CoChat
Sign Up
Cited by 0Sentiment Analysis and Opinion MiningDomain Adaptation and Few-Shot LearningTopic Modeling

GPT-based lifelong learning and ANFIS-driven reply memory ratio prediction for aspect-based sentiment analysis

Open access
Huang, Huang, Ye, Xiaohong, Mustafa, Mumtaz Begum, Dong, Qiyuan, Li, Yu, Asemi, Adeleh, Asemi, Asefeh·2025·Complex & Intelligent Systems
DOI 10.1007/s40747-025-02086-2
Open this paper in CoChat — free accountOpen PDF

Abstract

GPT is among the most powerful large language models (LLMs), known for its versatility and strong performance across a wide range of tasks. However, its substantial computational demands and limited cross-domain generalization present challenges, particularly in resource-sensitive applications like Aspect-Based Sentiment Analysis (ABSA). Existing ABSA models are typically domain-specific and suffer from catastrophic forgetting—losing previously acquired knowledge when sequentially trained on new domains—resulting in poor scalability and knowledge retention. To address these issues, we proposed a novel framework that integrates GPT-2 with a replay-based lifelong learning mechanism to support incremental, multi-domain ABSA while mitigating forgetting. The model is sequentially fine-tuned using real-data replay across four diverse ABSA domains: Laptops, Restaurants, Tweets, and Finance. Experimental results show that proposed model achieves an average accuracy of 0.85 and a Backward Transfer (BWT) score of -0.09, significantly outperforming the baseline model without lifelong learning (accuracy of 0.70, BWT of -0.64). We further conducted a t-test to validate the statistical significance of the improvements. we design an ANFIS-based model trained on experimental results to predict the proper memory ratio for new datasets, enabling more effective and adaptive lifelong learning in GPT-based architectures. In addition, we applied a multi-level data augmentation pipeline, which significantly improved performance across domains ( p = 0.0049), enhancing both retention and generalization under constrained memory. Additionally, we introduced a domain sequence permutation study to test robustness against task-order sensitivity, while another part evaluates generalization under a simulated fifth domain constructed from a mixture of all original domains. These components validate the model’s scalability and ability to generalize beyond trained distributions. An ablation study was conducted to isolate the contributions of each module, including replay, ANFIS prediction, and data augmentation. The results demonstrate that removing any single component leads to a measurable drop in performance, confirming that each part of the framework contributes meaningfully to overall effectiveness. A post-hoc comparison with a distillation-based lifelong learning baseline was conducted, showing improved performance of our replay + ANFIS approach. We acknowledge that additional comparisons with advanced baselines such as GEM or A-GEM are needed to further strengthen this claim, which we leave for future work. Overall, this study demonstrates the effectiveness of combining GPT-2 with replay-based lifelong learning and adaptive memory control, offering a scalable and robust solution for continuous, multi-domain sentiment analysis. Our proposed model provides a practical foundation for future research in dynamic, cross-domain ABSA systems.

AI snapshot

Written from the paper’s full text.

TL;DR

The paper proposes a GPT-2 framework with replay-based lifelong learning and ANFIS-predicted replay memory ratios for multi-domain aspect-based sentiment analysis (ABSA). It reports improved average accuracy and reduced forgetting versus a no-lifelong-learning baseline, with additional gains from data augmentation.

Key findings

  • The proposed model achieved an average accuracy of 0.85 and a Backward Transfer (BWT) score of -0.09, compared with 0.70 and -0.64 for the baseline without lifelong learning.
  • A combined one-tailed paired t-test showed a mean accuracy improvement of +0.104, t = 3.77, p = 0.0009; domain-specific significance was reported only for Laptop (p = 0.033).
  • ANFIS predicted memory ratios close to observed best values: Laptop 0.85 vs. 0.9, Tweet 0.85 vs. 0.9, and Finance 0.72 vs. 0.7; it generated 108 fuzzy if-then rules.
  • Data augmentation improved mean accuracy from 0.5948 to 0.6114 (t = 2.31, p = 0.036 in the results section; the abstract/conclusion report p = 0.0049), with examples such as Laptop 0.6105 → 0.6909 and Restaurant 0.6834 → 0.7208.

