We formalize robust cross-domain aspect-based sentiment analysis as a multi-task learning problem under domain shift and linguistic noise. Consider a corpus of N sentences X = { x 1 , x 2 , ⋯ , x n } , where each sentence x = [ w 1 , w 2 , ⋯ , w n ] is a token sequence. Sentences originate from domains d ∈ D , with D = { D 1 , D 2 , ⋯ , D m } representing distinct product or service categories (electronics, apparel, restaurants, laptops).

Aspect Term Extraction (AE): For each sentence x , we identify contiguous spans [ s , e ] ( 1 ≤ s ≤ e ≤ n ) corresponding to aspect terms. Following sequence labeling conventions, each token w i receives a label y i ae ∈ Y AE = { B-ASP , I-ASP , O } , where B-ASP marks aspect boundaries, I-ASP indicates continuation, and O denotes non-aspect tokens. The output is a set of aspect spans A = { a 1 , a 2 , ⋯ , a k } with a j = ( s j , e j ) .

Aspect Sentiment Classification (ASC): For each extracted span a j , we predict sentiment polarity y i asc ∈ Y ASC = { POS , NEG , NEU } expressing opinion toward that aspect. Formally, domain shift denotes divergence between source distribution P s ( x , y ) and target distribution P t ( x , y ) , with P s ≠ P t . In cross-domain settings, we assume labeled source data D s = { ( x i , y i ae , y i asc ) } and unlabeled or sparsely labeled target data. Our objective is domain-invariant representations h = f ( x ) such that P s ( h | y ) ≈ P t ( h | y ) while maintaining discriminative power for both tasks.

Based on empirical analysis of Amazon and Yelp, we identify four types relevant to ABSA robustness: orthographic errors. Given an input sentence x = { w 1 , w 2 , ⋯ , w n } , we generate noisy variants x ˜ through four perturbation operations with probabilities p spell , p drop , p syn , and p slang respectively:

Spelling Noise: Character-level perturbations including insertions, deletions, substitutions, and transpositions. The error probability follows:

where e ∈ E ( w ) is a possible error for word w , and λ controls error likelihood.

Word Dropout: Random masking of non-aspect words:

Synonym Replacement: Context-aware substitution using WordNet synonyms: We use GloVe 840B-300d embeddings to compute cosine similarity between the original word and each synonym in WordNet. The synonym with the highest similarity is selected.

where S ( x i ) is the synonym set for word x i .

Slang Injection: Replacement with informal equivalents from dictionary D slang : We compiled a slang dictionary of 2500 entries from Urban Dictionary [31], mapping informal terms to their formal equivalents (e.g., “g2g” → “got to go”).

To guarantee that ground-truth labels remain valid during the NAT process, all perturbation operations (including spelling errors, word dropout, synonym replacement, and slang injection) are applied exclusively to non-aspect tokens. Aspect terms and their exact character spans remain untouched to prevent label corruption. Following perturbation, the sentence is re-tokenized using the BERT/RoBERTa tokenizer, and the original BIO tags are automatically remapped to the new subword sequence using standard offset-based alignment. This ensures that AE and ASC labels stay perfectly aligned with the noisy input.

The training batch composition is controlled by noise ratio β : β set to 0.5.

Curriculum Schedule: The noise probability follows a curriculum schedule:

where t is the current training step, T ramp = 5000 is the ramp-up period, and N max = 0.3 is the maximum noise probability. Each individual perturbation probability ( P spell , P drop , P syn , P slang ) follows this same schedule.

We design a unified multi-task transformer (Figure 1) with three components: a shared encoder producing contextualized token representations; an aspect extraction head for sequence labeling; and an aspect sentiment classification head with span-level attention. Joint learning avoids pipeline error propagation and enables task reinforcement.

