What people think about fast food: opinions analysis and LDA modeling on fast food restaurants using unstructured tweets

Challenges in sentiment analysis

Sentiment analysis is a robust and efficient NLP technique that can analyze sentiments in a variety of data, including the publicly expressed views and comments of people on social networking platforms. Restaurants can substantially benefit from the automatic extraction of sentiments from large corpora of data (such as sentiments regarding fast food companies) in order to better serve the clients and their recommendations about the food. However several challenges require further research efforts. For example, labeling unstructured tweets using lexicon-based approaches still requires research. Although data cleaning helps to obtain better results, however, it also may cause information loss which is not very well investigated. Lexicon based-methods require extensive preprocessing of tweets to produce more accurate results and the limited availability of sentiment vocabulary across several domains makes it challenging. Similarly, sentiment analysis regarding food is not explored very well. Also, existing methods provide lower accuracy. The proposed method deals with low accuracy, appropriate feature extraction, and multiclass classification problem and provides better results. The BiLSTM+ GRU architecture outperforms single model gated recurrent units (GRUs) and makes the model more robust. The proposed method can reduce ‘variance’, and model bias, and so decreases the possibility of overfitting.

Contributions

This research aims to investigate how people generally feel about various fast food restaurants, their quality of service, foods, etc., so that more informed, quick, and intelligent decisions can be made by other users. Also, the provision or exclusion of specific food items can be made from restaurants in the light of user reviews sentiments. This will help restaurant owners gain the public’s trust in their restaurants as well as for customers to choose the best restaurant for enjoying the meal. Existing studies predominantly focus on the use of machine learning models for sentiment classification. Accordingly, the performance evaluation of deep learning models in general and deep ensemble models, in particular, are under-explored research areas. This study overcomes these limitations and makes the following contributions.

• A large dataset of unstructured tweets is gathered from Twitter using the Tweepy API regarding the top five fast food companies including McDonald’s, KFC burger, Subway, Burger King, and Pizza Hut. Preprocessing is carried out before applying deep ensemble models to the dataset. For this purpose, NLP is used for preprocessing, data cleaning, and labeling.

• Performance of lexicon-based approaches is evaluated including TextBlob and valency-aware dictionary for sentiment reasoning (VADER). This study makes use of deep learning models including recurrent neural network (RNN), long short-term memory (LSTM), Bidirectional LSTM (BiLSTM), and gated recurrent unit (GRU). It also deploys four customized deep ensemble models including BiLSTM+GRU, LSTM+GRU, GRU+RNN, and BiLSTM+RNN.

• Topic modeling is employed to acquire the most frequently discussed topics and words using the Latent Dirichlet Allocation (LDA) model.

The remaining of this article is divided into four parts. The following section presents the related work. The proposed approach is discussed in the ‘Material and Methods’ section which also includes a brief overview of deep ensemble models and the dataset used for experiments. Experimental results are discussed in the ‘Results and Discussions’ section. Finally, the conclusion is presented.

Dataset description and preprocessing

We used different hashtags for data collection like #McDonald’s, #KingBurger, #PizzaHut, #Subway, #Subway-sandwich, #KFCburger, etc. to collect a large number of tweets from Twitter using the Tweepy Python library. We collected 14,008 tweets related to McDonald’s, 12,000 related to KFC, 10,861 tweets about Pizza Hut, 14,002 tweets about BurgerKing, and 4,105 tweets about Subway restaurants. The tweets contain ‘user_name’, ‘date’, ‘user_location’, ‘friends’, and ‘text’. The text comprises the tweet and a clickable link to the relevant tweet on Twitter. Our study focused solely on the text variable, with no consideration given to the user or geographic location. The collected raw tweets contain some unnecessary and redundant information and are unlabeled. We used preprocessing steps to filter out irrelevant and redundant data from the tweets (Alzahrani et al., 2022; García, Luengo & Herrera, 2015). Preprocessing is important for efficient training of the models and precise results. First of all, all the tweets are converted into lowercase because the machine takes ‘food’ and ‘Food’ as separate words. Secondly, stopwords are removed from the tweets, and it is most important to remove useless information for improving the performance of models. Stopwords, for example, ‘is’, ‘of’, ‘this’, ‘to’, ‘that’, ‘be’, etc., provide no information about the topic. Also, punctuation like (,? # & []! *) is removed from tweets. However, this does not affect the performance of models. Following that, uniform resource locator (URL) links like ‘https://www.csthi.com’, duplicate text, null values, and numbers are removed from the tweets. In the stemming process, the words are converted into their base form, such as ‘walk’, ‘walked’, or ‘walking’ is transformed into ‘walk’. Details of tweets collected from Twitter are presented in Table 2.

