HTCNN-Attn: a fine-grained hierarchical multi-label deep learning model for disaster emergency information intelligent extraction from social media

Three-level hierarchical labeling architecture for disaster information

Different levels of emergency response organizations focus on varying levels of granularity in disaster information. Government organizations are more concerned with high-level information, such as determining whether a social media post is disaster-related. Emergency management departments focus on different categories of disaster awareness, such as aid requests, conditions of disaster victims, and so on. Ground-level emergency management action departments pay more attention to finer-grained categories, such as water, food, safety, transportation, medical assistance, etc.

To address these hierarchical needs, we designed a three-level labeling architecture for disaster information on social media, combining the existing categories in the Appen disaster response dataset with emergency management needs, as shown in Fig. 1. The first-level label consists of a binary classification to filter disaster-related information. The second-level labels consist of five categories: aid requests, conditions of disaster victims, donation requests, disaster, and facility conditions. The third-level labels correspond to the second-level labels and further subdivide into nine, four, five, nine and eight categories, totaling 35 labels.

Disaster perception framework

• (1)

  The disaster perception framework based on social media mainly consists of two processes: the model training process (represented by solid lines) and the real-time data classification and dispatch process (represented by dashed lines), as shown in Fig. 2.

• (2)

  Training data acquisition: Crawling disaster-related tweets from social media platforms such as Twitter or Weibo using web crawlers after a disaster occurs.

• (3)

  Training data preprocessing: Removing noise data, handling missing values, and processing duplicate content, etc. Text normalization is also carried out, including removing special characters and converting to lowercase.

• (4)

  Training data labeling: The training data is labeled with a three-level tag system for disaster-related information on social media.

• (5)

  Feature extraction for training: The text data is transformed into more informative feature representations.

• (6)

  Training the classification model: The extracted features are input into the model as word embeddings for training a hierarchical multi-label classification model. The model is then evaluated.

• (7)

  Real-time data acquisition: After a disaster occurs, a real-time web crawler is launched to collect relevant tweets.

• (8)

  Real-time data preprocessing: This process is the same as step 2.

• (9)

  Real-time data feature extraction: This process is the same as step 4.

• (10)

  Model inference: The real-time data features are input into the trained model for inference and classification.

• (11)

  Classification result dispatch: Based on the first-level tag classification result, information related to the disaster is filtered out. The second-level classification results are dispatched to the corresponding emergency response departments, and the third-level sub-classification results are sent to the next level of emergency response departments.

HTCNN-Attn hierarchical multi-label classification model

We have built the HTCNN-Attn hierarchical multi-label classification model, whose architecture is illustrated in Fig. 3. It consists of two main parts: feature extraction and the hierarchical fusion multi-label classification model.

Feature extraction

Before inputting social media tweets into a deep learning model, it is necessary to first extract meaningful features. Commonly used feature extraction methods include:

Bag of Words (BoW) model: The text is divided into individual words or phrases, and then the frequency of each word in the text is counted to form a vector representation. This method ignores the order and semantic information of the words.

TF-IDF: This method combines the concepts of term frequency (TF) and inverse document frequency (IDF). TF measures the importance of a word in a specific text, while IDF measures the importance of a word in the entire corpus. By multiplying TF and IDF, we obtain the TF-IDF value for each word as a feature. However, this method cannot capture word order information, ignores semantic similarity between words, and does not account for document length or the higher weight of frequent words.

N-gram model: The text is divided into sequences of continuous N words. This can be unigrams (single words), bigrams (two consecutive words), trigrams (three consecutive words), etc. The N-gram model can capture local relationships between words. However, it cannot handle long-distance dependencies between words and requires dealing with a large number of sparse features.

Word2Vec model: Based on a neural network, this is a word embedding model that maps each word to a continuous vector space (Li et al., 2021). Word2Vec represents semantic relationships between words as distances or similarities between vectors. However, its word vectors are static—i.e., the vector for each word is fixed after training and does not change based on context or tasks. Therefore, it lacks the ability to dynamically adapt to the context of the text. As a result, it is not suitable for handling polysemy or complex semantic relationships between word combinations.

Transformer model (Vaswani et al., 2017): The Transformer model, using multi-head self-attention mechanisms and feedforward neural network layers, is highly effective at capturing contextual and semantic information in long texts. It also offers excellent parallelism when handling sequence-based tasks.

