Source-linked AI summary
Sentiment Analysis Based on Deep Learning: A Comparative Study
Nhan Cach Dang, María N. Moreno-García, Fernando De la Prieta
TL;DR
Sentiment analysis is important for interpreting user opinions but remains difficult because of NLP-related accuracy challenges. This paper reviews deep-learning approaches and compares DNN, CNN, and RNN models using TF-IDF and word embedding across varied datasets. It reports that word embedding generally outperforms TF-IDF, while CNN offers the best accuracy–runtime balance and RNN is most accurate with word embedding but much slower.
Problem
Sentiment analysis faces NLP-related theoretical and technical issues that hinder polarity-detection accuracy, while practical comparisons across deep-learning techniques remain limited.
Method
The study reviews deep-learning sentiment-analysis research and evaluates DNN, CNN, and RNN models using TF-IDF and word embedding across datasets of tweets and reviews.
Results
Word embedding generally performs better than TF-IDF; CNN provides the best accuracy–CPU-runtime balance, while RNN is most accurate with word embedding but takes much longer.
Takeaways & Limitations
The comparison provides a broad perspective on combining deep-learning architectures with text-preprocessing techniques for sentiment analysis.
Abstract
from arXiv · showhide
The study of public opinion can provide us with valuable information. The analysis of sentiment on social networks, such as Twitter or Facebook, has become a powerful means of learning about the users' opinions and has a wide range of applications. However, the efficiency and accuracy of sentiment analysis is being hindered by the challenges encountered in natural language processing (NLP). In recent years, it has been demonstrated that deep learning models are a promising solution to the challenges of NLP. This paper reviews the latest studies that have employed deep learning to solve sentiment analysis problems, such as sentiment polarity. Models using term frequency-inverse document frequency (TF-IDF) and word embedding have been applied to a series of datasets. Finally, a comparative study has been conducted on the experimental results obtained for the different models and input features
A Comparative Study
The study is authored by Nhan Cach Dang, María N. Moreno-García, and Fernando De la Prieta. It was received in January 2020, accepted in March, and published on 14 March 2020.
- The authors are Nhan Cach Dang, María N. Moreno-García, and Fernando De la Prieta.
- The manuscript was received on 31 January 2020 and accepted on 10 March 2020.
- The paper was published on 14 March 2020 and focuses on sentiment analysis, deep learning, machine learning, neural networks, and NLP.
1. Introduction
The introduction motivates sentiment analysis through its importance for social-data applications and the difficulty of accurate polarity detection. It identifies limited practical comparisons of deep-learning methods and proposes evaluation across datasets, features, accuracy, F-score, and processing time.
- Sentiment analysis supports applications involving public opinions, including recommender systems, organizational surveys, political campaigns, and government analysis.
- Automatic sentiment analysis remains challenging because NLP issues, including negation and domain dependence, hinder polarity-detection accuracy.
- Few studies evaluate different deep-learning techniques with practical evidence across varied datasets and application domains.
- The study compares methods using overall accuracy, F-score, and processing time rather than reliability metrics alone.
- It applies DNN, RNN, and CNN with TF-IDF and word embedding to Twitter datasets and reviews recent deep-learning sentiment-analysis research.
2. Background
The background introduces deep-learning architectures and sentiment-analysis approaches, then describes preprocessing and the two feature representations used as model inputs. It distinguishes learned word representations from corpus-based TF-IDF weighting.
- Deep learning automatically learns and extracts features through multilayer neural-network architectures, unlike traditional approaches that define or select features manually.
- Deep Neural Networks (DNN): DNNs contain input, hidden, and output layers, with more than two layers including hidden layers.
- Convolutional Neural Networks (CNN): CNNs use convolution and pooling layers to extract features, reduce feature resolution, and feed a fully connected classification layer.
- Recurrent Neural Networks (RNN): RNNs process sequential information using feedback-based internal memory, while LSTMs use longer memory in hidden-layer activation functions.
- Sentiment Analysis: Sentiment analysis identifies positive, negative, or neutral opinions at aspect, sentence, or document level using lexicon-based, machine-learning, or hybrid approaches.
- Input Features: Word embedding maps similar-meaning words to similar real-valued vectors, whereas TF-IDF measures a word’s importance within a document corpus.
3. Related Work
Related work spans deep-learning sentiment analysis across models, languages, domains, and task types. The review identifies DNN, CNN, and hybrid models as popular polarity-analysis choices while noting a lack of direct CNN, RNN, and LSTM comparison.
- The review examines technical challenges, datasets, proposed methods, and application domains as references for future empirical studies.