TextBlob and VADER

TextBlob (Chandrasekaran & Hemanth, 2022) and VADER (Endsuy, 2021) are two of the most widely used lexicon-based approaches that include predefined dictionaries or rules. It makes NLP tasks simple for textual data. Polarity and subjectivity are two outputs received from a sentence when given to TextBlob. Polarity gives two sentiments; −1 for negative sentiment and +1 for positive sentiment. Subjectivity refers to subjects or judgments. TextBlob calculates the sentiments using the Algorithm #1:

VADER is used for both emotions and positive or negative polarities. It is very smart in NPL tasks, such as it does not require any processing like stopwords, stemming, etc. VADER calculates the sentiments using the Algorithm #2:

Ensemble of deep learning-based models

Deep learning is capable of automatically selecting the best features for a variety of tasks. Deep learning is more scalable than machine learning and requires less human intervention. Deep learning models solve classification problems from beginning to end, whereas machine learning needs to split the problems and solve them separately. Deep learning performs better when the dataset size is large. For the current study, the dataset is large and we expect better results using deep learning models, especially deep ensemble learning. In deep learning, overfitting issues can be solved by utilizing dropout layers, batch normalization, regularizers, and removing layers from the model to reduce its complexity. Several studies (Dhola & Saradva, 2021; Rupapara et al., 2021; Onan, 2021) used deep learning for sentiment analysis and showed better accuracy than machine learning models.

Ensemble learning is a useful method for combining the information from more than one conceding model to make the final prediction. It has been argued that ensemble learning would be an effective solution to the problem of overwhelming variance in single prediction models and might help decrease the generalization error (Rokach, 2009). The process of ensemble learning involves the combination of several categorization models rather than the use of a single model to solve a problem. By integrating the layers of several models, the overall performance of these models can become more effective. The four ensemble models are briefly discussed in the subsections below.

BiDirectional Long Short-Term Memory + Gated Recurrent Unit (BILSTM+GRU)

The BiLSTM process sequentially creates neural network information in two directions: past-to-future and future-to-past (Xu et al., 2019). Because of two input directions, a BiLSTM model is different from an LSTM model. We used 8 layers in the ensemble of BiLSTM with GRU. The embedding layer embeds the 6500 ×120 units in the first layer. One dropout of 0.5 is added after the GRU layer to prevent it from over-fitting initially. A GRU model layer with 128 units has been added as the 4th layer. The first dense layer is 64 units, and the second is three units with a softmax activation layer. Our dataset contains three classes, and we used categorical cross entropy as a loss function with the ‘Adam’ optimizer. The architecture of the proposed BILSTM+RU model is depicted in Fig. 2.