Therefore, after a disaster occurs, we use distributed crawlers to collect social media text data and extract features based on the Transformer model. The process is illustrated in Fig. 4, and the specific steps are as follows:

• (1)

  Calculating positional encoding.

  In the input sequence’s word embedding vectors, positional information is injected to capture the sequential order in the text. The calculation is shown in Eqs. (1) and (2). (1) $$\mathrm{P}{\mathrm{E}}_{2\mathrm{i}}\left(\mathrm{p}\right)=\mathrm{s}\mathrm{i}\mathrm{n}\left(\mathrm{p}/{10000}^{2\mathrm{i}/{\mathrm{d}}_{\mathrm{p}\mathrm{o}\mathrm{s}}}\right)$$ (2) $$\mathrm{P}{\mathrm{E}}_{2\mathrm{i}+1}\left(\mathrm{p}\right)=\mathrm{c}\mathrm{o}\mathrm{s}\left(\mathrm{p}/{10000}^{2\mathrm{i}/{\mathrm{d}}_{\mathrm{p}\mathrm{o}\mathrm{s}}}\right).$$

• (2)

  Calculating self-attention mechanism.

  The text data is quantified into features F, and the output and weight matrices ${\mathrm{W}}^{\mathrm{Q}},{\mathrm{W}}^{\mathrm{K}},\mathrm{a}\mathrm{n}\mathrm{d}{\mathrm{W}}^{\mathrm{V}}$ are calculated separately to increase the inter-feature relationships and enhance the generalization of the features. The calculation process is shown in Eqs. (3)–(5). (3) $$\mathrm{Q}=\mathrm{F}{\mathrm{W}}^{\mathrm{Q}}$$ (4) $$\mathrm{K}=\mathrm{F}{\mathrm{W}}^{\mathrm{K}}$$ (5) $$\mathrm{V}=\mathrm{F}{\mathrm{W}}^{\mathrm{V}}.$$

• (3)

  Calculating the output feature Z of the self-attention layer.

  The values after softmax normalization are multiplied by the V vector matrix, and the weighted vectors are summed to produce the self-attention output for that position. The calculation process is shown in Eq. (6). (6) $$\mathrm{Z}=\mathrm{S}\mathrm{o}\mathrm{f}\mathrm{t}\mathrm{m}\mathrm{a}\mathrm{x}\left({\displaystyle \frac{\mathrm{Q}{\mathrm{K}}^{\mathrm{T}}}{\sqrt{{\mathrm{d}}_{\mathrm{k}}}}}\right)\mathrm{V}$$where ${\mathrm{d}}_{\mathrm{k}}$ is the dimension of the model.

• (4)

  Calculating the output Matrix ${\mathrm{Z}}_{\mathrm{m}\mathrm{u}\mathrm{l}}$ of the multi-head self-attention layer.

  Where H denotes the number of attention heads, and ${\mathrm{Z}}_{\mathrm{i}}$ represents the (i + 1)-th attention head. [ $\cdots$] represents the concatenation of the H attention heads, and ${\mathrm{W}}^{\mathrm{O}}$denotes the concatenated matrix. Summing and layer normalization are performed, with layer normalization denoted by LN. The specific process is shown in Eqs. (7) and (8). (7) $${\mathrm{Z}}_{\mathrm{m}\mathrm{u}\mathrm{l}}=\left[{\mathrm{Z}}_0,{\mathrm{Z}}_1,\dots ,{\mathrm{Z}}_{\mathrm{i}}\right]{\mathrm{W}}^{\mathrm{O}}$$ (8) $${\mathrm{Z}}_{\mathrm{m}\mathrm{u}\mathrm{l}}=\mathrm{L}\mathrm{N}\left({\mathrm{Z}}_{\mathrm{m}\mathrm{u}\mathrm{l}}+\mathrm{F}\right).$$

• (5)

  Output features.

  The Z_mul is passed to the feedforward neural network, where it undergoes a fully connected layer and a nonlinear transformation using an activation function to obtain the final output features. Afterward, summing and layer normalization are applied again to further enhance the representational power and stability of the features.

Hierarchical fusion multi-label classification model

In traditional multi-label classification problems, each sample can be assigned multiple labels, typically using a flat structure, as shown in Fig. 5. In this problem, we propose a hierarchical multi-label tree structure, as shown in Fig. 6. In this tree structure, higher-level labels represent more general categories, while lower-level labels represent more specific subcategories. There is a hierarchical relationship between the labels.