- Recent work applies DNN, CNN, and RNN models to sentiment analysis and compares CNN, RNN, and LSTM techniques across studies.
- Deep-learning sentiment analysis has been applied in finance, weather tweets, trip reviews, cloud-service recommendations, and movie reviews.
- Polarity-based studies mainly analyze English tweets, with additional work on Spanish, Thai, and Persian tweets.
- Neural-network studies also address aspect sentiment, including semantic annotation, sentiment ratings, and product-aspect importance.
- Among 32 analyzed studies, DNN, CNN, and hybrid models were identified as popular polarity-analysis approaches, but comparative analysis of CNN, RNN, and LSTM was lacking.
- Many studies extract text features and convert them into Word2vec embeddings before sentiment analysis.
4. Comparative Study
The comparative study evaluates sentiment classification across datasets using three deep-learning models and two input-preparation approaches. It organizes the broader literature by study details, methods, datasets, and targets.
- Datasets: Eight datasets support experiments spanning tweets and user reviews with positive, negative, and neutral labels.The datasets include 1.6 million tweets at maximum, smaller tweet collections, and 125,000 review comments.
- Input features and models: Word embedding and TF-IDF are compared as alternative inputs for DNN, CNN, and RNN sentiment classifiers.Word embedding uses Word2vec, including skip-gram and continuous bag-of-words models.
- Related work: The related-work summary records publication year, authors, research work, methods, datasets, and study targets across 32 papers.The table is presented as a structured overview of important research related to the study.
- Study objective: The study compares model and feature performance across sentiment-analysis tasks to contribute to the state-of-the-art literature.The comparative framework covers the three selected deep-learning models and both preprocessing techniques.
- Preprocessing: The preprocessing pipeline cleans and lemmatizes text before converting sentences into feature vectors for the evaluated algorithms.Both feature-vector types serve as inputs, producing two induced models per algorithm.
5. Experimental Results
Experiments compare DNN, CNN, and RNN with TF-IDF or word embedding across sentiment datasets, evaluating accuracy, recall, precision, F-score, AUC, and processing time. Word embedding generally performs better, while CNN offers a strong accuracy–time tradeoff and RNN is most accurate but computationally costly.
- Experimental setup: DNN, CNN, and RNN were evaluated with TF-IDF and word embedding across tweet and review datasets.Accuracy, recall, precision, F-score, and AUC were used, with CPU time also considered.
- Dataset size: Word embedding was comparatively insensitive to dataset size, whereas TF-IDF showed a slightly greater and uneven dependence on sample size.For Sentiment140, using 50% of the sample left evaluation measures largely unaffected while reducing processing time by 50%.
- Feature representations: Word embedding generally outperformed TF-IDF across metrics and datasets, although some review datasets favored TF-IDF combinations slightly.For book reviews, CNN with word embedding achieved the highest accuracy, recall, F-score, and AUC; book and music reviews were exceptions to the broader pattern.
- Cross-dataset results: The additional datasets confirmed the Sentiment140 conclusions, with RNN and word embedding generally performing best despite dataset-specific exceptions.The Tweets Airline dataset showed high values across metrics in all cases.
- Model comparison: RNN with word embedding showed the highest reliability, but RNN was also the most time-consuming algorithm.RNN with TF-IDF took longer and produced lower accuracy, around 50%, on tweet and review datasets.
- Model comparison: CNN provided the best tradeoff between processing time and accuracy, while its processing time was slightly longer than DNN’s but its evaluation measures were much better.RNN with word embedding achieved the highest accuracy, but its processing time was 10 times longer than CNN’s.
6. Conclusions
The study combines TF-IDF and word embedding with DNN, CNN, and RNN models to compare sentiment-analysis performance across datasets. It finds that word embedding and CNN offer favorable practical tradeoffs, while dataset diversity and hybrid methods remain important directions for future work.
- 6. Conclusions: The study combined word embedding and TF-IDF with DNN, CNN, and RNN architectures to evaluate sentiment analysis across datasets with different topics.The experiments were designed to assess preprocessing techniques and deep learning models comparatively.
- 6. Conclusions: RNN achieved the highest accuracy with word embedding but took 10 times longer to process than CNN, limiting the significance of its improvement when computational cost matters.With TF-IDF, RNN was slower without significantly better results.
- 6. Conclusions: Word embedding generally produced better results than TF-IDF, while CNN offered the best balance between accuracy and CPU runtime.RNN was slightly more reliable with most datasets but required substantially more computation.
- 6. Conclusions: Because algorithm effectiveness depends largely on dataset characteristics, future evaluations should use more diverse datasets and examine hybrid approaches.The proposed future work also extends comparisons to new methods and social-media data types.