This study uses two dependable datasets: CrisisNLP [35] and CrisisLex [36]. These data sources include annotated tweets on a variety of real-world tragedies. A total of 33,370 tweets connected to disaster and non-disaster messages were collected. Task 1 is a binary classification, in which entire tweets are divided into two classes: disaster (Target = 1) and non-disaster (Target = 0), as depicted in Table 2. Furthermore, Task 2 involves creating a dataset that contains 15,435 tweets pulled from the dataset used in Task 1, presented in Table 3. After a tweet has been selected as disaster-related in Task 1, then it moves to Task 2 to be further categorized as a particular disaster type, i.e., earthquake or flood. Nonetheless, a balanced subset of non-disaster tweets is inserted to enhance textual discrimination. This approach avoids the model confusing general tweets with disaster-related tweets that might bear similar linguistic indicators. In addition, in Task 1, few tweets can be mistakenly registered as disaster-related (false positives). Therefore, their inclusion in Task 2 will enable the framework to reevaluate the ambiguous cases and ensure that it is robust enough to separate the real disaster-related and irrelevant text. As a result, Task 2 comes with earthquake, flood, and non-disaster categories in order to improve classification accuracy and model extrapolation. Consequently, Task 2 comprises three approximately balanced groups: earthquake (5779 tweets), flood (4873 tweets), and non-disaster (4783 tweets).

The resulting balanced multiclass dataset is organized and distributed, as shown in Table 3. Task 3 is based on 5779 earthquake-related tweets obtained as a result of Task 2, which is labelling them within aid categories, including food, medical help, infrastructure, and utilities, as presented in Table 4. Additional categories, i.e., sympathy, and support, and general information, are non-aid content commonly noticed during actual disaster communication. These categories assist the model in differentiating actionable requests from emotional expressions or general discourse. Several examples of relevant tweets are in Table 4. To protect the privacy of the users, usernames, URLs, locations, and hashtags in all the tweets were anonymized by substituting them with generic tokens. Tasks 1 and 2 include Tweet_ID, Tweet_text, and Target, whereas Task 3 includes an additional feature, Help, which identifies the kind of help that is requested.

Before extracting features from the data, we performed preprocessing in the following steps: Text normalization: For consistency, all tweet text is converted to lowercase. To reduce noise that could negatively impact model performance, punctuation marks, special characters, and numerical values are removed. Tokenization: Tokenization involves splitting each tweet into individual units, referred to as tokens. This step breaks sentences into smaller components, allowing the model to analyze textual data more effectively. Avoiding Stop Words: Common words such as “and”, “the”, and “is” which do not contribute meaningful information for tweet classification, are removed. Eliminating these terms enables the model to concentrate on more informative and discriminative words. Word Stemming: To normalize variations of the same word, the Porter stemming algorithm is applied. This process reduces words to their root forms (e.g., “flooding” to “flood”), helping the model capture semantic similarity across different word inflections. Table 5 presents examples of original tweets alongside their cleaned versions and corresponding target labels.

Table 6 shows a comparison of tweets before and after the cleaning process, along with their lengths. It clearly shows the importance of removing unnecessary features like, links, symbols, and special characters, which make tweets shorter and more organized. Tweets containing numerical values may include useful information; however, such values often appear inconsistently or in unusual contexts, which can confuse the model. Therefore, numerical values were replaced with a generic token to preserve their presence while improving textual consistency. The Porter stemming algorithm was selected because it is simple to implement and performs well on short, informal Twitter text. Although lemmatization preserves more semantic meaning, it requires accurate part-of-speech tagging, which is difficult to achieve with noisy and unstructured Twitter data. Consequently, stemming provides a practical and efficient alternative.

The textual data undergoes a pre-processing and cleaning; then, data are converted into numerical format that can be used in machine learning. In order to reflect the contextual relevance and meaning of words, this research opts to apply the TF-IDF methodology as a hybrid embedding model with Graph Convolutional Networks (GCN). TF-IDF is a measure of the level of significance of a word within a particular tweet in comparison to the entire corpus. It emphasizes distinctive and informative terms while reducing the impact of frequent but less meaningful words. Term Frequency (TF) is the frequency of the occurrence of a word within a tweet whereas the Inverse Document Frequency (IDF) is the frequency of a word in all the tweets. Mathematically, TF-IDF is written as: TF−IDF(ti,dj)=TF(ti,dj)xIDF(ti)where TF(ti,dj)=fi,j∑kfk,jand IDF(ti)=log⁡(Ndfi+1)

