Designing an intelligent push model for user emotional topics based on dynamic text categorization in social media news dissemination

Text sentiment analysis based on sentiment lexicons and statistical method

Turney (2002) posit that the adjectives or adverbs within the commentary text exert the predominant influence on its overall semantic inclination. They gauge the semantic orientation of a word by computing the mean semantic orientation of its adjectives and adverbs. The specific methodology involves calculating the disparity in mutual information between a given adjective or adverb and the benchmark terms “excel” and “poor” to ascertain the word’s semantic orientation. Sista & Polarized (2004) similarly contend that the emotional polarity of words present in the commentary text shapes the overarching emotional polarity of the commentary. Upon amalgamating the General Inquiry and WordNet sentiment dictionaries, the sentiment classification of textual expressions related to products and services is grounded in the consolidated sentiment dictionaries. The outcomes indicate a noteworthy enhancement in text classification accuracy by using sentiment dictionaries. Kamps et al. (2004) assert that word distance or similarity measures derived from WordNet sentiment lexicons predominantly align with WordNet’s classification relationships, thereby confining their applicability within noun and verb syntactic categories. Pang, Lee & Vaithyanathan (2002), utilizing extensive film reviews as a corpus, employed three machine learning algorithms—namely, support vector machine (SVM), naive Bayes (NB), and maximum entropy (ME)—for sentiment classification. The findings reveal that the SVM algorithm exhibits superior sentiment classification performance; nevertheless, all three algorithms underperform compared to topic-based classification methodologies. Ye, Zhang & Law (2009), leveraging blog comments within the tourism domain as their corpus, juxtaposed sentiment classification performance using three machine learning algorithms: SVM, NB, and a character-based N-gram model. The results demonstrate that naive Bayes manifests the least satisfactory classification performance among the trio.

Nevertheless, with a sufficiently extensive training set sample, the accuracy of these methods could surpass 80%. Mukwazvure & Supreethi (2015) adopted a hybrid approach for sentiment classification of news comments. Initially, sentiment tendencies were categorized into positive, negative, and neutral. Subsequently, the classification outcomes based on sentiment dictionaries served as training data for machine learning algorithms such as SVM and k-neural network (KNN). The findings highlight the superiority of the SVM algorithm over the KNN algorithm in the sentiment classification of news comments. Rohini, Thomas & Latha (2016) synthesized sentiment lexicons and decision tree classifiers to execute sentiment classification on a dataset of movie reviews. They compared sentiment classification results using the Kannada language movie review dataset and the machine-translated English review dataset.

Nevertheless, conventional methodologies for analyzing textual sentiment, relying on established sentiment dictionaries, manifest conspicuous deficiencies. Such strategies hinge on a restricted lexicon, posing a formidable challenge in encompassing diverse contexts and burgeoning lexicons. This limitation hinders capturing emotional nuances in specific domains or contemporary occurrences. Secondly, traditional approaches grapple with the inherent ambiguity of words, as a term may evoke varied emotions in disparate contexts. Resolving such ambiguity proves formidable, detrimentally impacting the precision of sentiment analysis.

Furthermore, conventional methods frequently disregard the contextual influence on emotion, neglecting to consider the amalgamation of adjacent words and syntactic structures adeptly. This oversight constrains a profound comprehension of the sentiment embedded within the text. Moreover, the emotional polarity inherent in sentiment dictionaries often stems from subjective judgments, introducing bias that impedes adaptability to diverse manifestations of emotion across varied demographics or cultures. Consequently, with the progression of deep learning technologies, textual content analysis can undergo marked improvement through protracted scrutiny of time-series data.