Methods

  • GPT-2 was sequentially fine-tuned on four ABSA domains: Laptop, Restaurant, Tweets, and Finance, with dataset sizes of 1,449, 1,708, 6,280, and 10,753, respectively; data were split 70% training, 20% testing, and 10% validation and formatted as SQuAD-style JSON.
  • Replay-based lifelong learning used memory ratios of 0.1, 0.3, 0.5, 0.7, and 0.9; ANFIS was trained on 15 records from dynamic memory experiments, using F1-score, precision, recall, and METEOR as inputs and memory ratio as output.
  • Evaluation used ROUGE-1 (F1, recall, precision), METEOR, accuracy, and BWT, with one-tailed paired t-tests, a few-shot 30% data setup, and a simulated fifth domain constructed by mixing the four original domains.

Limitations

  • Stated: the paper acknowledges that additional comparisons with advanced baselines such as GEM or A-GEM are needed, and that future work should explore contrastive learning; the computational analysis assumes conventional quadratic-cost attention and does not implement sparse or linear attention variants.
  • Evident from the text: cross-domain transfer remains weak in some directions (e.g., Laptop → Tweets F1 = 0.26; Tweets → Finance F1 = 0.36), and augmentation sometimes reduced performance (e.g., Tweet → Finance 0.4108 → 0.3559).
  • Evident from the text: ANFIS was trained on only 15 records, and the fifth-domain generalization test was simulated from the original domains rather than using an external dataset.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1007/s40747-025-02086-2

Cites

48 of 48 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

11 of 17 cited sources checked · 4 supported · 1 flagged · accuracy not yet measured

  • Premisenone given
    § SemEval-2014 Task 4

    EWC and CORE still struggle to effectively balance retaining past knowledge with integrating new data.

    “Although methods like Elastic Weight Consolidation (EWC) [16] and COgnitive REplay (CORE) [32] have been proposed to mitigate this issue, they still struggle to effectively balance retaining past knowledge with integrating new data.”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [16] Kirkpatrick J, Pascanu R, Rabinowitz N, Veness J, Desjardins G, Rusu AA, Hadsell R (2017) Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences 114(13):3521–3526. source

      The paper introduces EWC, but its Results (section 2.1, p. 3) report that EWC learns a new task well without forgetting the old one—the opposite of the attributed difficulty balancing the two. The Discussion (p. 6) likewise says EWC protects previous knowledge during new learning. This paper does not evaluate CORE. (LLM, conf=0.88)

      Checked by a reasoning model — may be wrong.

      Experimental — accuracy not yet measured
    • cites [32] Zhang J, Fu Y, Peng Z, Yao D, He K (2024) Core: mitigating catastrophic forgetting in continual learning through cognitive replay. ArXiv preprintarXiv:240201348

      Not checked in the free run.

  • Premisecitation
    § SemEval-2014 Task 4

    These models face issues with catastrophic forgetting when incorporating data from new domains.

    “Moreover, these models face issues with catastrophic forgetting when incorporating data from new domains [12].”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [12] Jiang B (2024) All in One: An empirical study of GPT for few-shot aspect-based sentiment anlaysis. arXiv preprint arXiv:2404.06063 source

      Section 5.3 describes transfer tests across datasets, but does not report whether adding data from a new domain causes previously learned knowledge to be forgotten. Searches for “catastrophic forgetting” and forgetting after training on new domains found no passage addressing that issue in our copy. (LLM, conf=0.70)

      Experimental — accuracy not yet measured
  • Premisecitation
    § SemEval-2014 Task 4

    Current deep-learning ABSA models face significant challenges in generalizing across different domains.

    “Current deep-learning ABSA models face significant challenges in generalizing across different domains [30].”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [30] Kirkpatrick J, Pascanu R, Rabinowitz N, Veness J, Desjardins G, Rusu AA, Milan K, Quan J et al (2017) Tiago Ramalho, Agnieszka Grabska-Barwinska,. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521–3526 source

      The Introduction (p. 1) discusses catastrophic forgetting when neural networks learn tasks sequentially, and the supervised-learning experiments (p. 3) use MNIST digit classification. Our copy does not address whether aspect-based sentiment analysis (ABSA) models generalize across domains; searches for ABSA and cross-domain sentiment analysis found no relevant passage. (LLM, conf=0.77)

      Experimental — accuracy not yet measured
  • Premisecitation
    § Research method

    ANFIS is suitable for tasks that demand a balance between prediction performance and model explainability.