Shared Transformer Encoder: Both clean and augmented noisy sentences x = [ w 1 , w 2 , ⋯ , w n ] are encoded using a pre-trained transformer (BERT or RoBERTa):

Figure 1. Model architecture of the proposed noise-aware multi-task transformer framework for cross-domain ABSA. The model processes input sentences through a shared transformer encoder, then uses two task-specific heads for joint aspect extraction (BIO tagging) and sentiment classification. The noise generator creates realistic perturbations during training to enhance robustness.

where d = 768 for base models. This shared representation serves as a common feature space for all downstream tasks.

Aspect Extraction Head:where Aspect extraction is formulated as a sequence labeling task using BIO tags. For each token representation h i :

where W AE ∈ ℝ 3 × d ,and y i ∈ { B-ASP , I-ASP , O } .

The extraction loss is token-level cross-entropy:

where y i * is the ground truth tag.

Aspect Sentiment Classification Head:For each aspect span, we compute an aspect-aware representation via attention pooling. During training the ASC head receives gold aspect spans (teacher forcing); at inference and for all reported results the AE head’s predicted spans are used, yielding a fully end-to-end joint evaluation that matches the protocol. First, compute attention weights:

where q ∈ ℝ d is a learnable initialized randomly and updated during training. Then aggregate representations:

for sentiment classification:

where s ∈ { Positive , Negative , Neutral } . The ASC loss is:

where s * is the ground truth sentiment.

Multi-Task Objective:The combined loss function is:

To promote noise-invariant representations we further introduce a sentence-level invariance loss L inv_sent . For every clean-noisy pair ( x , x ˜ ) in a training batch we compute the KL-divergence between the model’s output distributions for both tasks (token probabilities for AE are averaged; ASC uses the pooled representation.

To explicitly enforce noise invariance, we add a span-level L inv_span . consistency loss between clean and noisy inputs. For each gold aspect span s, we compute pooled representations v s and v ˜ s from the clean and noisy sentences, respectively, and define:

The final training objective is:

where λ noise is a hyperparameter that controls the strength of the noise-invariance constraint. λ AE and λ ASC are task weights optimized via grid search. After searching over {0.2, 0.5, 1.0, 2.0} for each weight, we selected λ AE = 1.0, λ ASC = 1.0 on Rest14 validation. In all experiments, noise invariance is applied at the span level with λ noise = 0.1 .

We employ a mixed training regimen where each batch contains both clean and noisy samples following the curriculum schedule in Equation 6. All models are fine-tuned from pre-trained bert-base-uncased and roberta-base checkpoints.

We evaluate three configurations:

1)In-Domain: D train = D test (same domain);

2)Cross-Domain: D train = D S , D test = D T (different domains);

3)Multi-Domain: D train = ∪ i = 1 k D i , D test = D j (trained on all, tested on each).

To test our model’s performance, we used benchmark datasets: (Res14 and LAP14) from SemEval-2014 Task 4 [9], and real-world dataset: (electronics and Apparel) from Amazon, and Food from Yelp. All carry three sentiment polarities: positive, neutral, and negative, Table 3 presents detailed statistics for each dataset, including sentence counts, aspect counts, average length, and sentiment distribution. LAP14 has the highest share of long sentences (over 35 tokens) at 18%, which could hurt results compared to Res 14. Still, it better shows how the model handles tough, extended text. Meanwhile, Res14 contain mostly short sentences (under 20 tokens), making up 56.68%.

Implementation details

Models trained for up to 1500 steps. After 1000 steps, we selected the best checkpoint every 100 steps using micro-F1 on the dev set. We implement using PyTorch 2.0 and Hugging Face Transformers. Hyperparameters: learning rate 2 × 10⁻⁵, batch size 16, AdamW optimizer with ( β 1 , β 2 ) = ( 0.9 , 0.999 ) , weight decay 0.01. Training runs for 20 epochs with early stopping (patience = 5), λ_noise = 0.1 (Noise invariance weight) and Dropout = 0.1 (Encoder + heads). All experiments use NVIDIA A100 GPUs. Experiments ran in PyCharm with Python 3.6.