Text sentiment analysis based on deep learning methods

McDonald et al. (2007) conducted an inquiry into a structured model for text sentiment classification that operates at different levels of granularity collaboratively. The distinctive advantage of this model lies in its ability to correlate and mutually influence classification results across various text levels. Empirical experiments have demonstrated a notable reduction in classification errors compared to isolated training models (McDonald et al., 2007). Fei, Liu & Wu (2004) introduced a classification method that combines phrase patterns with unsupervised learning algorithms to analyze the sentiment tendencies in sports comment data on specific websites. The results substantiate the method’s efficacy, showcasing commendable classification outcomes (Fei, Liu & Wu, 2004). Kennedy & Inkpen (2006) leveraging semantic and machine learning algorithms, proposed a sentiment classification method for online movie review texts. The findings underscore that the amalgamation of these two methodologies yields superior results compared to their individual use (Kennedy & Inkpen, 2006). Thet, Na & Khoo (2008) advocated combining information extraction techniques and machine learning in sentiment classification for in-depth analysis of movie review texts. Experimental results affirm the effectiveness of this method in the sentiment classification of online movie reviews (Thet, Na & Khoo, 2008). Hajmohammadi et al. (2015) introduced an innovative learning model that combines uncertainty-based active learning and semi-supervised self-training methods to classify sentiment in online book review data across three different languages.

The outcomes demonstrate a significant enhancement in the effectiveness of cross-lingual sentiment classification (Hajmohammadi et al., 2015). In recent years, the attention mechanism, recognized as an efficient technique, has seamlessly transitioned from image recognition to natural language processing. Its integration with deep learning techniques has proven to enhance the effectiveness of sentiment classification models. Usama et al. (2020) proposed a novel model based on RNN and CNN, incorporating the attention mechanism. Basiri et al. (2021) introduced an attention-based bi-directional CNN-RNN model that assigns weights to the bi-directional feature extraction layer using the attention mechanism. Liu et al. (2021) incorporated positional coding alongside a multi-head attention mechanism, employing graph convolutional networks to capture sentiment information from text, address interactions between contextual and aspectual words, and analyze long-distance dependencies between words. Zhu et al. (2021) introduced global and local textual information simultaneously into aspect-based sentiment classification, utilizing the attention mechanism to merge global and local features effectively.

The research above underscores that owing to the exponential surge in network information, traditional model-based approaches to social media opinion and sentiment analysis are increasingly inadequate for the requirements of pertinent departments. Consequently, the evolution of artificial intelligence-based methods, rooted in machine learning and deep learning, has emerged as a pivotal solution to address these challenges. Following data acquisition, the obtained data undergoes analysis using deep learning models for sentiment analysis. Enhancing the model’s efficacy through data reinforcement, achieved via reinforcement learning and other methodologies, markedly improves data performance and the model’s recognition capabilities. This progression offers valuable technical insights for the nuanced sentiment analysis of dynamic text.

The BI-LSTM for the text recognition

For sentiment analysis on textual data, the crux lies in addressing a time series classification problem, making RNNs the optimal solution within deep learning. To better facilitate the study of contextual semantics, this article employs a BI-LSTM network for model construction. LSTM is a specialized variant of recurrent neural networks expressly designed to overcome challenges associated with long sequence dependencies (Shi et al., 2022). By incorporating a gating mechanism, LSTM adeptly captures and manages prolonged dependencies in sequences, mitigating issues such as gradient vanishing or explosion that afflict traditional RNNs when handling extensive sequences.

The LSTM unit predominantly executes data optimization through an oblivious gate unit, ensuring effective long-term control. The implementation process of the oblivious gate is delineated in Eq. (1):

(1) $$f_t=\sigma \left(W_f\cdot \left[h_{t-1},x_t\right]+b_f\right)$$where $W_f$ is the weight matrix. $h_{t-1}$ is the hidden state. $x_t$ is the input of at current time. The computation of the hidden state involves the output from the preceding stage and the corresponding memory cells, as depicted by Eq. (2):

(2) $$h_t=o_t\cdot \mathrm{t}\mathrm{a}\mathrm{n}\mathrm{h}\left(C_t\right)$$where C_t is the state of the cell at time step t. It is related to the computation of the forgetting gate, which can be calculated by Eq. (3):

(3) $$C_t=f_t\cdot C_{t-1}+i_t\cdot \tanh \left(W_c\cdot \left[h_{t-1},x_t\right]+b_c\right).$$

Upon completing the aforementioned calculations for the pertinent state variables, the output of each cell, and consequently the final output, can be obtained:

(4) $$o_t=\sigma \left(W_o\cdot \left[h_{t-1},x_t\right]+b_o\right).$$

BI-LSTM represents an extension of LSTM, enhancing the capture of contextual information by assimilating both forward and reverse information from the input sequence. The hidden states of BI-LSTM are concatenated, incorporating information from both forward LSTM and reverse LSTM. By processing information in both left-to-right and right-to-left directions, BI-LSTM facilitates a more comprehensive understanding of the entire sequence. This bidirectional approach enables the model not only to consider past information but also to anticipate future information. The computational process of BI-LSTM is elucidated in Eqs. (5), (6):

(5) $$\overrightarrow{h_t}=\overrightarrow{\mathrm{L}\mathrm{S}\mathrm{T}\mathrm{M}}\left(x_t,h_{t-1}\right)=\sigma \left(W\cdot \left[h_{t-1},x_t\right]+b\right)$$

(6) $$\overleftarrow{h_t}=\overleftarrow{\mathrm{L}\mathrm{S}\mathrm{T}\mathrm{M}}\left(x_t,\overleftarrow{h_{t+1}}\right)=\sigma \left(\mathrm{W}\cdot \left[\overleftarrow{h_{t+1}},x_t\right]+\mathrm{b}\right)$$where $h_t$ is the time step $t$ of the hidden state, and $\overrightarrow{h_t}$ is the output of forward and $\overleftarrow{h_t}$ is the output of the backward, and the process of merging the two outputs in the sequence is shown in the right part of Fig. 2.

BI-LSTM exhibits significant advantages in sequence modeling tasks. Simultaneously processing forward and backward information, BI-LSTM excels at capturing contextual relationships more comprehensively, providing richer contextual information. This attribute is particularly advantageous for tasks involving long-distance dependencies. The bidirectional structure enhances the model’s ability to learn abstract representations of sequences, mitigates the rate of information loss, maintains symmetry balance, and elevates the stability and robustness of the model. BI-LSTM finds practical application in natural language tasks for recognizing patterns and analyzing data in context. Its capacity to consider contextual information on both the left and right sides of a word simultaneously enables a more comprehensive understanding of context and the capture of long-distance dependencies in text. Given the uncertainty and variability in text length, BI-LSTM exhibits flexibility in adapting to indeterminately long input sequences. It demonstrates the ability to learn features, rendering it well-suited for analyzing input and output in related data.

The attention mechanism for the model enhancement

The attention mechanism emulates how the human brain filters critical information and constructs probability distribution matrices to associate features that require learning. Initially proposed for application in image processing, the attention mechanism has since found widespread use in natural language processing tasks. Operating as a weight distribution mechanism, the attention mechanism can extract critical information from extensive data and adjust the significance of text features by modifying the weight coefficients (Niu, Zhong & Yu, 2021). The larger the weight coefficient, the more crucial the information is deemed in the comment text, exerting a more significant impact on the sentiment polarity classification results of the text. Assuming Q is a matrix composed of a set of queries, K is a matrix consisting of a set of keys, and V is a matrix composed of a set of values, the attention mechanism, denoted as Attention (Q, K, V), is delineated in Eq. (7):

(7) $$\mathrm{A}\mathrm{t}\mathrm{t}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}(Q,K,V)=\sum _{i=1}^m\frac{1}{z}\mathrm{e}\mathrm{x}\mathrm{p}\left(\frac{<q_t,k_i>}{\surd d_k}\right)v_i$$where $Z$ is the normalization factor, and $d_k$ is the dimension of the word embedding vector, which can play the role of adjusting factor in the formula so that the inner product will not be too large. $q_t$ (query) is the task to be queried, and $k_i$ (key) and $v_i$ (value) are the corresponding key-value pairs. $q_t$ by calculating the inner product. The entire process can be visually represented in Fig. 3. These scores are normalized into attention weights using the softmax function for different key values Q, K, and V. These weights are then utilized to weigh and sum the values in the input sequence, ultimately generating the final attention output. This dynamic attention mechanism enables the model to selectively focus on information at various locations within the input sequence, enhancing its capability to capture key information. In this article, incorporating the attention mechanism involves augmenting the feature vocabulary based on the results obtained from the BI-LSTM method, thereby completing the enhanced training of the model.