Here f(i,j) denotes the raw frequency of the term ti in tweet dj, dfi represents the number of tweets in which the term ti occurs, and N is the total number of tweets. This representation allows the model to emphasize informative words, thereby improving its ability to identify critical disaster-related content. As a result, each tweet is represented as a fixed-length sparse vector, where non-zero entries correspond to highly ranked terms and their associated relevance scores. The top twenty terms in the dataset are illustrated in Fig. 1, which is based on the average TF-IDF scores assigned to each word across the entire corpus. Examples of terms with higher average TF-IDF values include “nepal,” “earthquake,” and “rubyph,” indicating their strong association with tweet-specific content. The objective of this analysis is to provide initial insight into prominent topics by highlighting statistically significant terms that effectively distinguish tweets within the dataset. The selection parameter k=20 was determined empirically through experimentation with values of k=20,50,100, and 200 features. Increasing k beyond 20 resulted in a classification accuracy improvement of less than 1%, while considerably increasing model complexity and training time. Given the limited length of tweets, a compact set of representative terms was sufficient to capture meaningful vocabulary patterns. Consequently, selecting k=20 achieves a suitable balance between accuracy, computational efficiency, and representational adequacy. Sensitivity analysis confirms that the reduced feature space retains essential lexical information without introducing redundancy.

While TF-IDF effectively highlights the importance of individual words, it does not account for word co-occurrence patterns, which may result in a loss of contextual information. To address this limitation, a graph-based embedding approach is introduced that constructs a word co-occurrence graph to model contextual relationships between terms. This allows word dependencies to be represented more comprehensively before classification.

We use NPMI to make a graph of word co-occurrences so that our embeddings can include a global perspective. This process includes:

After creating a co-occurrence graph, we create word embeddings using GCN. Unlike traditional text representations like TF-IDF, which treat words as distinct features, GCN uses the graph structure to capture both local dependencies (within individual tweets) and global dependencies (across multiple tweets). This enables the model to enhance word connections based on graph connections, resulting in more context-aware embeddings that are useful for classifying tweets related to disasters. Fig. 2 illustrates the two-step GCN-based graph embedding process.

Step 1: Aggregation of the neighborhood (message passing): By combining information from its linked neighbours, each word (or node) in the network modifies its embedding. Node A in the figure receives data from its neighbours B, C, D, E, and F as part of the message-passing process. This step makes sure that each word’s representation acquires more context awareness by incorporating semantic information from other words in the graph. This aggregation step is represented mathematically as: (5)H(l+1)=σ(D~−1/2A~D~−1/2H(l)W(l))

The graph’s adjacency matrix is represented by A~, which includes self-loops to preserve some of the original semantic meaning of each word. The degree matrix, or D~, is used for normalization in order to reduce bias from highly connected words. The feature matrix at layer l1, which includes the existing word embeddings, is represented by H(l). A learnable weight matrix that is optimized during training is represented by W(l). W(l) represents a learnable weight matrix that is optimized throughout the training process. σ is an activation function, like ReLU, that makes the model non-linear. This step makes sure that words with strong co-occurrence relationships have an impact on each other’s embeddings.

Fig. 3 shows the hybridisation of TF-IDF and GCN. First, a method called TF-IDF converts each word in the dataset to a numerical vector. This makes words unique to a tweet more important while making frequently used words less important. In parallel, a graph representation is created, with words (ω1,ω2,ω3,…,ωn) acting as nodes and their co-occurrence associations acting as edges. NPMI, which aids in identifying meaningful word correlations, is used to find these relationships. This word graph is analysed by the GCN, which gathers structural connections between words as well as contextual meaning. This aids the model in comprehending the relationships between words in different tweets. The final feature vector incorporates both global and local word associations by integrating the GCN-generated vectors with the TF-IDF feature vectors. The computed embeddings of TF-IDF and GCN are fed into the classification model once they have been concatenated into a single feature vector:Z=ZTF-IDF⊕ZGCN

By using ⊕ to represent the concatenation operation, the final representation retains both local word importance (TF-IDF embeddings) and global contextual information (GCN embeddings). In order to extract high-level features, the concatenated vector Z is then fed into a FCNN, which has numerous hidden layers and ReLU activation functions. The output probabilities are guaranteed to add up to 1 by a softmax function: Y=Softmax(WFCNN⋅Z+b)

Here b represents the FCNN’s bias term, and WFCNN represents the weight matrix.

