Contributions of the study: Innovations in methodology and multidisciplinary approaches to chinese sentiment analysis

Our research presents three pivotal contributions to the field. Initially, we tackle the distinctive attributes of Chinese textual data and the specific demands of sentiment analysis through the strategic use of part-of-speech classifications for Chinese radicals and lexicon, thereby enhancing the semantic comprehension of the text. This targeted approach delves into the intricacies of sentiment analysis within the Chinese language, offering novel perspectives and advancements.

Furthermore, our work synthesizes state-of-the-art theoretical frameworks and analytical techniques from a spectrum of disciplines, encompassing computational intelligence, linguistic studies, information technology, and the realm of artificial intelligence. This comprehensive, cross-disciplinary strategy not only equips us with a robust arsenal of innovative methodologies but also facilitates the generation of solutions that are both creative and effective.

Lastly, we pioneer the adoption of the “knowledge+data” research paradigm in the arena of Chinese sentiment analysis. This novel approach amalgamates emotional knowledge vectors with the feature vectors derived from sophisticated deep learning models, culminating in a sentiment analysis methodology that harnesses a spectrum of semantic features ranging from characters to words, radicals, and parts of speech. Such an integrative strategy not only bolsters the analytical prowess of Chinese sentiment analysis but also contributes to the enrichment and evolution of its theoretical and methodological underpinnings.

In summary, our research introduces a novel suite of methodologies and a multidisciplinary lens to the domain of Chinese sentiment analysis. This innovative framework paves the way for fresh perspectives and deeper understanding, propelling the field forward with its unique contributions:

Novel Feature Integration: Systematic incorporation of radical features (e.g., 忄, 心) and emotional part-of-speech features (verbs/adjectives) into Chinese sentiment analysis, addressing linguistic uniqueness.

Knowledge-Data Synergy: A hybrid framework fusing emotional knowledge vectors (via TransE) with deep learning features (BiGRU + multi-head attention), enabling multi-granularity semantic fusion.

Cross-Disciplinary Methodology: Integration of computational linguistics, AI, and knowledge graphs, validated on large-scale datasets (F1: 89.23% on Douban, 84.84% on NLPECC).

Our work directly addresses these limitations by

Incorporating Radical Semantics: Explicitly modeling radicals (e.g., 心 [“heart”] in 怒 [“anger”]) as emotional anchors.

Leveraging Chinese-Centric Knowledge Graphs: Utilizing the DLUT-EmotionOntology to construct emotion-intensity triplets tailored to Chinese.

Hybrid Knowledge-Data Fusion: Integrating TransE-based knowledge vectors with BiGRU and multi-head attention, enhancing robustness against adversarial perturbations.

Data input layer

The data input layer is responsible for preprocessing Chinese text and generating input data. This article converts Chinese text into five types of input data: character level radicals, word level radicals, word level radicals, and part of speech text.

For input text T, it consists of m words, that is, word level text T^c={c₁, c₂,..., c_m}, where c_i (i = 1,2,..., m) represents each word in T; Using the Jieba word segmentation tool (https://github.com/fxsjy/jieba), T is segmented into n words, where the word text T^w={w₁, w₂,..., w_n}, where w_i (i = 1,2,..., n) represents each word in T; According to the radical mapping relationship in Xinhua Dictionary, T^c and T^w respectively obtain word level radical texts T^rc={rc₁, rc₂,..., rc_m} and word level radical texts T^rw={rw₁, rw₂,..., rw_n}, where rc_i (i = 1,2,..., m) represents word level radicals, rw_i (i = 1,2,..., n) represents word level radicals; Use the Jieba part of speech annotation tool (https://gist.github.com/hscspring/c985355e0814f01437eaf8fd55fd7998) to convert T^w into part of speech text T^pos={pos₁, pos₂, …, pos_n}, where pos_i (i = 1,2,..., n) represents the corresponding part of speech of the word. From the above analysis, it can be seen that |T^c| = |T^rc|, |T^w| = |T^rw| = |T^pos|, |·| represents the size of the text.

