Model for sentence classification. The model applies a convolutional filter over the sentence, where each filter size represents a different type of n-gram. The information of the collected n-grams is summarized by a pooling layer. Afterward, the summarized vector is fed into a neural network that generates the classification. Image from 
The idea of this model is to apply a 1-D convolutional filter over the word embeddings, where each filter corresponds to a different n-gram filter. For example, if we a have a size 2 filter, we would apply it to two consecutive words and generate a new value; If we have 20 size 2 filters, we would be computing 20 different computations for bigrams. One filter could look at two consecutive positive words, while a separate filter might look at a negation word followed by an adjective. Therefore, it is useful to have different filter size to distinct type of n-grams(trigrams, 4-grams, …).
In my model I have created 32 different filters for each distinct size (3, 4 and 5 in my model). After applying the convolutions, we use the pooling layer summarize the information provided by the convolutions. This will generate a vector representation of the sentence that will used to fed a fully connected network, which is responsible for creating the predictions.
The best accuracy achieved by this model is 34% using the Full dataset. This result can be best visualized through the confusion matrix:
We can see that the model suffers from the same problems as the ones presented by the Recurrent Model, where it overfits for 3 and 4-star reviews. This problem is present in the model trained with the Full (underflow) dataset as well.
I believe that the filters learned by the model do not generalize because of the small dataset I have, allowing the filters to overfit for the traning data.
Single Source Datasets
The results for single source datasets resembles the results of both models using the Full dataset. For example, looking at the confusion matrix produced by the Convolutional model for the Cineclick dataset, we can see that the model clearly overfits.
Back to Bag Of Words
After a close look at the results, the Bag Of Words model obtained the best result, as can be seen by the last past post of this series. I tried to improve this model by restricting the vocabulary for words that appeared in the training data more than 10 times ( A behavior already implemented in both Recurrent and Convolutional models). However, this approach did not provide any improvements to the model.
I also tried to apply the oversampling technique to the datasets used for the Bag Of Words model, to handle ratings with fewer reviews; but again, no improvement to the model was perceived. Instead, the model was worse at generalizing to the validation dataset of all available datasets.
There are other approaches which I could try to use to make improvements in the models. I could create a language model based on movie reviews from different sources (e.g., reviews which do not have a rating) and try to check if the model generates a “sentiment unit” as has happened with OpenAI. However, I still believe that I will not have enough data to do that.
Another approach I could try is to add sentiment information to the word embeddings. Since word embeddings are generated with the idea to capture semantic information, there is no guarantee that they carry sentiment information already. Although I believe this would be beneficial to the Bag Of Words model, I don’t think that the model accuracy would radically improve. That’s because the dataset used is too small and there are few examples of each classification type. Furthermore, the reviews are complex text structures, making finding patterns that generalize a difficult task.
In conclusion, I realized that I underestimated this task. I did not think that the collected dataset would be that much problem as my models have proven me. However, this was still a great experience to have. I have learned a lot not only about Natural Language Processing; but also about Tensorflow and different models for text classification. I have confidence that I will have better judgement on my next NLP task and also be faster to implement new solutions.
(If you want to better understand how I implemented the models and how I generated the results presented in this article, please take a look at the Github page for this project.)