We designed and implemented a hierarchical multi-label classification model by integrating CNNs to fuse hierarchical information. Figure 7 illustrates the workflow of the proposed model. This model was designed to classify three hierarchical categories (cat_1, cat_2, and cat_3).

First, the input layer (main_input) accepts text sequences as a 1D tensor of shape (max_len,). The extracted feature vectors are then used for word embedding operations. The embedding layer converts integer word indices into dense vector representations. Next, three convolutional layers—cnn1, cnn2, and cnn3—are defined for three different convolutional window sizes (3, 4, 5), followed by max-pooling operations. Convolution operations are performed in all three layers, with the calculation process shown in Eqs. (9)–(11).

(9) $${\mathrm{r}}_{\mathrm{i}}=\mathrm{r}\mathrm{e}\mathrm{l}\mathrm{u}\left({\mathrm{W}}_{\mathrm{l}\mathrm{k}}\bullet {\mathrm{E}}_{\mathrm{i}:\mathrm{i}+{\mathrm{h}}_{\mathrm{l}}-1}+\mathrm{b}\right)$$

(10) $$\mathrm{R}=\left[{\mathrm{r}}_1,{\mathrm{r}}_2,...,{\mathrm{r}}_{\mathrm{n}-{\mathrm{h}}_{\mathrm{l}}+1}\right]$$

(11) $${\mathrm{M}}_{\mathrm{l}}=\left[{\mathrm{R}}_1,{\mathrm{R}}_2,...,{\mathrm{R}}_{\mathrm{K}}\right].$$

Here, ${\mathrm{W}}_{\mathrm{l}\mathrm{k}}$ represents the weights of the k convolutional kernels at the l-th layer. ${\mathrm{r}}_{\mathrm{i}}$ represents a new feature of a word vector sequence, b is the bias term, and ReLU is the activation function. After passing through the convolutional kernels, the word vector features are transformed into a feature map ${\mathrm{M}}_{\mathrm{l}}$ at the l-th layer, which consists of multiple feature vectors.

The attention weights for each word vector in the feature map ${\mathrm{M}}_{\mathrm{l}}$ are computed as follows, with the calculation process described in Eqs. (12)–(15).

(12) $$a\nu g=mean\left(M_l\right)$$

(13) $$\mathrm{m}\mathrm{a}\mathrm{x}=\mathrm{M}\mathrm{a}\mathrm{x}\left(M_l\right)$$

(14) $$A\mathrm{t}\mathrm{t}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}=sigmoid\left(W_t\bullet concat\left(a\nu g,\mathrm{m}\mathrm{a}\mathrm{x}\right)\right)$$

(15) $$M_l=M_l\bullet attention.$$

Here, ‘avg’ denotes the average across the channel dimension of ${\mathrm{M}}_{\mathrm{l}}$, while ‘max’ indicates the maximum value along the channel dimension of ${\mathrm{M}}_{\mathrm{l}}$. We concatenate the average and maximum values and then multiply them by a weight matrix with a kernel size of t. After applying the sigmoid activation function, we obtain the attention weights for each word. Next, we perform max pooling on each feature vector ${\mathrm{R}}_{\mathrm{i}}$ of ${\mathrm{M}}_{\mathrm{l}}$, generating a new feature representation ${\mathrm{F}}_{\mathrm{l}}$. After passing through a fully connected layer, we obtain the classification results for layer l. The calculation process is described in Eqs. (16)–(18).

(16) $$\stackrel{\wedge }{r_i}=\mathrm{m}\mathrm{a}\mathrm{x}\mathrm{\_}pool\left(R_i\right)$$

(17) $$F_l=\left[\stackrel{\wedge }{r_1},\stackrel{\wedge }{r_2},...,\stackrel{\wedge }{r_K}\right]$$

(18) $$F_{out}=linear\left(F_l\right).$$

Experimental data

We utilize a publicly available disaster response dataset from Appen and HumAID. The Appen dataset contains 26,248 labeled messages sent during various historical disasters worldwide, including those in Chile and Pakistan. HumAID consists of 77,196 annotated tweets that have been collected during 19 major natural disaster events, including earthquakes, hurricanes, wildfires, and floods, which happened from 2016 to 2019 across different parts of the World. Appen’s data is annotated with one or more of 36 categories. HumAID’s data is annotated with 11 categories, including caution and advice, displaced people and evacuations, don’t know cant judge, infrastructure and utility damage, injured or dead people, missing or found people, not humanitarian, other relevant information, requests or urgent needs, rescue volunteering or donation effort, sympathy and support.