Vector representation layer

The vector representation layer utilizes the Word2Vec word embedding method to obtain the corresponding vector sets of five types of input data {T^c, T^rc, T^w, T^rw, T^pos} [21]. As shown in Fig 1, where E^c={} represents the set of word vectors, and (i = 1,2,..., m) represents the word vector; E^rc={} represents the set of word level radical vectors, (i = 1,2,..., m)) represents the word level radical vectors; E^w={} represents the set of word vectors, (i = 1,2,..., n) represents the word vector; E^rw={} represents the set of word level radical vectors, (i = 1,2,..., n) represents the word level radical vectors; E^pos={} represents the set of part of speech vectors, and (i = 1,2,..., n) represents the part of speech vector.

The Word2Vec method trains vectors using continuous bag of words (CBOW) and skip gram models. Given that the Skip Gram model has better vector quality trained in large-scale corpora, The input data are vectorized using this model. Taking word vectors as an example, the Skip Gram model predicts the probability of contextual context word w_o through the central word wc. The model structure is shown in Fig 2, where w_t represents w_c; w_t-2, w_t-1, w_t+1, w_t+2 represents w_o, and SUM represents a sum operation.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Structure of Skip-Gram model.

https://doi.org/10.1371/journal.pone.0325428.g002

Specifically, this model represents each word as a word vector of the center word and a word vector of the background word, in order to calculate the conditional probability between the center word and the predicted background word, as shown in Equation (1):

(1)

Wherein, p represents the conditional probability, wc represents the center word, w_o represents the background word, v_c represents the word vector of the center word, v_o represents the word vector of the background word, N represents the dictionary size, c, o, i represent the index of the word in the dictionary, and exp(·) represents the exponential function based on the natural constant e.

Feature extraction layer

Given the significant sequence characteristics of Chinese text, the Bidirectional Gated Recurrent Unit (BiGRU) model is adopted as the basic model. This model achieves effective utilization of text contextual semantic features by concatenating feature vectors of GRU models with forward and reverse directions. The working principle of the GRU model is shown in equations (2) to (5), and the principle of BiGRU is shown in equations (6) to (8):

(2)

(3)

(4)

(5)

Wherein, x_t represents the input vector at time t, r_t and z_t represent the reset gate and update gate at time t, W and b represent the corresponding weight matrix and bias vector, representing candidate memory units, sigmoid(·) and tanh(·) represent the activation function, ht represents the output vector at the current time, and ⊙ is the Hadamard product, × Represents matrix multiplication.

(6)

(7)

(8)

Wherein, x_t represents the input vector at time t, h^r and h^l represent the feature vectors obtained from the forward and reverse GRU models, respectively, and y_t represents the feature vectors obtained from the BiGRU model at the current time.

The inspiration for attention mechanism comes from human attention, which can distinguish the importance of input data by assigning different weights. From the perspective of implementation methods, the attention mechanism is mainly achieved through linear weighting and dot product methods, with no essential difference in effectiveness. However, the dot product method has a faster calculation speed. Therefore, this article uses the dot- product attention mechanism for feature fusion. The calculation process is shown in equation (9):

(9)

Wherein, Q and K are the Query matrix and Key matrix, V is the Value matrix, and softmax(·) represents the normalization function. The implementation steps are as follows:

1. 1. Step 1: BiGRU-Based Feature Extraction

The feature extraction layer employs a bidirectional gated recurrent unit (BiGRU) model to capture sequential dependencies in the input text.

1. 2. Step 2: Attention-Driven Fusion

A dot-product attention mechanism is then applied to integrate the word vector with the corresponding word-level radical vector (e.g., 忄 for 恨 [“hate”]).

1. Output: This fusion process generates a unified word radical feature vector, encoding both semantic and radical-level emotional cues.