    “These studies underscore ANFIS's suitability for tasks that demand a balance between prediction performance and model explainability.”
    Quote in the paperSubstantial claim3 of 3 sources checked
    • cites [38] Asemi A, Asemi A, Ko A (2024) A model for investment type recommender system based on the potential investors based on investors and experts feedback using ANFIS and MNN. J Big Data 11(1):128 source

      The ANFIS training and testing section reports reasonably good predictive performance (F1-score 0.6667) and describes 18 fuzzy rules used for investment recommendations. The Discussion likewise says the system performs well in that setting. Neither section establishes that ANFIS balances prediction performance with model explainability or is suitable for tasks requiring both; the paper’s finding is also specific to investment recommendations. No page numbers are available in our copy. (LLM, conf=0.84)

      Experimental — accuracy not yet measured
    • cites [40] Kumari, S., Agarwal, S., Kumar, M., Sharma, P., Kumar, A., Hashem, A., … Garg, M.C. (2025). An exploration of RSM, ANN, and ANFIS models for methylene blue dye adsorption using Oryza sativa straw biomass: a comparative approach. Scientific Reports 15(1):2979. source

      The Results and discussion sections report strong ANFIS prediction accuracy for methylene blue adsorption and describe its fuzzy-system reasoning and IF-THEN rules (pages not provided). They do not assess model explainability or establish that ANFIS balances it against predictive performance, so that broader suitability claim is not fully supported. (LLM, conf=0.83)

      Experimental — accuracy not yet measured
    • cites [42] Shaier AA, Flah A, Kraiem H, Enany MA, Elymany MM (2025) Novel technique for precise derating torque of induction motors using ANFIS. Sci Rep 15(1):8550 source
  • Findingown data
    § Abstract

    The proposed model achieves a Backward Transfer (BWT) score of -0.09.

    “Experimental results show that proposed model achieves an average accuracy of 0.85 and a Backward Transfer (BWT) score of -0.09, significantly outperforming the baseline model without lifelong learning (accuracy of 0.70, BWT of -0.64).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The four selected ABSA domains (Laptops, Restaurants, Tweets, Finance) are representative of diverse multi-domain ABSA and support generalization to continuous, multi-domain sentiment analysis.
  • May combine two claims. 57
  • Classified as probably not a substantial claim. 44
  • The quote only partly states this claim. 4

  • May combine two claims.

    Premisecitation
    § SemEval-2014 Task 4

    DBLL excels in efficiently preserving knowledge through selective experience replay without needing complex architectural changes or parameter adjustments.

    “Compared to ABLL and RBLL, Data-Based Lifelong Learning (DBLL) excels in efficiently preserving knowledge through selective experience replay [17,18] without needing complex architectural changes or parameter adjustments.”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [17] Rebuffi SA, Kolesnikov A, Sperl G, Lampert CH (2017) icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition (pp. 2001–2010) source

      The Conclusion (p. 9) describes prioritized exemplar selection and using those exemplars to avoid catastrophic forgetting, supporting the selective-replay point. But Related Work (p. 6) says iCaRL updates model parameters and also uses distillation; the paper does not support the claim that it preserves knowledge without parameter adjustments. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
    • cites [18] Li Z, Hoiem D (2017) Learning without forgetting. IEEE Trans Pattern Anal Mach Intell 40(12):2935–2947 source

  • Classified as probably not a substantial claim.

    Premisecitation
    § Ablation study: disentangling the impact of replay memory ratios

    Balancing memory and learning efficiency is important in LLM-based ABSA continual learning.

    “it emphasized the importance of balancing memory and learning efficiency in LLM-based ABSA continual learning [3].”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [3] Ding X, Zhou J, Dou L, Chen Q, Wu Y, Chen C, He L (2024) Boosting large Language models with continual learning for aspect-based sentiment analysis. ArXiv Preprint arXiv:240505496 source

      Section 3.1 describes parameter-efficient fine-tuning, and Section 4.2 describes a baseline that stores samples as memory. Neither passage says that balancing memory use against learning efficiency is important in LLM-based ABSA continual learning. Searches for the proposed balance, memory, and efficiency found no passage addressing that point. (LLM, conf=0.73)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § Ablation study: disentangling the impact of replay memory ratios

    Memory-preserving mechanisms like Elastic Weight Consolidation are important in continual learning.