Data preprocessing

We integrated the Appen and HumAID datasets. After cleaning (removing duplicates and invalid entries), we reorganized labels into a three-level structure. Table 2 illustrates the hierarchization process of a single message.

To address data imbalance, we balanced the dataset while preserving its hierarchical structure, as illustrated in Fig. 8. At the first level, we applied stratified sampling to balance the “Disaster-related” (58%, 27,144 samples) and “Non-disaster” (42%, 19,656 samples) categories, resulting in a total of 46,800 balanced samples without discarding hierarchical relationships. For the second and third levels, we employed a hybrid approach. The Synthetic Minority Over-sampling Technique (SMOTE) was used to generate synthetic samples for underrepresented subcategories, such as “Medical Help” (increased from 200 to 700 samples in the balanced data) and “Other Infrastructure” (increased from 500 to 944 samples). Random undersampling reduced overly dominant subcategories, e.g., “Shelter” (decreased from 2,000 to 820 samples) and “Water” (decreased from 4,500 to 1,180 samples), to avoid skewed model training. During training, a class-weighted loss function was applied, assigning higher weights to rare labels. For example, in the third-level subcategories, labels like “Medical Help” (700 samples) and “Other Weather” (480 samples) received greater weight in loss calculations compared to dominant labels like “Flood” (870 samples) and “Refugees” (1,500 samples).

Then we cleaned the data, which included:

• (1)

  Basic text normalization: removing special characters, punctuation, converting words to lowercase, removing mentions, hashtags, URLs, and replacing contractions. We substituted numbers with words, removed non-ASCII characters, and extra punctuation.

• (2)

  User identity anonymization: All social media account identifiers (e.g., @usernames, profile URLs) were irreversibly hashed using SHA-256 with dataset-specific salts to prevent re-identification.

• (3)

  Geolocation obfuscation: Precise GPS coordinates were generalized to city-level regions through coordinate truncation (2 decimal places ≈ 1.1 km accuracy) and GeoNames API mapping, ensuring GDPR compliance.

• (4)

  PII redaction: We employed a rule-based NLP pipeline to detect and remove personally identifiable information (PII) such as phone numbers, email addresses, and physical addresses. Additionally, named entities (e.g., individuals’ full names) identified via spaCy’s NER model were masked with generic placeholders (e.g., “[NAME]”).

The study utilized only publicly available social media data and adhered to the platform’s terms of service. No private or sensitive user content was retained after preprocessing.

Model training and parameter settings

Experiments were conducted on a Tesla T4 GPU running Linux. The preprocessed data were fed into the HTCNN-Attn model for training under the PyTorch 1.8.1 framework. To ensure optimal performance and reproducibility, we conducted systematic hyperparameter tuning using grid search on a validation subset (10% of the training data). Table 3 summarizes the experimental validation process for critical parameters. For CNN kernels, comparative testing of 128, 256, and 512 units revealed diminishing returns beyond 256 kernels, where the configuration achieved a micro-F1 of 0.740 vs. 0.738 with 512 kernels, while reducing training time by 34%. The learning rate was optimized to 2e−6 through stability analysis—this value maximized classification performance (micro-F1: 0.740) while preventing the overshooting observed at 5e−6, which caused a 3.8% micro-F1 degradation. Transformer depth was constrained to four layers after determining that six-layer architectures provided only marginal gains (+0.5% micro-F1) at a 15% increased computational cost. Batch size optimization was settled at 128 to balance GPU memory limitations of the Tesla T4 platform with stable gradient estimation. As shown in Table 3, these parameters collectively maximize classification accuracy while maintaining computational efficiency critical for real-time deployment.

Final parameters were set as: epoch = 50, batch_size = 128, word_embedding_dim = 128, convolution_kernels = 256 (three windows of 3/4/5), learning_rate = 2e–6, dropout = 0.1, activation = Sigmoid, optimizer = Adam, and hierarchical cross-entropy loss.

Model evaluation

Figure 9 shows the accuracy of the three hierarchical levels over training epochs. From the line plots, we observe that the accuracy at the first level remains relatively stable at the beginning of training, and then gradually increases and stabilizes. The accuracy curve shows a steady upward trend until converging to a relatively high level. This suggests that after sufficient training, the first-level model can gradually learn better feature representations, thereby improving classification accuracy.