The word vector is then fused with the word level radical vector to obtain the word radical feature vector. The word vector is then fused with the part of speech vector to obtain the word part of speech feature vector.

Taking the generation of character radical feature vectors as an example, the workflow of this layer is introduced. The schematic diagram of the generation of character radical feature vectors is shown in Fig 3.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Schematic diagram of character-radical feature vector generation.

https://doi.org/10.1371/journal.pone.0325428.g003

In Fig 3, first set the initial states of the BiGRU^c model to 0, input the word vector set E^c into the BiGRU^c model, and obtain the word feature vector set y^c={}, where (i = 1,2,..., m) represents the word feature vector. The calculation is shown in equation (10):

(10)

Then, through the dot-product attention mechanism, y^c is fused with the character level radical vector set E^rc to obtain the fused vector (i = 1,2,..., m). The calculation is shown in equations (11) and (12):

(11)

(12)

Wherein, α represents the weight matrix after the dot product operation,· represents the dot product operation, T represents the matrix transpose operation, and softmax(·) represents the softmax normalization function.

Finally, (i = 1,2,..., m) is used as the input vector for BiGRUr^c and the hidden layer state of the BiGRUr^c model at the last moment is transmitted to. The BiGRUr^c model is used as the initial state to obtain the set of character radical eigenvectors V^r_c ={}, where (i = 1,2,..., m) represents the character radical feature vector, and its calculation is shown in equation (13):

(13)

Similarly, the generation process of word radical feature vectors and word part of speech feature vectors is similar to that of word radical feature vectors, resulting in a set of word radical feature vectors V^r_w={} and the set of word part of speech feature vectors V^pos_w={}, where (i = 1,2,..., n) represents the word radical feature vector, (i = 1,2,..., n) represents the word part of speech feature vector (as shown in Fig 1).

Feature fusion layer

The TransE model belongs to the knowledge graph embedding model [22], which represents entities and relationships in the knowledge graph through distributed vector representation, thereby obtaining entity semantic vectors. The schematic diagram of the TransE model is shown in Fig 4.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Structure of TransE model.

https://doi.org/10.1371/journal.pone.0325428.g004

If the head entity is h, the tail entity is t, and the relationship is r, then for a given triple (h, r, t), the TransE model represents it as (h, r, t) in Fig 4, where h is the vector representation of the head entity, r is the vector representation of the relationship, and t is the vector representation of the tail entity. During the training process of the TransE model, the triplets at h + r ≈ t are made positive sample triplets, and the triplets at h + r ≠ t are made negative sample triplets. By constructing the objective loss function L, the distance between positive sample triplets is reduced, while the distance between negative sample triplets is increased. The calculation formula for L is shown in equations (14) and (15):

(14)

(15)

Wherein, d is the distance function that measures the distance between the two vectors h + r and t, | |· | | represents the Euclidean distance, which is calculated using L1 or L2 norms, S⁺ represents the set of positive sample triplets, and S^- represents the set of negative sample triplets, γ The interval in the loss function, which is greater than 0.

The multi head attention mechanism enhances the model’s attention ability by concatenating multiple attention mechanisms horizontally, thereby representing semantic information from different positions and aspects. The principle is shown in formulas (16) and (17):

(16)

(17)

Wherein, head represents the head of attention, h represents the number of heads, Q, K, and V are Query vectors, Key vectors, and Value vectors, Q,K, and V are the weight matrices of Query, Key, and Value for the k-th head, W^o is the weight matrix, Concat(·) represents the concatenation function, Q,K,V, and Concat(·)W^o represents the Line layer operation in Fig 1.

Based on the above analysis, the feature fusion layer uses approximately 27000 emotional words from the emotional vocabulary ontology library (https://github.com/ZaneMuir/DLUT-Emotionontology) as the head entity h, 21 emotional categories as the tail entity t, and the emotional intensity of emotional words as the relationship r to construct an emotional knowledge graph. This layer uses the TransE model to represent the triplets in the emotional knowledge graph using distributed vectors, resulting in the emotional knowledge vector K^r_c, K^r_w and K^pos_w in Fig 1 (these three are the same vector). With K^r_c as an example, Fig 5 shows the generation process of the output vector .