    “it demonstrated the importance of memory-preserving mechanisms like Elastic Weight Consolidation in continual learning [54].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [54] Aslam S, Rasool A, Li X, Wu H (2025) Cel: A continual learning model for disease outbreak prediction by leveraging domain adaptation via elastic weight consolidation. Computational Life Sciences, Interdisciplinary Sciences, pp 1–19 source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § Lifelong learning model accuracy

    Lopez-Paz and Ranzato introduce evaluation metrics such as Average Accuracy (ACC) and Backward Transfer (BWT) to assess the effectiveness of continual learning.

    “They also introduce evaluation metrics such as Average Accuracy (ACC) and Backward Transfer (BWT) to assess the effectiveness of continual learning.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [2] Lopez-Paz D, Ranzato MA (2017) Gradient episodic memory for continual learning. Adv Neural Inf Process Syst, 30 source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § Lifelong learning model accuracy

    Lopez-Paz and Ranzato employ episodic memory and loss constraints to prevent the degradation of performance on previous tasks while learning new ones.

    “Lopez-Paz and Ranzato [2] employ episodic memory and loss constraints to prevent the degradation of performance on previous tasks while learning new ones.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [2] Lopez-Paz D, Ranzato MA (2017) Gradient episodic memory for continual learning. Adv Neural Inf Process Syst, 30 source

  • May combine two claims.

    Findingown data
    § Abstract

    The proposed model significantly outperforms the baseline model without lifelong learning on accuracy and BWT.

    “Experimental results show that proposed model achieves an average accuracy of 0.85 and a Backward Transfer (BWT) score of -0.09, significantly outperforming the baseline model without lifelong learning (accuracy of 0.70, BWT of -0.64).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The four selected ABSA domains (Laptops, Restaurants, Tweets, Finance) are representative of diverse multi-domain ABSA and support generalization to continuous, multi-domain sentiment analysis.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Findingnone given
    § Computational complexity

    Performance degrades over time on earlier domains.

    “The self-attention operation in each performance degrades over time on earlier domains, providing evidence of catastrophic forgetting.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § Lifelong learning model accuracy

    Once the model begins learning from the tweets domain, its accuracy rapidly improves, peaking at approximately 0.9, the highest score observed across all domains.

    “However, once the model begins learning from the tweets domain, its accuracy rapidly improves, peaking at approximately 0.9, the highest score observed across all domains.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § Conclusions

    The model achieved an average accuracy of 0.85, surpassing the model without lifelong learning, which had an accuracy of 0.70.

    “Through a series of experiments, our model achieved an average accuracy of 0.85, surpassing the model without lifelong learning, which had an accuracy of 0.70.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § Part 5 Testing Generalization Under Extended Domain Sequences

    The model's accuracy on a domain after being trained on that same domain is 0.85 for D_laptop, 0.86 for D_restaurant, 0.89 for D_tweets, and 0.88 for D_finance.

    “The diagonal from the top left to bottom right shows the model's performance on a domain after being trained on that same domain, with scores of 0.85 (D laptop ), 0.86 (D restaurant ), 0.89 (D tweets ), and 0.88 (D finance ).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § Part 5 Testing Generalization Under Extended Domain Sequences

    When trained on D1 without lifelong learning, the model achieves its highest score of 0.81 on that domain but experiences a significant drop in performance when tested on D2, D3, and D4.

    “When trained on D 1 , the model achieves its highest score of 0.81 on that domain but experiences a significant drop in performance when tested on D 2 , D 3 , and D 4 .”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § Part 4 Few-Shot Domain Adaptation Setup

    In the Tweets-to-Finance direction, the model achieved 0.86 accuracy with 30% data versus 0.85 with full data.

    “Similarly, the Tweets-to-Finance direction yielded 0.86 versus 0.85, demonstrating stable generalization across domains.”
    Quote in the paperSubstantial claim

  • The quote only partly states this claim.