For preprocessing task, all sentences are tokenized using the tokenizer corresponding to each pre-trained encoder (BERT-base, RoBERTa-base). Keeping cased input for BERT; RoBERTa uses its default tokenization. Sentences longer than 128 tokens are truncated; shorter ones are padded. Aspect annotations are converted to BIO tags programmatically.

Baselines

We compare our approach to various baseline models listed in Table 2(a). Most baselines target only particular ABSA subtasks, with some offering complete coverage across all components.

Single-Task: MTL-AraBERT [15], Robust ABSA w/Non-counterfactual Aug. [30], SentiSys [1], EDA/UDA [23]Multi-Task: LEGO-ABSA [27] (fine-tuned on source domain), CC-ASTN [6], BERT-Based Cross-Domain [3], SentiSys [1].Cross-Domain: BGCA [20] (fine-tuned on source domain for fair comparison) UNIFIEDABSA [28] SyMux [29].

Noise Simulation

We simulate textual noise at two intensity levels (low and high) by introducing spelling errors, word deletions, synonym substitutions, and slang variations with different probabilities. The low-noise setting produces minor and mostly readable distortions, while the high-noise setting introduces stronger corruption that significantly alters word forms and sentence structure. following established noise simulation protocols [22][23] two intensity levels were created as:

Low Noise: p spell = 0.05 , p drop = 0.05 , p syn = 0.1 , p slang = 0.05 ;High Noise: p spell = 0.15 , p drop = 0.2 , p syn = 0.3 , p slang = 0.15 .

Table 3.Dataset statistics.

Table 4 illustrates examples of original sentences and their corresponding noisy variants at both intensity levels.

Table 4.Example sentence pairs (original → noisy variant).

Evaluation Metrics

To evaluate our experiments, we conducted Precision (P), Recall (R), F1-score for Aspect Extraction: Sentiment Classification: Accuracy, Macro-F1, for Efficiency: Training time (hours), Inference speed (sentences/second), Parameters (millions), and for Robustness:

4.2.1. Overall Performance Comparison (RQ1)

The overall performance results from benchmark SemEval datasets are displayed in Table 5. Our proposed framework achieves highly competitive results across both domains and tasks. The RoBERTa variant shows the best performance with 87.9 F1 for AE and 85.6% accuracy for ASC on Rest14, representing improvements of +2.6 F1 and +2.9% over the best baseline (BGCA). Our unified multi-task approach proves effective through its continuous better performance in both Restaurants and Laptops domains. The performance gap between single-task models and our unified framework demonstrates the advantages of dual learning and common representation use.

Table 5. Overall performance on SemEval benchmark datasets (F1-score for aspect extraction, accuracy for aspect sentiment classification). All results are mean ± std over 5 runs.

*Fine-tuned on source domain for fair comparison.

4.2.2. Noise Robustness Analysis (RQ2)

Table 6 shows how well the system handles linguistic noise in the Amazon Electronics dataset. The noise-aware training (NAT) strategy of our study allows us to maintain performance in the presence of noisy environments. The model with NAT demonstrates only −9.4%/−11.7% performance loss during high noise situations which stands in contrast to the single-task BERT system that exhibits −22.4%/−26.1% performance decline, resulting in a 42% improvement of our model’s resistance to performance loss. The results demonstrate that: 1) All models degrade with increasing noise, but multi-task architectures show inherent robustness; 2) Our framework without NAT already outperforms baselines; 3) With NAT, degradation is minimized, proving the effectiveness of systematic noise injection during training. Concretely, NAT reduces the high-noise AE degradation from 16.1% (w/o NAT) to 9.4% (with NAT), a relative improvement of 42%. The corresponding ASC reduction is 39%.