In sentiment analysis of social media content, the introduction of attention mechanisms is precious due to user-generated text’s diverse and often unstructured nature. Social media posts vary widely in length, structure, and emotional complexity, with important sentiment cues sometimes embedded within specific phrases or keywords. Attention mechanisms enable the model to dynamically focus on these relevant portions of text, significantly enhancing its ability to capture critical features and contextual information essential for accurate sentiment interpretation. Furthermore, attention mechanisms improve the model’s adaptability to the informal language and mixed sentiments commonly found in social media, leading to more robust sentiment classification. This dynamic focus also contributes to the model’s interpretability, allowing researchers and practitioners to gain insight into which words or phrases most influence the model’s decision-making process, strengthening confidence in the analytical results.

The ATT-BI-LSTM for the social media text recognition

After completing the introduction of the BI-LSTM model and the attention mechanism, we carried out the construction of the model, and the overall model framework is shown in Fig. 4.

The model first processes social media-related data collected from public datasets, involving data cleaning, keyword analysis, and text vectorization to ensure the data is high quality and structured appropriately for model input. During data cleaning, redundant characters, stopwords, punctuation, and irrelevant information are removed to reduce noise and maintain data purity. For keyword analysis, techniques like word frequency statistics and TF-IDF are used to extract core terms from each text, allowing for a more accurate reflection of the text’s themes and sentiments. To efficiently represent text semantics, we vectorize the text data using pre-trained word embedding models such as Word2Vec or GloVe, ensuring each word is mapped into a multidimensional vector space, preserving the semantic and contextual information of the text. The dimensionality of the word vectors is specifically set to capture emotional features, balancing generalization ability and ease of data processing. The processed word vectors are then fed into the model, with the input layer’s maximum text length set to 512 to capture the complete information in most comments. The model uses a two-layer BI-LSTM structure, each containing 128 LSTM units, designed to capture contextual sentiment information in both forward and backward directions. Additionally, an attention mechanism is incorporated, allowing the model to adaptively focus on essential words and phrases, thereby enhancing the capture of emotional features and improving sentiment classification accuracy. In the classification stage, Softmax and a fully connected layer are used to perform binary classification of sentiment, identifying positive or negative emotions within the text.

Experiment setup

To enhance the model’s validation and training, a public dataset was utilized for analysis and pre-training under migration learning, facilitating the testing of the model on real-world data. The microblogging dataset employed in this study is the GitHub public microblogging sentiment analysis dataset (https://github.com/SophonPlus/ChineseNlpCorpus/blob/master/datasets/weibo_senti_100k/intro.ipynb.), containing sentiment annotations for Sina microblogs, totaling over 100,000 items. The dataset exhibits an equal distribution of negative and positive sentiments, each accounting for half of the entries. For practical application considerations, data entries were filtered, and 2,000 positive comments and 2,000 negative comments were selected for analysis.

In addition, the IMDB review dataset (Tripathi et al., 2020) was employed, representing a binary categorical sentiment dataset with 50,000 IMDB film reviews. Reviews with a negative score of <= 4 and a positive score of >= 7 were included. Given the volume of data, 2,000 positive and 2,000 negative reviews were chosen for analysis, thereby completing the model evaluation.

Following the dataset determination, corresponding indicators for model evaluation were selected. Precision, Recall, and F1 score were chosen as the evaluation metrics in this article. The calculation methods for these indicators are delineated in Eqs. (8)–(10):

(8) $$P={\displaystyle \frac{TP}{TP+FP}}$$

(9) $$R={\displaystyle \frac{TP}{TP+FN}}$$

(10) $$F1={\displaystyle \frac{2\times P\times R}{P+R}}$$where $TP$ denotes a true example; $FP$ denotes a false positive example; and $FN$ denotes a false negative example. In this model, careful tuning of hyperparameters was essential to achieve optimal performance in sentiment classification. The input layer was set with a maximum text length of 512, balancing the need to capture context with computational efficiency. Each of the two layers in the BI-LSTM model was configured with 128 LSTM units, a choice aimed at maintaining a balance between capturing sufficient sequential information and managing model complexity to avoid overfitting. The attention mechanism enhanced the model’s focus on crucial features, improving sentiment detection. The batch size for training was set to 10, allowing efficient updates while maintaining stability in gradient descent. The Adam optimizer was chosen for its adaptability to varying learning rates, enhancing convergence speed. The classification was finalized with a Softmax layer, clearly distinguishing between positive and negative sentiments. This combination of hyperparameters helped the model achieve robust performance while efficiently processing the social media data for accurate sentiment classification.