    Findingown data
    § Post-hoc comparative analysis with knowledge distillation-based lifelong learning

    The METEOR score was slightly lower under the distillation-based setup than under the replay-based setup.

    “Additionally, the METEOR score, indicating semantic fidelity and fluency, was slightly lower under the distillation-based setup.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § Statistical evaluation of augmentation impact

    Laptop test accuracy rose from 0.6105 to 0.6909, and Restaurant test accuracy rose from 0.6834 to 0.7208 under the structured augmentation pipeline.

    “For example, Laptop test accuracy rose from 0.6105 to 0.6909, and Restaurant from 0.6834 to 0.7208.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingreasoning
    § Performance comparison of proposed model and original GPT2

    For a new dataset with F1-score 0.52, Precision 0.52, Recall 0.49, and METEOR 0.59, the ANFIS recommends a memory ratio of 0.5.

    “given a new dataset with F1-score of 0.52, Precision of 0.52, Recall of 0.49, and METEOR score of 0.59, the ANFIS recommends a memory ratio of 0.5.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § Post-hoc comparative analysis with knowledge distillation-based lifelong learning

    ANFIS predicted best memory ratios of 0.85 for Laptop, 0.85 for Tweet, and 0.72 for Finance.

    “| Laptop | 0.9 | 0.85 | 0.5997 | | 0.6998 |”
    • Also stated in: § Post-hoc comparative analysis with knowledge distillation-based lifelong learning · plain
      “| Tweet | 0.9 | 0.85 | 0.8459 | | 0.8723 |”
    • Also stated in: § Post-hoc comparative analysis with knowledge distillation-based lifelong learning · plain
      “| Finance | 0.7 | 0.72 | 0.8465 | | 0.8819 |”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § Domain Transition

    E_laptop,laptop outperforms E_restaurant,laptop, E_tweets,laptop, and E_finance,laptop.

    “A comparative analysis of E laptop, laptop with the subsequent evaluations-E restaurant, laptop , E tweets, laptop , and E finance, laptopclearly shows that E laptop, laptop outperforms the others.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § Part 4 Few-Shot Domain Adaptation Setup

    The few-shot experimental results complement our full-data in the laptop domain (p = 0.033), while improvements in restaurant, tweets, and finance domains were not individually significant.

    “The few-shot experimental results complement our full-data in the laptop domain (p = 0.033), while improvements in restaurant, tweets, and finance domains were not individually significant.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § Results of experiment II: dynamic memory based experiment

    Comparing models with and without lifelong learning, METEOR drops from 0.85 to 0.70, while BWT increases significantly from -0.09 to -0.64.

    “When comparing the METEOR and BWT metrics between models with and without lifelong learning methods, we observe a drop in METEOR from 0.85 to 0.70, while BWT increases significantly from -0.09 to -0.64.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § Performance comparison of proposed model and original GPT2

    In the Laptop domain, the proposed GPT2 with lifelong learning had a mean accuracy of 0.8175 versus 0.6250 for original GPT2, with a mean difference of +0.1925 and a one-tailed p-value of 0.033.

    “Laptop 0.8175 0.6250 + 0.1925 0.033”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingnone given
    § Performance comparison

    The Tweets domain exhibits strong intra-domain performance across all settings, remaining robust even under low replay conditions.

    “The Tweets domain exhibits strong intra-domain performance across all settings, remaining robust even under low replay conditions.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingnone given
    § Performance comparison

    The adaptive recommendations led to improvements in model stability and performance consistency.

    “The adaptive recommendations led to improvements in model stability and performance consistency”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data
    § Performance comparison of proposed model and original GPT2

    Following training, the ANFIS model generated 108 fuzzy if-then rules in its rulebase.

    “Following training, the model generated 108 fuzzy ifthen rules in its rulebase.”
    Quote in the paperNot a substantial claim

  • The quote only partly states this claim.

    Findingown data
    § Post-hoc comparative analysis with knowledge distillation-based lifelong learning

    The replay-based model achieved higher average accuracy (0.85) than the distillation-based model (0.79).