Long Short-Term Memory + Gated Recurrent Unit (LSTM+GRU)

The GRU neural network (Raza, Hussain & Merigó, 2021) simulates the operation of the LSTM. LSTM contains three gates: input, output, and forget, whereas GRU has only the update and reset gates. GRU uses less power than LSTM while having a smaller number of gates. We employed a 7-layer model for the LSTM + GRU ensemble. The 6500 ×120 units in the first layer are embedded in the embedding layer. In this case, 100 and 64 units are used to activate the LSTM layer. There is now a 128-unit GRU model layer available. 32 units in the first dense layer are activated using the l2 regularization kernel known as ‘ReLU’ and three units in the second dense layer are activated using a softmax activation. We utilized the ’Adam’ optimizer and the categorical cross-entropy loss function to evaluate the constructed model on three classes. The architecture of the ensemble model (LSTM+RU) is depicted in Table 3.

Gated Recurrent Unit + Recurrent Neural Networks (GRU+RNN)

The RNN model is specially designed for processing sequential information (text, audio, video, etc.). The output is generated to follow the previous samples in the RNN model. It works similarly to that of the GRU, but the operations and gates are different. To solve this, we combined GRU with CNN to enhance the model results. A 7-layer model for the GRU+RNN ensemble with 6500 ×120 units in the first layer is embedded in the embedding layer. In this case, 256 units are used to activate the GRU layer and 128 for simple RNN layers. In the first dense layer, 32 units are activated by the ReLU l2 regularization kernel, and 3 units in the second dense layer are activated by a softmax activation. The architecture of BILSTM+RU is depicted in Table 3.

BiDirectional Long Short-Term Memory + Recurrent Neural Network (BILSTM + RNN)

We used an 8-layer model for the BiLSTM + RNN ensemble. The 6500 ×4120 units in the first layer are embedded in the embedding layer. 100 units are used to activate the first LSTM layer and 64 for the second. A 0.5 unit dropout is added in the 4th layer. 32 units in the first dense layer are activated using the l2 regularization kernel and 3 units in the second dense layer are activated using a softmax activation. We utilized the Adam optimizer and the categorical cross-entropy loss function to evaluate the constructed model. The architecture of the BILSTM+RU model is depicted in Table 3.

Transformer based models

Recently, NLP witnessed rapid growth led by pre-trained supervised models. Such models contain hundreds of millions of parameters and show very good performs for NLP tasks. Particularly, the BERT model has achieved an iconic position recently. These models are significant advancements over their previous models (Devlin et al., 2018). In 2018, Google’s AI team developed a model for sentiment analysis (SA), text prediction (TP), and question-and-answer tasks called BERT. The BERT attention mechanism is trained to recognize associations between words in a phrase based on their context. The encoder receives input text and the decoder outputs it, but the encoder mechanism is extremely crucial.

BERT is one of the most comprehensive models for NLP and several variants have been proposed. It has a large size and a high number of parameters which makes it effective. However, it requires large computation time which inhibits the model’s ability to operate quickly and efficiently on low-power devices. Similarly, it has reliability problems on low-powered devices. In some real-time scenarios, any accuracy gains from an improvement may be overturned if the system’s response time is too slow. To reduce the time required to execute the model’s computations, several modifications have been introduced.

The BERT model is complex and hard to train model and may take days for training depending on the size and nature of the data. On the other hand, XLNet transformer-based auto-regressive, as well as the auto-encoding model achieved state-of-the-art performance in many NLP applications. Similar to BERT, XLNet uses bidirectional text and avoids the limitations posed by masking in the BERT model. The XLNet manages to capture a greater number of dependency pairs than BERT. Both models are used in this study for experiments with the hyperparameters given in Table 4.

In this study, sentiment classification was also performed using the BERT model. We used 400 words from each text as input layers, a Keras layer with 109,482,241 parameters, two dense layers with units of 16 and 32, one for multiclass classification, and two dropout layers with units of 0.5 and 0.2.

LDA based topic modeling

Topic modeling aims at finding a group of similar words or topics which are frequently discussed in tweets. Contrary to the classification of sentiments from machine learning models which show only the type of sentiments, topic modeling provides frequent words used in the text. In the context of food-related tweets, it provides the positive and negative aspects of food companies’ menus and services which can be used by the companies to improve the quality of services.