The accuracy at the second level increases rapidly during the early training stages and then gradually stabilizes. The accuracy curve shows a rapid rise followed by saturation. This indicates that, building on the first-level features, the second-level model can further extract richer representations and thereby significantly enhance classification accuracy.

The third-level accuracy also shows a gradual increase followed by stabilization. The accuracy curve steadily rises and eventually converges to a relatively high level. This suggests that, inheriting the features from the previous two levels, the third-level model leverages the self-attention mechanism to further enhance feature representation, achieving higher classification accuracy.

Figure 10 illustrates the model loss across the three hierarchical levels and the total loss over training epochs, highlighting the optimization process at each stage. The Level 1 loss starts relatively low and decreases rapidly during the initial epochs, reflecting the simplicity of the binary classification task at this level. By the later epochs, the loss stabilizes near zero, indicating that the model effectively learns high-level representations. At Level 2, the loss starts at a moderately higher value but decreases significantly in the early epochs. It stabilizes around epoch 20, demonstrating the model’s ability to leverage Level 1 features for optimizing intermediate classification. The Level 3 loss, dealing with the most fine-grained and complex categories, starts at the highest value and decreases more gradually, stabilizing after sufficient training, indicating the model’s capacity to capture nuanced patterns using hierarchical attention mechanisms. The total loss, aggregating all levels, exhibits a steep decline in the early epochs followed by a gradual stabilization, reflecting the overall convergence and robustness of the HTCNN-Attn model. These trends, combined with the accuracy improvements shown in Fig. 9, demonstrate the model’s effectiveness in progressively refining representations and achieving logical consistency across the hierarchical levels.

To validate the HTCNN-Attn model, we compared it with five baselines (TF-IDF-Logistic Regression, TF-IDF-Naive Bayes, ALBERT-BiLSTM, ALBERT-BiGRU, stella_en_400M_v5) using 30 independent trials, reporting metrics as mean ± SD (Table 4). A one-way ANOVA revealed significant overall differences in classification metrics (Accuracy: F(5, 174) = 28.7, p < 0.001; Micro-F1: F(5, 174) = 19.5, p < 0.001) and computational efficiency (Training Time/Memory: Kruskal-Wallis H = 89.3/92.1, p < 0.001). Post-hoc Tukey’s HSD tests showed HTCNN-Attn significantly outperformed all baselines in Accuracy (e.g., +1.25% vs. stella_en_400M_v5, q(6, 174) = 8.10, p < 0.001, Cohen’s d = 0.87) and micro-F1 (e.g., +1.87% vs. stella_en_400M_v5, q(6, 174) = 6.50, p < 0.01), with the largest improvements against traditional models (e.g., +4.38% Accuracy vs. TF-IDF-Naive Bayes). For efficiency, it required 210 ± 15 min training time and 0.78 ± 0.05 GB memory, significantly faster and more memory-efficient than deep learning baselines (e.g., −26% time/−52% memory vs. stella_en_400M_v5, p < 0.001). Shapiro-Wilk and Levene’s tests confirmed normality and variance homogeneity (all p > 0.05), validating parametric tests. Results collectively demonstrate HTCNN-Attn’s robust performance and efficiency in hierarchical disaster information extraction.

Error analysis

The hierarchical multi-label nature of the task introduces unique challenges for error analysis. Unlike traditional single-label classification, each social media post can be assigned multiple third-level labels (e.g., “Medical Help” and “Medical Products” simultaneously), rendering standard confusion matrices inappropriate. Given the complexity of 35 fine-grained third-level categories, we focus on the top 8 most frequent error pairs (Fig. 11) to ensure clarity while capturing dominant error patterns.

The model exhibited significant confusion between semantically overlapping categories. For instance, “Cold” was frequently misclassified as “Storm” (78 errors, 29.1% rate), often due to contextual ambiguity in posts like “freezing temperatures during the storm”. Similarly, “Medical Help” and “Medical Products” showed bidirectional errors (63–78 errors, 23.6% rate) when posts mentioned both “doctors” and “medicines”. Rare categories with limited training data, such as “Military Aid” (480 samples), suffered reduced recall (0.55).

To address these issues, we propose (1) refining ambiguous labels (e.g., splitting “Other Aid” into subcategories), (2) injecting domain-specific lexicons (e.g., meteorological terms) to enhance semantic distinction. These adjustments directly target the observed error patterns while maintaining computational efficiency.