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Generation process of output vector.

https://doi.org/10.1371/journal.pone.0325428.g005

In Fig 5, the character radical feature vector set is set as a Query vector for the multi head attention mechanism, and an emotional knowledge vector K^r_c is used as feature fusion for the corresponding Key vector and Value vector to obtain the knowledge enhanced feature output vector . The workflow is shown in equations (18) and (19):

(18)

(19)

Wherein, TransE(·) represents the TransE model, and MultiHead(·) is the multi head attention mechanism. Similarly, the generation process of the output vectors and shown in Fig 1 is similar to . The implementation steps are as follows:

1. 1. Knowledge Graph Embedding

The TransE model maps entities (e.g., emotional words) and relationships (e.g., intensity) from the knowledge graph into distributed vector representations.

1. 2. Multi-Head Attention Integration

These semantic vectors are dynamically aligned with BiGRU-derived features through a multi-head attention mechanism. Specifically:

Query: Feature vectors from BiGRU (e.g., contextual embeddings).

Key/Value: TransE-generated knowledge vectors (e.g., 愤怒 [anger] → 强度 [intensity]).

1. Output:The fused vectors encode both data-driven contextual semantics and knowledge-driven emotional relationships.

Result output layer

The result output layer is responsible for generating emotion recognition results, as shown in Fig 6.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Output schematic.

https://doi.org/10.1371/journal.pone.0325428.g006

The specific process is as follows: first, perform maximum pooling on the output vectors,, and, and then perform feature fusion through vector concatenation to obtain the fused feature vector V_y; Then, input V_y into the fully connected neural network and normalize it using the softmax function to obtain the probability output P; Finally, select the value with the highest probability as the emotion recognition result y. The workflow is shown in equations (20) to (22):

(20)

(21)

(22)

Wherein, Concat(·) represents the concatenation function, max(·) represents the maximum pooling operation, W and b represent the weight matrix and bias, softmax(·) represents the normalization function, and argmax(·) represents the probability maximization function.This process is described as follows:

1. 1. Pooling and Concatenation

Max pooling is applied to each feature vector (e.g.,,) to extract salient emotional signals.

The pooled vectors are concatenated into a unified representation (V_y).

1. 2. Classification

The concatenated vector V_y is fed into a fully connected neural network.

Activation: Softmax normalization computes probability distributions over emotion classes.

Prediction: The class with the highest probability is selected as the final output (e.g., “positive” or “negative”).

Radical and POS Feature Implementation

Radical Extraction: To capture the semantic-emotional significance of Chinese radicals, we developed a systematic pipeline:

1. 1. Xinhua Dictionary API Integration
   1. Each character (e.g., 怜 [“pity”]) was mapped to its radical (忄 [“heart”]) using the Xinhua Dictionary API.
   2. For 99.2% of common characters, radicals are successfully retrieved via the Xinhua Dictionary API (e.g., 怒 [“anger”] → 心 [“heart”]).
2. 2. Edge Case Handling
   1. Rare or complex characters (e.g., 龘 [“dá,” a rarely used Unicode character]) lacking API coverage were resolved via manual annotation or cross-referenced with the Unihan Database.
   2. For ambiguous cases (e.g., 赢 [“win”] containing multiple radicals), priority was given to the dominant semantic component (贝 [“shell/currency”]).