The LDA model has become the most researched and extensively used model in many disciplines since it addresses the shortcomings of other models such as latent semantic indexing (LSI), latent semantic analysis (LSA), and probabilistic latent semantic indexing (PLSI). LDA classifies, interprets, and evaluates a large collection of textual information, exposing hidden topics in the process (Yang & Zhang, 2018). It divides the topic into different topics and divides the large documents into smaller dimensions. When we do this manually, it will take a long time, but the LDA model can extract this in a very short time and play important role in sentiment analysis. The words assigned to the topic and their distribution according to the Dirichlet distribution. The LDA model provides the best results. The LDA model is shown in Fig. 3.

We use a bag of words to extract features from the cleaned tweets, setting the number of features to 500, the maximum number of features to 50, and the stop words to English. Then, we fit countVectorizer to the clean text data containing sentiment values and apply the LDA model with 10 components and a maximum of five iterations. The LDA model then extracts relevant topics.

Evaluation metrics

The performance of the models is evaluated using different evaluation metrics (Hossin & Sulaiman, 2015). The evaluation is performed using the test data to check the performance of models according to their efficiency. Accuracy, precision, and recall parameters are used in this study. Accuracy is determined by dividing true positive (TP) plus true negative (TN) predictions over total predictions. Precision is determined by dividing TP predictions divided by TP plus false positive (FP) predictions. The recall is computed by dividing TP predictions over TP plus false negative (FN) predictions.

Sentiment analysis using lexicon-based approaches

The two most widely used lexicon-based approaches are used for sentiment analysis of all tweets like McDonald’s tweets, KFC tweets, etc. Table 5 represents the sentiment of the public towards fast food-related tweets using two lexicon-based approaches. All tweets contain mostly neutral sentiments. Tweet sentiments related to different fast food restaurants using the TextBlob approach are represented. A total of 3,253 tweets are used for positive sentiments and another 2,453 are negative sentiments about KFC. Public sentiments towards McDonald’s are 4,616 neutral, 2,952 positive, and 1,852 negative as shown in Table 5. There are 3,343 neutral tweets, 2,012 positive tweets, and 1,077 negative tweets about Pizza Hut. Following that, Burger King and Subway restaurant tweets, people have 7,048 and 1,558 neutral sentiments, 3,896 and 1,388 positive sentiments, and 2,788 and 1,026 negative sentiments, respectively. Using the VADER approach, the public shows mostly neutral sentiments about different restaurant foods, as in TextBlob sentiments. People have 2,424 sentiments against the fast food service and quality out of 11,769 towards KFC. Only 3,280 sentiments have positive emotions towards McDonald’s.

Results reveal that using the VADER, the number of positive sentiments is higher as compared to using TextBlob except for Subway food sentiments where the positive sentiments using TextBlob are higher. It can be observed that although the same tweets are used for sentiment analysis using TextBlob and VADER, the sentiments assigned by the two are substantially different.

Table 6 displays the most commonly used words identified in the tweets. Good, love, better, best, gift, start, pizza, hut, kfc, burger, happy, customer, sandwich, order, win, etc are words that are mostly used for positive sentiments, and words like bad, sorry, fuck, dead, shit, food, please, unpleasant are treated as negative sentiments in the tweets.

Results of transformer-based models

The BERT model was used by many researchers to classify the sentiments. Abdelgwad (2021) used the BERT model for sentiment classification on the hotel-review dataset, fitted the model on 10 epochs, and achieved an 89.50% accuracy by employing 10% dropout layers and 24 batch sizes. Singh, Jakhar & Pandey (2021) used the BERT model for emotion classification on Twitter data and attained 93.80% accuracy. Similarly, Pota et al. (2021) compared the performance of the BERT model with the proposed model which was trained using five epochs with an eight-batch size. Results showed that the performance of BERT is superior. In view of the reported accuracy of the BERT model, this study also employs BERT for experiments. Since training a BERT model requires a lot of processing power, most studies only use 5, 8, or 10 epochs. We trained the BERT model using 10 epochs.