Ablation study

To thoroughly investigate the contribution of each component within our HTCNN-Attn model, we conducted a series of ablation experiments. These experiments involved systematically removing or modifying specific modules to assess their impact on the overall performance. We used the full HTCNN-Attn model as the baseline, which integrates hierarchical information, a Transformer encoder, and a CNN layer for feature refinement. The metrics considered include Accuracy, Micro_F1, Macro_F1, P@1, P@3, and hierarchical consistency. Table 5 presents the results of the ablation study. The experimental process is structured as follows:

• (1)

  No-Hier: We remove the hierarchical structure modeling module, treating the classification as a flat multi-label problem. This tests the importance of leveraging parent-child relationships between labels.

• (2)

  No-Transformer: We eliminate the Transformer encoder and rely solely on CNN (and embeddings), thereby assessing the role of self-attention and global context modeling.

• (3)

  No-CNN: We remove the CNN layer to determine its contribution to capturing local n-gram patterns and complementing the Transformer’s global attention.

• (4)

  No-Pretrained-Embeddings: We replace ALBERT or other pre-trained embeddings with randomly initialized embeddings to gauge the value of external language knowledge.

• (5)

  No-Attention-for-Hierarchy: we remove it to understand how it affects the alignment between label hierarchies and model predictions.

The study shows that the hierarchical structure is crucial for maintaining consistent predictions across different levels of the label taxonomy. When this module is removed, the hierarchical consistency drops by approximately 14.36%. Removing the Transformer encoder leads to a decrease of about 3.82% in micro-F1 and 2.87% in Macro-F1. Taking away the CNN layer causes P@1 and P@3 to drop by about 1.87% and 1.07%, respectively. Discarding pre-trained embeddings leads to a significant decline in all metrics, with hierarchical consistency dropping by about 6.06%. Removing the hierarchical attention mechanism results in a 1.9% decrease in hierarchical consistency. Each component of the HTCNN-Attn model contributes significantly to its overall effectiveness, jointly enabling the model to perform better.

Generalization experiments

To assess the cross-event generalization ability of HTCNN-Attn, we conducted experiments on the CrisisBench dataset (Firoj Alam et al., 2021), focusing on the 2012 Hurricane Sandy, a disaster event not included in our original training dataset, HumAID. This choice allows us to rigorously test the model’s adaptability to unseen scenarios.

The data processing steps followed the methods described in “Preprocessing”, with additional normalization for event-specific language features. We fine-tuned HTCNN-Attn using an AdamW optimizer with an initial learning rate of 0.0001 for 20 epochs. A stratified 70/15/15 split was applied across five independent trials to ensure balanced data representation.

As shown in Table 6, HTCNN-Attn outperformed baseline models, achieving an accuracy of 0.935, a micro-F1 score of 0.702, and an HC score of 0.678. Compared to ALBERT-BiGRU, it improved micro-F1 by 4.2% and HC by 5.5% (both p < 0.05). Although the micro-F1 score decreased by 3.7% relative to HumAID, the stable HC score demonstrates the model’s robust hierarchical reasoning ability across different disaster events, validating its generalization performance.

Deployment and inference

Taking the Nepal earthquake that occurred on April 25, 2015, at 14:11 as an example, we collected post-earthquake tweets from Twitter. We utilized the Flask framework to build a system that loads the trained model. By inputting a tweet into the system, accurate classification was achieved. Furthermore, by integrating the location information of the tweet with the extracted category information, we were able to enhance the perception of disaster-related information, as illustrated in Fig. 12.

To address real-time deployment requirements, we compared model inference efficiency (Table 7). Traditional Term Frequency-Inverse Document Frequency (TF-IDF) baselines are fast (7.2–8.5 ms latency, 118–139 samples/sec) but lack hierarchical text modeling, limiting accuracy for complex disaster texts. Sequence models like ALBERT-BiGRU offer moderate speed (17.0 ms, 59 samples/sec) but require high GPU memory (4.8 GB), challenging edge use. The large MTEB baseline stella_en_400M_v5, despite strong semantics, is inefficient (22.5 ms latency, 6.3 GB memory), unsuitable for resource-constrained environments. Our HTCNN-Attn balances speed and efficiency with 12.0 ms latency (29% faster than ALBERT-BiGRU), 83 samples/sec throughput (meeting disaster data peaks), and low memory (3.5 GB, 44% less than stella_en_400M_v5). These metrics validate its readiness for real-time emergency response, proving superior efficiency while maintaining hierarchical classification accuracy.