This paper describes the system submitted to "Sentiment Analysis at SEPLN (TASS)-2019" shared task. The task includes sentiment analysis of Spanish tweets, where the tweets are in di erent dialects spoken in Spain, Peru, Costa Rica, Uruguay and Mexico. The tweets are short (up to 240 characters) and the language is informal, i.e., it contains misspellings, emojis, onomatopeias etc. Sentiment analysis includes classi cation of the tweets into 4 classes, viz., Positive, Negative, Neutral and None. For preparing the proposed system, we use Deep Learning networks like LSTMs.
CuDNNLSentiment Analysis (SA) refers to the use of Natural Language Processing (NLP) to systematically identify, extract, quantify, and study a ective states and subjective information. The Sentiment Analysis at SEPLN (TASS)-2019 3 was a classi cation task where it was required to classify a Spanish tweet on basis of its sentiment,into various classes like, Positive, Negative, Neutral and None. It was further divided into two subtasks;the rst subtask being monolingual testing of system while the second task being cross-lingual testing of system. However, the task threw some additional challenges. The given tweets involved lack of context, where the number of words were less than 240. Moreover, the tweets were in an informal language and contained multi-linguality. Also, the classi cation system that would be prepared for the task, needed to be generalized for various test corpora as well.
To solve the task in hand, we built a bidirectional Long Short Term Memory (LSTM) based neural network, for prediction of the sentiments present in the provided dataset. For both the subtasks, our system categorized the instances into P, N, NEU and NONE.
The rest of the paper has been organized as follows. Section 2 describes the data, on which, the task was performed. The methodology followed is described in Section 3. This is followed by the results and concluding remarks in Section 4 and 5 respectively. 2
The dataset that was used to train the model was provided by InterTASS[
The dataset originally comprised of Spanish tweets of various dialects, namely ES(Spain), PE(Peru), CR(Costa Rica), UR(Uruguay) and MX(Mexico). The tweets were also tagged with their respective sentiments. We merged all this data and shu ed them. The resulting dataset had 7,265 sentiment tagged tweets, which were splitted into 5,086 instances of training data and 2,179 instances of development data. Our approach was to convert the tweets into a sequence of words and convert them into word embeddings. We then run a neural-network based algorithm on the processed tweet. Language and label based categorical division of data is given in Table 2, 3, 4 and 5.
The provided training and development data were merged and shu ed to create a bigger training set, and we refer to the same as training data in the methodology section. 3
The rst stage in our model was to preprocess the tweets. For the preprocessing
steps, we took inspiration from the work on Hate Speech against immigrants in
Twitter[
The last step (step 5) consists of taking advantage of the Pascal Casing of hashtags (e.g. #TheWallStreet). A simple regex can extract all words; we ignore a few errors that arise in this procedure. This extraction results in better performance mainly because words in hashtags, to some extent, may convey sentiments of hate. They play an important role during the model-training stage.
We treat each tweet as a sequence of words with interdependence among
various words contributing to its meaning. We convert the tweets into one-hot
vectors. We also include certain manually extracted features listed below:
1. Counts of words with positive sentiment, negative sentiment and neutral
sentiment in Spanish
2. Counts of words with positive sentiment, negative sentiment and neutral
sentiment in English
3. Subjectivity score of the tweet
4. Number of question marks,Exclamations and full-stops in the tweet
For this, we used SenticNet5[
The use of BiLSTM networks is a key factor in our model. The work of [
We use a bidirectional LSTM based approach to capture information from both the past and future context followed by an Attention layer consisting of initializers and regularizers.
Our model is a neural-network based model. Initially, the manual feature vectors are appended with the feature vector obtained after converting the processed tweet to one-hot encoding. It is then passed through an embedding layer which transforms the tweet into a 128 length vector. The embedding layer learns the word embeddings from the input tweets. We pass the embeddings through a Batch Normalization layer of dimensions 10 X 128. This is followed by one bidirectional LSTM layer containing 128 units with its dropout and regular dropout set to 0.4 and activation being a sigmoid activation. This is followed by a Bidirectional CuDNNLSTM layer with 64 units for better GPU usage. This is followed by the nal output layer of neurons with softmax activation, where, each neuron predicts a label as present in the dataset.
For both subtasks 1 and 2, we train a model containing 4 neurons for predicting P(0/1), N(0/1) and NEU(0/1) and NONE(0/1) respectively. After the CuDNNLSTM layer we have added a regularizing layer which is initialized with glorot uniform initializer. This layer has both W regularizers and b regularizers to prevent the model from over tting. This provides better validation and test results by generalizing the feature learning process. The model is compiled using the Adam optimization algorithm with a learning rate of 0.0005. Categoricalcrossentropy is used as the loss function. The working is depicted in Figure 1.
We note that the dataset is highly skewed in nature. If trained on the entire Data
Concatenate
Embedding Layer
Batch Normalization Bidirectional LSTM Bidirectional CuDNN
LSTM Output
Dense Layer
Regularizer training dataset without any validation, the model tends to completely over t to the class with higher frequency as it leads to a higher accuracy score.
To overcome this problem, we took some measures. Firstly, the training data was split into two parts | one for training and one for validation comprising 70 % and 30 % of the dataset respectively. The training was stopped when two consecutive epochs increased the measured loss function value for the validation set.
Secondly, class weights were assigned to the di erent classes present in the data. The weights were approximately chosen to be proportional to the inverse of the respective frequencies of the classes. Intuitively, the model now gives equal weight to the skewed classes and this penalizes tendencies to over t to the data. 4
We participated in subtasks 1 and 2 of Sentiment Analysis at SEPLN (TASS)2019 and our system ranks rst among the competing participants.
We have included the automatically generated tables with our results. The results(rounded o to 3 decimal places) are depicted in Table 6, 7 and 8.
One-Hot Encoding
Blob Score, Punctuation count, No. of +ve words, No. of -ve words, No. of neutral words
In this system report, we have presented a model which performs satisfactorily in the given tasks. The model is based on a simple architecture. There is scope for improvement by including more manually extracted features (like those removed in the preprocessing step) to increase the performance. Another fact is that the model is a constrained system, which may lead to poor results based on the modest size of the data. Related domain knowledge may be exploited to obtain better results. Use of regularizers led to proper generalization of model, henceforth increasing our task submission score.