The results of the BERT model using the TextBlob and VADER approaches are given in Table 7. Using TextBlob features and VADER features extracted from fast food tweets, the BERT model achieved 95.41% and 93.45% accuracy scores, respectively. TextBlob features led the BERT model to obtain 97% precision, 95% recall, and 92% F1 score on neutral tweets, 95% on positive tweets, and 92% on negative tweets. Using the VADER features, precision, recall, and F1 scores are low.

The results of the TextBlob and VADER-based XLNet model are also included in Table 7. The XLNet model obtained an accuracy of 94.12% and 93.54%, respectively, when provided with TextBlob and VADER features, respectively. Recall scores of 96% for neutral tweets and 90% for negative tweets are achieved by the XLNet model using TextBlob and VADER, respectively. The XLNet model outperforms the BERT model in terms of accuracy and precision on VADER features.

Results of ensemble deep learning models

Experiments are performed using both the standalone deep learning models, in addition to the designed ensemble models. Figure 4 shows the training and loss curves of ensemble deep learning models. At epochs 40, 41, and 46, the training accuracy is at its highest with 99.92% accuracy, while the validation accuracy is at 95.62% using the BiLSTM+GRU model. In the LSTM+GRU model, the highest training accuracy of 99.93% is at epochs 45, while the validation accuracy is 95.41%. The training accuracy and loss curves increase after epoch 20 and the highest values are obtained at epoch 41 for all the models.

The ensemble deep learning models, BilSTM+GRU, GRU+RNN, LSRM+GRU, and BiLSTM+RNN perform better as compared to individual models, which are listed in Table 8. The simple RNN model used 816,099 parameters for training the model and obtains a 93.15% accuracy with the TextBlob approach and 89.3% accuracy with the VADER approach. The other three models LSTM, BiLSTM, and GRU are trained with 912,819, 677,027, and 880,227 parameters, and they got 94.67%, 94.63%, and 94.63% accuracy, respectively with TextBlob. The ensemble BiLSTM+GRU achieved 95.62% accuracy which is better as compared to any single model.

The precision, recall, and F1 score of different standalone and ensemble models are presented in Table 9 using extracted features from the TextBlob approach. The individual model RNN achieved very low precision, recall, and F1 score on positive tweets, but 89% precision, 86% recall, and 87% F1 score on negative tweets. The GRU achieved a 94% overall score on positive tweets, with 91% precision, an F1 score, and a 92% recall score. Both the ensemble BILSTM+GRU and BILSTM+RNN models achieved 97% precision, recall, and F1 scores which shows the superiority of ensemble models for sentiment classification for food-related tweets.

Table 10 presents the precision, recall, and F1 score of standalone and deep ensemble models using the VADER approach. Similar to its performance with the TextBlob, the RNN model also achieved a low score on positive, negative, and neutral tweets using VADER. Individual BiLSTM models achieved an F1 score of 81% on negative tweets, while BiLSTM+RNN and BiLSTM+GRU achieved the highest F1 score of 87%. The GRU+RNN achieved an 89% F1 score and a 93% recall score. Both TextBlob and VADER in Tables 9 and 10 showed that the ensemble models perform better than the standalone models.

Topic extraction using LDA topic modeling

Topic modeling (Mujahid et al., 2021) in text-mining is used to understand the large corpus of text and provide useful insights about the topics discussed in the text. For topic modeling, LDA (Farkhod et al., 2021) is used on the preprocessed data and vectors created from the BoW. Then, we extract 10 highly discussed topics from the document that contains the keywords. Topic modeling intends to categorize the text into documents and words. Each topic includes a research-related keyword and a weighting. The positive and negative results of topic modeling based on the LDA approach are presented in Table 11.