Example Workflow:

1. Input Character: 恨 (“hate”)
2. → Xinhua API Query → Radical: 忄 (“heart”)
3. → Emotional Label: Negative (intensity: 0.92)

POS Tagging with Emotion-Specific Labels: We extended the Jieba tokenizer to incorporate emotion-aware part-of-speech (POS) tags:

1. Step 1: Raw text segmentation using Jieba’s default lexicon.
2. Step 2: POS tagging with custom emotion labels:
   1. VA (affective verb, e.g., 热爱 [“adore”]),
   2. NA (negative adjective, e.g., 丑陋 [“ugly”]),
   3. PA (positive adverb, e.g., 快乐地 [“happily”]).
3. Step 3: Manual validation of ambiguous tags (e.g., 感动 [“moved”] tagged as VA instead of generic verb V).
4. Implementation: Custom dictionaries were merged with Jieba’s default lexicon to prioritize emotion-specific tags. For example, 愤怒 (“anger”) is tagged as NA (negative adjective) rather than NN (noun).

Dataset and experimental environment

For the experimental phase, we utilized two distinct datasets: the comprehensive NLPECC dataset and an in-house compilation, the Douban Movie dataset [23]. The NLPECC dataset encompasses a rich collection of 44,875 samples, categorized under six distinct emotional labels—ranging from likes and sorrows to dislikes, anger, happiness, and others. In our experimental design, this dataset was meticulously partitioned into training, validation, and testing subsets, adhering to a 6:2:2 ratio.

Our curated Douban Film Dataset is an aggregation of movie review excerpts extracted from Douban Films, encompassing five diverse films: “The Return of the Holy One,” “Charlotte’s Troubles,” “Captain America 3,” “Little Times 3,” and “July and Anson.” Initially compiled at approximately 300,000 entries, this dataset underwent a rigorous preprocessing regimen. This included data cleansing, character width standardization, punctuation normalization, case conversion of English letters, script simplification, and deduplication, yielding a refined dataset of roughly 250,000 commentaries. A detailed statistical breakdown of the preprocessed dataset is presented in Table 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Dataset information after preprocessing.

https://doi.org/10.1371/journal.pone.0325428.t001

As delineated in Table 1, the comments were stratified based on their star ratings: 1-star and 2-star comments were designated as negative, 3-star comments as neutral, and 4-star and 5-star as positive. However, the experimental inclusion of neutral comments was precluded due to the inaccuracies in their emotional labeling. This exclusion resulted in a dataset of 67,000 negative and 140,000 positive comments, revealing a distribution disparity. To mitigate this imbalance, we performed a random sampling of 50,000 comments from each category, curating a final experimental dataset encompassing 100,000 film review entries.

In the execution of our experiment, the dataset was meticulously segmented into a decade of parts, adhering to the 6:2:2 allocation for training, validation, and testing phases, respectively. To elaborate, the training subset was populated with 60,000 entries, evenly split between positive and negative comments. Similarly, both the validation and testing subsets were composed of 20,000 comments each, maintaining the balanced 50−50 distribution between the two sentiment categories. The experimental milieu, along with its specific configurations, is delineated in Table 2, providing a comprehensive overview of the tools and settings employed in this study.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Experimental environment and configuration.

https://doi.org/10.1371/journal.pone.0325428.t002

Hyperparameter settings and evaluating indicator

Table 3 elucidates the hyperparameter configurations and the metrics utilized for performance evaluation in our experiments, where epoch represents the number of iterations of the experimental dataset during training, batch_size is the number of samples input into the model during training, lr represents the learning rate of the model, hidden_dim represents the number of neurons in the hidden layer of the model, num_heads represents the number of heads in multi head attention, while dropout represents the dropout rate of model neurons during the training phase, used to avoid overfitting and improve the model’s generalization ability.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Hyperparameter settings.

https://doi.org/10.1371/journal.pone.0325428.t003

Our experimental evaluation hinged on the critical metrics of Precision (P), Recall (R), and the F1-score, which is the Harmonic Mean of the two, to quantitatively assess the efficacy of our emotion recognition model. The formulae governing these calculations are articulated in equations (23) through (25).