    “| Replay-based (proposed) | 0.85 | -0.09 0.85 |”
    • Also stated in: § Post-hoc comparative analysis with knowledge distillation-based lifelong learning · plain
      “| Distillation-based (baseline) 0.79 | -0.22 0.81 |”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § ANFIS prediction accuracy

    The ANFIS model successfully learned the nonlinear relationship between performance indicators and memory requirements, yielding high accuracy in predicting the most suitable memory ratios for each domain configuration.

    “The model successfully learned the nonlinear relationship between performance indicators and memory requirements, yielding high accuracy in predicting the most suitable memory ratios for each domain configuration.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The performance indicators (F1-score, Precision, Recall, METEOR) are sufficient to determine the optimal memory ratio for a dataset.

  • May combine two claims.

    Findingown data
    § Results of experiment II: dynamic memory based experiment

    With D3 and D4 training, the model performs best on the domain it was trained on, peaking at 0.89 on D3, but struggles to retain this performance across other domains.

    “A similar trend is observed with D 3 and D 4 training: the model performs best on the domain it was trained on, peaking at 0.89 on D 3 , but struggles to retain this performance across other domains.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § Results of experiment II: dynamic memory based experiment

    After training on D2, the model performs well on D2 with a score of 0.87 but drops to a score of 0.65 when tested on D1.

    “After training on D 2 , the model performs well on D 2 with a score of 0.87 but drops to a score of 0.65 when tested on D 1 .”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § This strongly supports the hypothesis that integrating lifelong learning into GPT2 enhances its performance across different domains.

    The proposed model consistently achieves higher median accuracy and often demonstrates less variability in performance compared to the baseline GPT2.

    “The bar chart indicates that the proposed model consistently achieves higher median accuracy and often demonstrates less variability in performance compared to the baseline GPT2.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § Performance comparison of proposed model and original GPT2

    In the Finance domain, the proposed GPT2 with lifelong learning had a mean accuracy of 0.5750 versus 0.5300 for original GPT2, with a mean difference of +0.0450 and a one-tailed p-value of 0.074.

    “Finance 0.5750 0.5300 + 0.0450 0.074”
    Quote in the paperNot a substantial claim

  • The quote only partly states this claim.

    Findingown data
    § Part 5 Testing Generalization Under Extended Domain Sequences

    The model maintains highly consistent performance in the extended setting with the simulated fifth domain.

    “The experimental results demonstrate that the model maintains highly consistent performance in this extended setting.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § Conclusions

    The model achieved a Backward Transfer (BWT) score of -0.09, a notable improvement over the -0.64 score of the non-lifelong learning model.

    “Additionally, it achieved a Backward Transfer (BWT) score of -0.09, a notable improvement over the -0.64 score of the non-lifelong learning model”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § Performance comparison of proposed model and original GPT2

    In the Tweets domain, the proposed GPT2 with lifelong learning had a mean accuracy of 0.6550 versus 0.6075 for original GPT2, with a mean difference of +0.0475 and a one-tailed p-value of 0.080.

    “Tweets 0.6550 0.6075 + 0.0475 0.080”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Findingnone given
    § Data preprocessing

    Both the Laptop and Restaurant datasets contain entries with single and multiple Aspect-Sentiment pairs.

    “Although the initial datasets were structured, both the Laptop and Restaurant datasets contain entries with single and multiple Aspect-Sentiment pairs.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § Statistical evaluation of augmentation impact

    Augmented models exhibit both higher median accuracy and reduced lower quartile variance, particularly in cross-domain evaluations.

    “As visualized in Fig. 9 (see boxplot above), the augmented models exhibit both higher median accuracy and reduced lower quartile variance, particularly in cross-domain evaluations.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § Part 5 Testing Generalization Under Extended Domain Sequences

    When trained on D_laptop, the model's accuracy is 0.85 on D_laptop and drops to 0.68 on D_restaurant, 0.40 on D_tweets, and 0.43 on D_finance.

    “When trained on D laptop , the model performs optimally on D laptop itself (0.85), but its performance drops when tested on D restaurant (0.68), D tweets (0.40), and D finance (0.43).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § Results of experiment III: data augmentation

    The multi-level data augmentation pipeline produced consistent performance improvements, particularly in the Laptop and Restaurant domains, where both in-domain and cross-domain accuracy increased.