The model combines TF-IDF and GCN embeddings to find a balance between statistical word importance and structural contextual understanding. This makes it a strong and reliable way to classify tweets about disasters.

In the concluding phase of the suggested model, a FCNN is used to execute three sequential classification tasks. This multi-task framework enables the model to incrementally enhance predictions, beginning with general categorization and advancing to more precise insights. The tasks are structured in a hierarchical cascade, whereby the outcome of one task dictates the initiation of the subsequent activity. The input for all three tasks is the same: the concatenated vector of TF-IDF and GCN embeddings that the FCNN processes. There is a different output layer for each task, and the softmax activation for each one is based on the number of classes in that task. By letting the model learn generic patterns once and utilize them for many goals, this shared representation increases efficiency by eliminating duplication and enhancing generalization. This step-by-step design is like how people really analyse tweets: first figure out whether they are relevant, then what kind they are, and lastly what kind of help they need. Algorithm 1 presents all the steps of our proposed model.

We examined our suggested method with a number of well-known baseline models in order to assess its efficacy. These models’ specifics are as follows:

For identifying sequential patterns in text, article [16] uses a Bi-LSTM that processes tweets both forward and backward.

Table 7 provides a summary of the findings from these comparisons. With an accuracy of 94.47%, our hybrid TF-IDF+GCN model outperforms all baselines. The strength of combining statistical and structural information in disaster tweet classification is demonstrated by this significant improvement over individual models like TF-IDF (75.60%) and GCN (80.98%), as well as over other deep learning techniques.

The accuracy patterns of the GCN and Hybrid models over 50 epochs are depicted in Fig. 4. It shows that the GCN stabilizes at about 80%, and a hybrid model shows improved learning ability, reaching up to 95% accuracy. Confusion matrices showing the classification performance of the TF-IDF, GCN, and Hybrid models are shown in Fig. 5. While TF-IDF and GCN show slightly higher values of false positives as well as false negatives, the proposed hybrid model has the lowest misclassification rate.

Based on the obtained results, the TF-IDF + GCN hybrid model emerges as the most effective approach for disaster tweet classification in this study, outperforming baseline TF-IDF, standalone GCN, and LSTM-based models. After establishing baseline performance for binary disaster vs. non-disaster classification (Task 1), the evaluation progressed to a more challenging multi-class scenario. In Task 2, the model was required to distinguish between different disaster types, such as floods and earthquakes, which introduced greater complexity compared to binary classification. This task demanded finer-grained discrimination, as the model needed to capture subtle linguistic differences between disaster categories rather than simply determining the presence or absence of a disaster.

The training, validation loss, and accuracy curves over fifty epochs are displayed in Fig. 6. It gives the accuracy vs. loss while the model gets trained. Epochs are displayed on the x-axis, accuracy (%) is shown on the right y-axis, and loss values are shown on the left y-axis. The training and validation loss decreases with increasing epochs, indicating improved learning and generalization of the model. Accuracy increases concurrently, indicating an improvement in classification ability. With a distinct emphasis on striking a balance between lowering loss and increasing accuracy, the image illustrates how the model’s learning has evolved over time.

The confusion matrix heatmaps in Fig. 7 illustrate how well the model detects three disasters. The model generates a large number of accurate predictions, as indicated by the high diagonal values (909, 893, 1031). The framework’s accuracy ranges from 93.39% to 95.82%, its precision ranges from 88.80% to 95.92%, and its F1-scores for each class range from 91.00% to 93.37%, according to the metrics table.

Table 8 shows how three different models can be used to predict the type of disaster. The TF-IDF-only model is the worst, but the GCN-only model is a little better because it captures word relationships. Our hybrid model of TF-IDF and GCN gets the best results: 95.49% for No Disaster, 86.70% for Flood, and 93.31% for earthquake, with an overall accuracy of 91.78%. This shows that using both TF-IDF features and GCN embeddings together gives better results for classifying disasters.

Task 3 was about figuring out what kind of disaster it was, and Task 3 makes the analysis even better by sorting tweets into groups based on what kind of help and support they need. This fine-grained classification is harder because tweets often have signals of help that are unclear or overlap.