(23)

(24)

25)

Wherein, true positive (TP) represents the correctly classified positive sample, false positive (FP) represents the misclassified positive sample, and false negative (FN) represents the misclassified negative sample. And P represents the proportion of positive samples predicted correctly by the model to the samples predicted as positive, while R represents the proportion of positive samples predicted correctly by the model to the actual positive samples.

Comparative experiment

This article proposes a Chinese text sentiment analysis method called KEAMM, which integrates multi granularity semantic features driven by knowledge and data collaboration. It is compared with existing Chinese text sentiment analysis models to demonstrate the superiority of the proposed model. These benchmark methods are as follows:

• BiGRU [6]. This model uses bidirectional gated loop units to extract features from word vectors and achieve text sentiment analysis.
• DCCNN [5]. This model uses different channels for convolution operations, with one channel being a word vector and the other being a word vector. Emotional analysis is performed by fusing the features of the two channels.
• FG_ MCCNN [9]. This model comprehensively utilizes word vectors, word vectors, and the fusion of word vectors and part of speech to perform sentiment analysis on vectors through the multi-channel integration of CNN models.
• BERT BiLSTM [10]. This model uses the BERT model to construct word vectors, and then uses BiLSTM for feature extraction to achieve emotion recognition.
• BERT-MCNN [13]. This model is an emotion analysis model that combines the BERT model with a multi-layer collaborative convolutional neural network.

Ablation experiment

This article proposes the KEAMM sentiment analysis method using features such as characters, words, radicals, and parts of speech as inputs. In order to verify the effectiveness of each module in the KEAMM model, this article adjusts the model structure and conducts the following ablation experiments.

• Two BiGRUs. Use two BiGRU models to model word text and word text, concatenate their output vectors, and perform emotion recognition.
• Four BiGRUs. Use four BiGRUs to model word text, word level radical text, word text, and word level radical text, concatenate the four channels through the output vectors of BiGRU, and perform emotion recognition.
• Five BiGRUs. This model abandons the feature fusion of characters, words, radicals, and parts of speech in the KEAMM model through attention mechanism, and instead extracts features from character text, character level radical text, word text, word level radical text, and part of speech text through five BiGRUs, concatenates the output feature vectors of the five channels through BiGRU, and performs emotion recognition.
• Attention based multi granularity model (AMM). Discard the knowledge introduction part of the KEAMM model, that is, remove the feature fusion layer, concatenate the feature vectors output by the feature extraction layer, and perform emotion recognition.
• KEAMM. The model proposed in this article.

Result and analysis

Table 4 presents the experimental results of the above models on the experimental dataset.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Experimental results (%).

https://doi.org/10.1371/journal.pone.0325428.t004

After conducting a comprehensive analysis of the experimental results from the comparison and ablation experiments on the Douban Film dataset and NLPECC dataset (as shown in Table 4), the following conclusions can be drawn.

The BiGRU model, which relies exclusively on word-level features, achieved F1-scores of 85.63% and 77.70% on the Douban and NLPECC datasets, respectively. On the other hand, the Two BiGRUs model, a dual-channel model that incorporates both word and word features, achieved F1 values of 86.58% and 79.35% on the same datasets. Comparing these results, it can be observed that the Two BiGRUs model outperformed the BiGRU model by 0.95 and 1.65 percentage points, respectively.

These findings suggest that incorporating both word and word features in the Two BiGRUs model allows for the combination of their individual strengths, resulting in the extraction of more comprehensive semantic features that facilitate improved sentiment analysis.

To quantify the contribution of radical features, we conducted ablation studies (Table 5) on the NLPECC dataset:

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Ablation validation.

https://doi.org/10.1371/journal.pone.0325428.t005

Key findings:

Removing radical features caused a 1.21% F1 drop, underscoring their critical role in disambiguating sentiment (e.g., distinguishing 恨 [“hate”] from 狠 [“ruthless”] via radical 忄 vs. 犭).

POS features contributed a 0.87% F1 gain, validating their utility in identifying emotional carriers (e.g., VA verbs like 赞美 [“praise”]).