Table 11 presents the most frequently discussed topics in McDonald’s tweets. The favorable words show that kids love to eat fast food at McDonald’s. Today, McDonald’s fries really drive better food fast. Kids are very happy to eat burgers and fries. The negative sentiments like kids eat fries at Mcdonald is year old wrong, chicken mcdonald nuggets meal order not arrive in time, a little mcdonald old girl shot because of threw water on employee. These are topics extracted from the McDonald’s tweets. Figure 5 shows the word cloud of the most frequently used words in tweets for McDonald’s. Table 12 presents the most common discussed topics in KFC tweets. The negative sentiments are: really bad fried chicken, sandwiches and fries are fucking India, KFC, hate chicken burger, sorry contact number worst hear, etc.

Figure 6 shows the word cloud for KFC tweets. LDA extracted the most common words from Burger King-related tweets are presented in Table 13. Burger King-related main positive and negative words are listed here. It shows that tweets use several bad words in negative sentiment-containing tweets like ‘bad amp work’ ‘hate time’, ‘shit bag’, ‘fucking white got black’, etc. It shows that negative words are more frequent than good words.

The word cloud for the most frequently used words in Burger King tweets is given in Fig. 7. The word cloud shows that several words are frequently used like ‘king emplouee’, ‘king burger’, ‘viral burger’, ‘behind wall’, etc.

Table 14 shows that Pizza Hut started a short survey to win a gift card for pizza domino’s, and it is better for the company. Pizza Hut had the same disgusting thing about three years ago. Pizza Hut should bring back the Sicilian-style pizza they had back in 1998. I ordered Pizza Hut and Wing Stop to be delivered at the same time. This is going to be so embarrassing. I could crush some Pizza Hut wings; apparently, there’s a Pizza Hut in my city now, so God bless the United States.

Table 15 shows the frequently used negative and positive words in the tweets for Subway food. Topics are extracted using LDA modeling on the preprocessed data. Results show that ‘killed’, ‘complaint’, ‘woman killed’, ‘fucking’, ‘wtf wrong’, ‘shot employee’, and several other highly negative and extremely harsh words are used in negative tweets. The word cloud for Pizza hut is given in Fig. 8.

Word cloud is used to show how textual information looks, and it is mostly used to analyze data from social platforms (Kabir, Ahmed & Karim, 2020). It is the summary of the whole document, and the word’s size denotes the frequency and prominence of data. It is used to know how much our data is related to research. Figure 9 presents the word cloud for Subway-related tweets.

Choice of fast food

Topic modeling is also carried out to find the sentiments of the people regarding the choice of fast food. For this purpose, the gathered tweets for the considered five companies are taken into account. Results indicate that a higher ratio of people remains neutral regarding selecting fast food with 49% of the total tweets. The people that prefer and like fast food are only 31% which have a positive view of fast food. The remaining 20% do not seem satisfied with fast food and show a negative attitude toward fast food, as shown in Fig. 10.

Comparison of proposed method with existing studies

We compared the performance of the proposed approach to several state-of-the-art approaches. The findings of the comparison with previously conducted research for sentiment analysis are presented in Table 16. All the results mentioned in Table 16 are taken from only sentiment analysis studies and deep learning models. Srinivas et al. (2021) deployed neural networks and LSTM to analyze the sentiments of tweets. The study achieved 87% accuracy with LSTM. The study employed CNN, NN, and LSTM models. The authors did not use ensemble models to improve accuracy. Similarly, Singh et al. (2022) combined LSTM and RNN models with enhanced attention layers to obtain better performance. The authors used different hyper-parameters and activation layers to increase the model performance and achieved an accuracy of 84.563%. Omara, Mousa & Ismail (2022) conducted a study on sentiment analysis using BiGRU+CNN. In comparison to these studies, the proposed BiLSTM + GRU model performs much better and shows better accuracy.