Fig. 8 displays the proposed model’s accuracy and loss values with 50 epochs. The test loss drops from 0.1158 to 0.032168, the testing accuracy rises from 91% to 95% in epochs 1 to 50. The figure shows a steady decrease in loss and a steady rise in accuracy, indicating successful model learning and convergence.

Tables 9 and 10 show how well our suggested hybrid TF-IDF+GCN model works for help categorization tasks. Table 9 shows that the hybrid model consistently gets high precision, recall, and F1-scores across all five help categories. The best results are in Other Useful Information (F1 = 0.9836). The overall accuracy is 94.64%, which shows that it is extremely trustworthy. On the other hand, Table 10 compares several baselines, such as TF-IDF-only, GCN-only, LSTM, Bi-LSTM, and BERT. All of these do much worse across categories, especially when it comes to tweets about infrastructure.

Fig. 9 demonstrates that the confusion matrix for Task 3 indicates the model’s efficacy in categorizing tweets into five assistance-related classifications. Despite the fact that there were only a few small misclassifications between labels that were closely connected to one another, such as Sympathy and Other Useful Information, its great performance is highlighted by the concentration of predictions along the diagonal.

The outcomes from all three tasks indicate that the suggested hybrid model functions effectively. It can be used to find disasters in general and to sort them into different types and help categories. This shows that our model can handle both simple and complicated cases, which is important for real disaster response. To gain a clearer understanding of these results, we will examine the contribution of each model component in Section 3.4. This ablation study looks at the TF-IDF only, the GCN only, and the TF-IDF+GCN model together.

To illustrate the contributions of TF-IDF, GCN, and their combination, we present the overall accuracy for each of the three tasks in this ablation study. To determine how much each component contributed, we tested three different iterations of the model: TF-IDF-only, GCN-only, and the hybrid TF-IDF+GCN. The results are displayed in Table 11. For Task 1 (Disaster vs. Non-Disaster), the hybrid model achieved 94.47% accuracy, significantly higher than TF-IDF-only (75.60%) and GCN-only (80.98%). This demonstrates that binary detection is improved by combining structural and statistical features. For Task 2 (Disaster Type Classification), the hybrid model received 91.78%, compared to 76.09 percent for the TF-IDF-only model and 82.30% for the GCN-only model. This shows that combining features helps with multi-class classification. Our hybrid model got 94.64% accuracy for Task 3 (help categorization), which is much better than TF-IDF-only (54.01%) and GCN-only (62.89%). The ablation analysis shows that TF-IDF finds out how important words are, while GCN finds out how words are related to each other. Together, they always make things better on all tasks.

Table 12 shows the full error analysis of the suggested hybrid model for all three classification tasks. It shows that the model works well and consistently, even though the tasks are quite challenging. In binary classification (Task 1, 6674 samples), the model achieves 94.47% accuracy with 372 total errors (5.57%), which is acceptable in disaster detection when missing actual catastrophes (FN) should be reduced. False positives (202, 3.03%) outnumber false negatives (170, Effective, bias-free learning is shown by the balanced error distribution. The model achieves 91.78% accuracy in classifying 3-class disasters (Task 2, 3087 samples), with the top issue being FloodrightarrowEarthquake misunderstanding (10.39%), accounting for the biggest error in the analysis. Large inter-disaster misunderstanding shows that flood and earthquake tweets share grammatical and contextual characteristics, while confusion with non-disaster content remains modest (43–74 instances). The misunderstanding is asymmetric, with FloodrightarrowEarthquake (107) being six times greater than EarthquakerightarrowFlood (18), suggesting that flood-related jargon is more commonly misinterpreted as earthquake terms. Despite five semantically overlapping categories and extreme class imbalance (517 to 75 data), the model achieves the maximum accuracy (94.64%) in help category categorization (Task 3, 1156 samples). Most categories had uniform error distributions (0.97%–1.51% every confusion pair), with only the smallest class (Infrastructure, 75 samples) having a 14.67% error rate with 11 absolute mistakes. The model’s good discrimination is shown by the continuously low inter-category confusion (2–6 instances per pair) across semantically comparable categories like “Food & Medicine” and “Infrastructure & Utilities”. Real semantic ambiguity and linguistic overlap cause errors, not model defects or class imbalance, and the model performs well across all class sizes when accounting for task difficulty.