Technical Enhancements

Radical Coverage: Expanded to 98.5% of characters via Unihan integration.

POS Tag Accuracy: Achieved 92.3% agreement with human annotators after custom lexicon tuning.

Discuss

Upon examining the DCCNN and FG_MCCNN models, it’s clear that the latter, a three-channel CNN model, integrates a comprehensive set of features-characters, words, and part-of-speech - yielding higher F1-score of 86.92% and 80.47% respectively, compared to the dual-channel DCCNN model’s 86.53% and 79.32%. This disparity underscores the positive impact of part-of-speech features on emotion recognition performance, with the FG_MCCNN outperforming the DCCNN by 0.39 and 1.15 percentage points respectively.

In the ablation study, the Five BiGRUs model showed a slight increase in F1-score by 0.34 and 0.43 percentage points over the Four BiGRUs model, further substantiating the beneficial role of part-of-speech features in sentiment analysis.

Comparing the DCCNN and Two BiGRUs models, which both utilize word features but differ in their foundational models—CNN for DCCNN and BiGRU for Two BiGRUs—reveals a marginal superiority of the Two BiGRUs model, with F1-score of 86.58% and 79.35% respectively, surpassing the DCCNN by a minimal 0.05 and 0.03 percentage points. This suggests that while CNN models are adept at feature extraction, they might overlook certain semantic nuances of text, whereas BiGRU models, with their proficiency in capturing long-term dependencies, can extract richer semantic features, thus improving performance.

The Four BiGRUs model demonstrated its effectiveness with F1-score of 87.07% and 80.80%, outperforming both the Two BiGRUs and DCCNN models. Similarly, the Five BiGRUs model, with its incorporation of radical features, showed an enhancement in performance with F1-score of 87.41% and 81.23%, indicating the significance of these features in Chinese sentiment analysis.

The AMM model, leveraging the BiGRU and dot-product attention mechanism, achieved F1-score of 88.24% and 82.20%, an increase of 0.83 and 0.97 percentage points over the Five BiGRUs model. This highlights the advantage of feature fusion over mere concatenation, allowing for a deeper and more nuanced understanding of text semantics.

The BERT-BiLSTM and BERT-MCNN models, with their dynamic text vector representation, achieved F1- score of 88.25% and 82.99%, and 88.68% and 83.14% respectively, outperforming other models and coming close to the performance of our proposed KEAMM model. This illustrates the advantage of BERT’s capability to fine-tune semantic representations for sentiment analysis tasks.

Ultimately, the KEAMM model, integrating emotional knowledge vectors through a multi-head attention mechanism, achieved the highest F1-score of 89.23% and 84.84%, surpassing all other models. Compared to the AMM model, the KEAMM model showed an increase of 0.99 and 2.64 percentage points, and when compared to the BERT-BiLSTM and BERT-MCNN models, it demonstrated an increase of 0.98 and 1.85 percentage points, and 0.55 and 1.7 percentage points respectively. This underscores the importance of incorporating explicit emotional knowledge in enhancing the accuracy of sentiment analysis models.

In summary, the comparative analysis of various models reveals the KEAMM model’s superiority in Chinese sentiment analysis, highlighting the value of integrating emotional knowledge and multi-granularity features for improving performance.

Dataset bias and domain generalization

The current model is primarily trained and validated on movie review datasets (e.g., Douban Film Reviews and NLPECC). While the dataset provides rich emotional expressions, it inherently reflects the linguistic patterns and vocabulary specific to entertainment contexts. Consequently, the model may struggle to generalize to specialized domains such as healthcare or finance, where sentiment expressions often involve technical jargon and implicit emotional cues (e.g., patient feedback or investment risk assessments). For instance, terms like “疗效” (“treatment efficacy”) in medical reviews or “波动” (“market volatility”) in financial texts carry domain-specific sentiment nuances that are underrepresented in our training data. Future work should explore cross-domain adaptation techniques, such as domain-adversarial training or few-shot learning, to enhance the model’s versatility. Additionally, curating high-quality datasets from diverse domains (e.g., clinical narratives, stock market commentaries) will be critical for robust evaluation.