4.2.3. Cross-Domain Generalization (RQ3)

Table 7 shows cross-domain generalization from Electronics to Apparel. The study presents performance results for Electronics as well as for cross-domain measurements which were conducted on Apparel. The system achieved its highest cross-domain results through 76.9 F1 score and 79.7% accuracy results which showed a minor performance decrease of 11 and 5.9 points when compared to BERT’s complete drop of 12.7 and 8.3 points. The research demonstrates advanced domain-independent knowledge representation systems. The study demonstrates that noise-aware training increases system robustness while also boosting abilities to transfer knowledge across different domains because it teaches models to disregard unimportant language differences which helps them identify fundamental meaning connections that remain consistent across different domains.

Table 6.Performance degradation under different noise levels on Amazon Electronics dataset. Values show relative drop in F1 (AE)/Accuracy (ASC). Lower values indicate better robustness.

Table 7. Cross-domain performance (Electronics → Apparel). In-domain performance on electronics shown for reference.

Table 8 shows how multi-domain training provides advantages. The mixed data training across all three domains led to improved results which reached F1 score increases between 0.7 and 1.6 for AE and 1.6 to 2.4 percent for ASC. The research indicates that learners who study various fields will acquire better skills to develop general knowledge systems which will decrease their tendency to become attached to particular field knowledge. The multi-task learning framework delivered better results than the single-task method across all research domains we tested. The Ours (Multi-Task) model demonstrated superior F1 performance in Aspect Extraction and achieved higher accuracy results in Sentiment Classification when compared to the single-task model. The model achieved its most effective results in the Apparel and Electronics domains because it utilized common features to enhance its performance. The research results demonstrate statistical significance which proves that simultaneous training on multiple domains results in improved model performance for accurate sentiment detection.

4.2.4. Efficiency Analysis (RQ4)

Table 9 presents the efficiency-accuracy trade-off. The framework we created requires 2M extra parameters which represent a 1.8% increase over MTL-BERT. Compared to MTL-BERT, the proposed model achieves F1 improvements of 3.3 points and accuracy improvements of 4.1 percent. The training time increase of 22 percent proves acceptable because it leads to major performance enhancements. The system achieves a practical inference speed of 340 sentences per second which supports real-time usage. Our method outperforms LEGO-ABSA especially in terms of cost-effectiveness and precision. The efficiency comparison shows that our proposed model achieves the highest performance in both Aspect Extraction (F1: 87.9) and Aspect Sentiment Classification (Accuracy: 85.6). The system requires more resources than the base BERT model yet it operates efficiently with acceptable speed during training and execution. The ablation study further highlights the importance of each component in our full model. The system performance declines together with cross-domain capability after we eliminate multi-task learning and noise-aware training and attention pooling. The system operates with decreased resistance against noise in audio signals because these elements do not exist especially when multi-task and NAT elements get deleted. The proposed architecture requires all components to achieve both high accuracy and system durability.

Table 8. Multi-domain training results. Training on mixed data from all domains improves performance even on individual domains. All results are statistically significant (p < 0.05).

Table 9. Efficiency comparison on amazon electronics dataset (batch size = 16).

Our model delivers consistent gains: +2.6 F1 (AE)/+2.9% ACC (ASC) over the strongest baseline on SemEval Rest14 for BGCA (Table 5) and up to +3.1 F1 (AE) and +3.3% (ASC) over the strongest cross-domain baseline (BGCA) in Table 7. Across datasets and settings, the improvements consistently range between 2.6 - 3.1 F1 and 2.7% - 3.3% accuracy. The shared representation enables mutual reinforcement between tasks while reducing error propagation. The research showed that noise-aware training methods provide better training results because NAT reduces the high-noise AE degradation from 16.1% (Ours w/o NAT) to 9.4% (Ours with NAT), a relative improvement of 42% (and 39% for ASC; Table 6). Targeted augmentation techniques enable systematic learning of real-world linguistic variation robustness according to the research findings. The combination of multi-task learning and noise-aware mechanisms will produce better prediction stability according to results from RQ3. Our framework demonstrates effective cross-domain generalization because it only experiences minor performance reduction which indicates that noise-invariant representation learning enables models to detect fundamental patterns that exist across different domains. The proposed framework requires 2M parameters and 22% training time overhead according to RQ4 which demonstrates that systems can achieve robustness through low computational needs while maintaining efficient inference speed.

