The proliferation of electronic documents in the digital environment necessitates efficient methods for extracting and categorizing user opinions. Sentiment analysis, the computational identification and categorization of opinions, plays a crucial role in understanding user sentiments and providing improved online services. While various methods exist, accurate and fast multi-label automatic classification remains a challenge. Existing dictionary-based approaches are labor-intensive and slow to adapt to emerging words, while machine learning methods, especially those based on sequential processes, struggle with multi-label classification and potential biases from earlier steps. This study addresses these limitations by focusing on improving multi-label classification for short texts, specifically using tweets as a representative dataset. Short text classification is increasingly important due to the prevalence of concise online communication in social media and other applications. The researchers aim to develop an efficient classifier requiring minimal training samples and costs for sentiment analysis, acknowledging the need to balance efficiency and model completeness.

The literature review surveys existing emotion classification methods, primarily focusing on dictionary-based and machine learning approaches. Dictionary-based methods, while offering theoretical and practical achievements, are hampered by the limitations of corpus resources, slow updates to emotional word lists, and the inability to identify new or altered words. Several studies using dictionary-based methods are reviewed; these demonstrate varying degrees of accuracy (e.g., Aman and Szpakowicz, 2007, achieved over 66% accuracy; Paltoglou and Thelwall, 2012, achieved 86.5% accuracy on Twitter and MySpace data; Taboada et al., 2011, reached 85.6% accuracy on a Twitter corpus). The review then shifts to machine learning methods, examining supervised and semi-supervised learning approaches. Supervised learning commonly utilizes word-level, sentence-level, and chapter-level features, with examples cited such as Keshtkar and Inkpen’s (2012) multi-level mood classification. Semi-supervised learning methods that use unlabeled samples are also discussed, noting the challenges of sensitivity to initial classification results. The literature review also covers neural network approaches, including those incorporating convolutional neural networks, recurrent neural networks (RNNs), and transformer-based models like BERT, highlighting their strengths and weaknesses in addressing fine-grained sentiment analysis and contextual information. The review concludes by noting the increasing recognition of the need for multi-label learning to capture the multifaceted nature of human emotions in text, citing related works such as Yang et al. (2014) and Liu and Chen (2015). While these multi-label methods show promise, they often lack iterative correction mechanisms.

This study modifies the MLkNN algorithm for improved multi-label emotion classification in short texts. The process begins by dividing the short text into sentences and extracting in-sentence features. These features are used with the MLkNN classifier as a base classifier to provide an initial emotion classification. The algorithm then incorporates the emotional transfer relationship characteristics between adjacent sentences and between the sentences and the full tweet text. This introduces a first adjustment step that modifies the overall emotion classification based on these inter-sentence relationships. The Average Precision (AVP) is calculated to evaluate the results. If convergence is not reached, a second adjustment is applied using label correlation, following a second-order strategy to investigate pairwise label relationships. This strategy utilizes a co-occurrence matrix to quantify label correlations. The co-occurrence frequency is normalized, and a parameter α (0 ≤ α ≤ 1) weights the combination of the initial MLkNN results and the label correlation adjustments. Finally, sentence-level emotion classifications are integrated to determine the overall tweet-level emotion classification. The methodology diagram illustrates these steps. The dataset used is the Sentiment140 Twitter corpus, with 6500 tweets selected after filtering and annotation. The dataset is divided into a 7:3 split for training and testing. Evaluation metrics include Subset Accuracy (SA), Hamming Loss (HL), One-Error (OE), Ranking Loss (RL), Average Precision (AVP), Accuracy (AC), Precision (PR), Recall (RE), and F-score. Three experimental groups are designed: one using basic MLkNN with varying features (unary grammar, with and without adjacent and full-text features); a second employing S-MLkNN, combining unary and binary grammar features with adjacent and full-text features; and a third using L-MLkNN, which adds label correlation adjustments to S-MLkNN.

The experimental results reveal several key findings. Firstly, the basic MLkNN classifier, when using only unary grammar features, achieves a relatively low average accuracy (around 42%). However, integrating features from adjacent sentences and the full tweet significantly improves accuracy, with increases of approximately 13% (K=5) and 15% (K=8). The second set of experiments demonstrates that using binary grammar features in addition to unary features enhances the initial classification results. Interestingly, the choice of K-value (number of nearest neighbors) appears to have less impact on accuracy in this case. The third set of experiments, using L-MLkNN with label correlation adjustments, demonstrates significant improvements. The optimal parameters (K=8 and α=0.7) achieve the highest accuracy and lowest Hamming Loss (HL) and One-Error (OE), indicating an enhanced ability to correctly classify multi-label emotions. The improved L-MLkNN algorithm outperforms the other methods in overall performance, particularly showing a considerable improvement in recall (RE), reaching 0.8019. A comparative analysis of the performance across various text lengths (Figure 2) reveals that while performance is consistent for short texts, the improved algorithms offer significantly better results for longer texts. Table 2, 3, and 4 detail the findings for each experimental group, showcasing performance across several metrics.

The findings demonstrate that the proposed improved MLkNN algorithm effectively improves the accuracy and speed of multi-label emotion classification for short texts. The incorporation of contextual information (adjacent sentences and full tweet) and label correlation significantly enhances classification accuracy, surpassing the performance of the basic MLkNN classifier. The results highlight the importance of considering both inter-sentence relationships and label dependencies when dealing with multi-label sentiment analysis. While the method achieves high accuracy on the chosen dataset, the study acknowledges the dependence on supervised learning and the potential limitations associated with the relatively small sample size of the dataset. The relatively small proportion of multi-labeled samples in the corpus might affect the generalization ability of the model. The superior performance of the L-MLkNN model, especially with K=8 and α=0.7, suggests a more refined approach to the classification problem, as compared to the simpler S-MLkNN and basic MLkNN.

This study successfully presents an improved multi-label emotion classification method based on MLkNN. The incorporation of contextual information (adjacent sentences and full tweet) and label correlation adjustments significantly improve the accuracy and recall of the classification process. The L-MLkNN model, with optimized parameters (K=8 and α=0.7), demonstrates superior performance compared to baseline methods. Future research should investigate the applicability of this method to larger, more diverse datasets and explore techniques to reduce reliance on large labeled datasets, potentially extending the approach to semi-supervised or unsupervised learning scenarios.

The primary limitation of the study is the relatively small size of the dataset used. The results may not fully generalize to larger or more diverse datasets, possibly impacting the robustness and external validity of the findings. The model’s performance might vary considerably when confronted with outliers or unexpected data patterns. Additionally, the study primarily uses supervised learning, limiting the applicability of the model to situations with limited or no labeled data. Future research should consider applying the model to a much larger, more diverse set of data to assess the generalizability of the results and explore semi-supervised or unsupervised learning approaches to address the reliance on labeled data.