Emotion classification result and model comparison

Following the completion of the model’s dataset and evaluation indicators, we proceeded with the model testing. In the comparative analysis of models, we selected commonly used methods in natural language processing, including CNN, RNN, LSTM, and BI-LSTM. The model-building parameters, the number of layers, and the proposed model have been aligned. The specific results obtained from Weibo data are illustrated in Figs. 5 and 6.

In Figs. 5 and 6, it is evident that the method proposed attains elevated recognition results in both positive and negative emotion recognition. The precision is 0.821 for positive emotion recognition and 0.832 for negative emotion recognition, demonstrating commendable recognition outcomes. Conversely, without adding the attention module, the method exhibits a diminished recognition rate. The recognition results for another dataset, IMDB, are depicted in Figs. 7 and 8.

In Figs. 7 and 8, the addition of the attention module contributes to an overall improvement in recognition rates. After incorporating the attention module, the precision for positive and negative emotions is 0.829 and 0.817, respectively. Both metrics demonstrate significant enhancements due to the individual contributions of the attention module. In contrast, the precision of BI-LSTM without the attention module is 0.781 and 0.758, indicating a substantial improvement in the model’s overall performance after incorporating the relevant module. To further validate the impact of the attention mechanism module on model performance, ablation experiments were conducted, and the results are illustrated in Figs. 9 and 10.

In Figs. 9 and 10, it is evident that whether on the Weibo dataset or the IMDB dataset, the addition of the attention module consistently yields superior overall recognition results across different batches compared to the standalone BI-LSTM model. The recognition accuracy is more consistent, resulting in an overall improved performance. Hence, the multi-faceted model proposed proves to be more adept at accomplishing the task of emotion recognition in dynamic text comments on social media news, thereby enhancing opinion analysis.

The practical test for the model

As per this article’s requirements, we analyzed the APP data of interest sourced from news comments over the past 3 years. Approximately 1,000 entries were selected for data analysis, and the labels for these data entries were manually assigned. Simultaneously, to augment the dataset, we utilized Weibo data and IMDB data for separate training and observation through transfer learning. The recognition rates of different models obtained after transfer learning are presented in Table 1 and Fig. 11.

From the curves shown in the figure, we can observe that the transfer effect obtained after pre-training the model on the same data is similar, indicating that the proposed ATT-Bi-LSTM method achieved favorable recognition performance. The average precision for both positive and negative classifications exceeded 90%, demonstrating the effectiveness of this method in the required application scenarios of this study. We constructed a Python-based deep learning environment for model deployment, primarily using TensorFlow and PyTorch frameworks, running on high-performance servers to ensure efficient model training and testing. The data preprocessing pipeline includes tokenization, stopword removal, and lemmatization, leveraging libraries like NLTK and SpaCy for accurate text processing. For word vector initialization, we used pre-trained embeddings such as GloVe and Word2Vec to capture deep semantic relationships within the text, thereby improving the model’s initial performance. To ensure consistency and ease of deployment across various environments, the model was encapsulated in a Docker container, enabling rapid deployment in development, testing, and production environments. Kubernetes was employed for container orchestration, allowing for automatic scaling as needed. Data processing and model training operations were conducted within the container to maintain a consistent environment, reducing dependency issues.

Additionally, model performance evaluation was conducted using cross-validation, with continuous hyperparameter tuning to enhance model accuracy. For transfer learning, we initially trained the pre-trained model on source domain data, fine-tuning the target domain and leveraging Weibo and IMDB datasets to enhance cross-domain sentiment classification. This deployment environment enabled efficient data processing and model updating, establishing a robust foundation for future large-scale applications.