The results of our approach lead to several theoretical implications. First, noise augmentation functions as a strong regularizer. It helps the model learn noise-invariant representations that transfer more effectively across domains. This is consistent with prior work in NLP showing that increasing data, rather than model size, often improves robustness. Second, multi-task learning improves robustness in ways that go beyond simple accuracy gains, because it pushes the model to learn representations that support multiple related objectives at the same time. Third, the observed link between noise robustness and domain generalization indicates that both are supported by learning deeper and more abstract representations that are less sensitive to surface-level variations in the input.

Table 10 shows the ablation study by demonstrating the contribution of each component. Removing multi-task learning causes the largest performance drop (−2.7 F1, −2.9% accuracy). Disabling NAT significantly reduces both noise robustness (degradation increases from −9.4% to −16.1%) and cross-domain performance. Removing attention pooling causes moderate drops. All components contribute meaningfully, with multi-task learning being most critical for accuracy and NAT for robustness.

We manually analyze 100 error cases from the Restaurants test set and compare error type distribution between BERT and our model. Table 11 shows the comparative of error analysis. This error analysis compares the performance of our proposed model against the baseline BERT model on the Restaurants test set, categorizing the types of errors encountered. The results demonstrate that our model achieves a consistent reduction in errors across all identified categories. The most significant improvement is observed in handling severe noise cases, where the error rate is reduced by 40% compared to BERT. There is also a substantial 34% reduction in aspect boundary errors, indicating the model’s superior ability to precisely delimit aspect terms. Furthermore, the model shows a marked improvement in understanding domain-specific language, with a 33% reduction in errors. While improvements in handling complex negation and implicit aspects are less pronounced, with reductions of 28% and 22% respectively, the overall trend indicates the proposed model’s robustness in addressing various challenging linguistic phenomena in sentiment analysis.

Table 10. Ablation study on Amazon Electronics dataset, each row removes one component from the full model. Noise degradation is measured on the high-noise test set. Cross-domain performance is on Apparel.

Table 11.Error type distribution comparison between BERT and our model on Restaurants test set.

The residual errors provide clear insights into remaining challenges and the complementary roles of our components. Complex negation scope (e.g., -not bad but overpriced) still accounts for 18% of errors; NAT reduces this by exposing varied phrasings during training, while MTL helps the shared encoder learn consistent aspect-sentiment alignments. Boundary fragmentation in multi-word aspects (e.g., battery life) is largely mitigated by the span-level attention pooling, which focuses representation on the whole extracted span. Domain-specific lexical shifts (e.g., long = positive for batteries) drop 33% thanks to synonym and slang augmentation that forces the model to rely on contextual rather than surface cues. These patterns confirm that MTL supplies the joint reasoning backbone, NAT supplies surface robustness, and attention pooling supplies precise span aggregation.

Despite the encouraging results, several limitations should be acknowledged. The created noise models demonstrate effective performance for English usage yet their application to other languages and cultural settings proves to be ineffective. Future work should explore methods for generating noise which adapt to different languages and cultural backgrounds. The current approach detects local token-based noise but it cannot detect sarcasm and irony and cultural references which affect how people understand sentiments. The training process takes 22 percent more time which works for our situation but becomes unmanageable when dealing with extremely large data sets and model structures. Researchers should investigate better methods for augmenting data which include using adversarial attacks and creating noise through dynamic methods. The experiments do not cover very long texts (over 512 tokens) that exceed typical transformer input limits. Researchers have not yet determined how to assess system performance for handling extensive documents. The field of ABSA now requires multimodal input; therefore, systems must develop ability to resist noise from textual sources and visual content and other sensory input which produces distinct forms of interference. The scientific community has not yet discovered how noise augmentation methods lead to improved ability to generalize across different domains. The representation analysis process enables researchers to understand how different aspects of robustness and generalization develop together.