Static radical extraction

Our radical feature extraction relies on predefined mappings from the Xinhua Dictionary, which introduces two key limitations:

Coverage Gaps: Rare or dialect-specific characters (e.g., 龘 [“dá,” a complex Unicode character]) may lack radical annotations in static dictionaries.

Context Insensitivity: The current approach treats radicals as fixed semantic units, ignoring dynamic compositional semantics. For example, the character 愁 (“sorrow”) contains the radical 心 (“heart”), but its emotional valence can shift depending on context (e.g., 愁绪 [“melancholy”] vs. 愁容 [“worried expression”]).

To address this, future research should investigate dynamic radical parsing using deep learning models (e.g., convolutional networks for stroke pattern recognition) or hybrid approaches combining dictionary-based rules with contextual embeddings. Integrating resources like the Unihan Database—which provides radical decompositions for over 90,000 CJK characters—could further enhance coverage.

Computational efficiency

The BiGRU-based architecture, while effective in capturing sequential dependencies, incurs significant computational overhead for long texts (e.g., reviews exceeding 500 tokens). On the NVIDIA Tesla K80 GPU, KEAMM processes 100,000 tokens in approximately12 seconds, which may hinder real-time applications. To mitigate this, we propose three directions:

Lightweight Architectures: Replace BiGRU layers with efficient variants like TinyBERT or MobileBERT, which reduce parameter counts by 60–80% while retaining performance.

Model Compression: Apply knowledge distillation to transfer KEAMM’s knowledge to a smaller student model.

Hardware Optimization: Leverage FPGA acceleration or quantization-aware training to expedite inference on edge devices.

Dialect and informal language

The model currently focuses on standard Mandarin texts, neglecting regional dialects (e.g., Cantonese) and internet slang (e.g., “栓Q” [“thank you” in Mandarinized English]). These forms of communication are prevalent in social media but pose challenges due to non-standard orthography and cultural specificity. Future work should incorporate dialectal corpora (e.g., Weibo Cantonese posts) and adopt code-switching detection mechanisms to improve robustness. Future Work Roadmap is in Table 6.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 6. Future work roadmap.

https://doi.org/10.1371/journal.pone.0325428.t006

Broader implications

Cross-Linguistic Relevance: The radical-centric paradigm extends to other logographic systems (e.g., Japanese Kanji’s 心 [“kokoro”] in 怒 [“ikari,” anger]).

Adversarial Robustness: KEAMM exhibits resilience to noise, with only a 1.2% F1 drop under synonym substitution attacks, outperforming baseline models by ~3%.

Future directions

Domain Adaptation: Extending KEAMM to specialized domains (e.g., healthcare, finance) via domain-adversarial training and curated corpora.

Dynamic Radical Parsing: Replacing static dictionaries with stroke-based CNN models to handle rare characters (e.g., 龘) and dialectal variations.

Syntactic Integration: Incorporating dependency trees via graph convolutional networks (GCNs) to model syntactic-semantic interactions (e.g., negation phrases like 不愉快 [“unpleasant”]).

Lightweight Deployment: Optimizing inference speed through TinyBERT distillation and hardware-aware quantization.

Societal impact

KEAMM’s context-aware sentiment detection holds transformative potential across industries:

Entertainment: Real-time analysis of film reviews to gauge audience reactions.

Healthcare: Monitoring patient feedback for emotional undertones in clinical narratives.

Finance: Detecting market sentiment shifts in investor commentaries.

By unifying radical semantics, knowledge graphs, and deep learning, KEAMM establishes a new benchmark for Chinese sentiment analysis. Future work will expand its linguistic, technical, and ethical horizons, ensuring robust and equitable applications in an increasingly data-driven world.