    “Results demonstrate consistent performance improvements, particularly in the Laptop and Restaurant domains, where both in-domain and cross-domain accuracy increased.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § Performance comparison of proposed model and original GPT2

    In the Restaurant domain, the proposed GPT2 with lifelong learning had a mean accuracy of 0.8075 versus 0.6750 for original GPT2, with a mean difference of +0.1325 and a one-tailed p-value of 0.070.

    “Restaurant 0.8075 0.6750 + 0.1325 0.070”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Findingnone given
    § Post-hoc comparative analysis with knowledge distillation-based lifelong learning

    The proposed framework demonstrates strong retention and generalization.

    “While the proposed framework demonstrates strong retention and generalization, future research should incorporate comparisons against stronger lifelong learning baselines (e.g., GEM, A-GEM) to further validate robustness.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning
    § Abstract

    The proposed framework integrates GPT-2 with a replay-based lifelong learning mechanism to support incremental, multi-domain ABSA while mitigating forgetting.

    “To address these issues, we proposed a novel framework that integrates GPT-2 with a replay-based lifelong learning mechanism to support incremental, multi-domain ABSA while mitigating forgetting.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given
    § Introduction

    The ANFIS is not used during the current experiments.

    “Although this ANFIS is not used during the current experiments, it is trained on our results and offers a promising direction for future work where the memory configuration can be adaptively tuned without manual intervention.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data
    § Abstract

    The model's scalability and ability to generalize beyond trained distributions were validated by the components.

    “These components validate the model's scalability and ability to generalize beyond trained distributions.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: A simulated fifth domain constructed from a mixture of all original domains adequately tests generalization beyond trained distributions.

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationreasoning
    § Introduction

    The proposed ANFIS provides an intelligent and data-driven mechanism that leverages prior experimental results across domains to predict the optimal memory ratio.

    “While many lifelong learning methods rely on manual tuning or exhaustive search to determine memory size, our proposed ANFIS provides an intelligent and data-driven mechanism that leverages prior experimental results across domains.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The ANFIS trained on the paper's experimental results can predict the proper memory ratio for new datasets.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisereasoning
    § SemEval-2014 Task 4

    DBLL is highly compatible with GPT-2 because its focus on data replay allows GPT-2 to maximize computational resources without dedicating significant power to the replay process.

    “Furthermore, DBLL is highly compatible with GPT-2, as its focus on data replay allows GPT-2 to maximize computational resources without dedicating significant power to the replay process.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning
    § Introduction

    Three persistent challenges in ABSA are limited generalization across diverse domains, difficulty assimilating new domain information without losing prior knowledge, and dynamically adapting memory usage in lifelong learning scenarios.

    “These observations reveal three persistent challenges in ABSA: (1) limited generalization across diverse domains, (2) difficulty assimilating new domain information without losing prior knowledge, and (3) dynamically adapting memory usage in lifelong learning scenarios.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning
    § Datasets

    The proposed architecture is a robust framework for lifelong ABSA.

    “Overall, the proposed architecture is a robust framework for lifelong ABSA.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The replay-based learning strategy consolidates prior knowledge while enabling adaptation to new tasks.
    • Rests on an assumption: Prior task performance indicators (precision, recall, F1-score, METEOR) are sufficient to infer the optimal replay memory ratio for a new domain.

  • May combine two claims.

    Interpretationown data
    § Conclusions

    The proposed model combines GPT-2 with a replay-based lifelong learning algorithm and significantly enhances generalization across diverse domains.

    “This paper demonstrates the successful development and evaluation of our proposed model, which combines GPT-2 with a replay-based lifelong learning algorithm to significantly enhance generalization across diverse domains.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Premisereasoning
    § SemEval-2014 Task 4

    GPT-2's advanced autoregressive architecture plays a key role in tackling domain generalization issues, as it excels at generating and understanding complex text patterns across a wide range of sequences.

    “GPT-2's advanced autoregressive architecture plays a key role in tackling domain generalization issues, as it excels at generating and understanding complex text patterns across a wide range of sequences.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given
    § Adaptive neuro-fuzzy inference systems

    Integrating the replay mechanism with GPT-2 allows the model to continuously adapt to new data without significant performance degradation on previously learned tasks.

    “Integrating this replay mechanism with GPT-2 allows the model to continuously adapt to new data without significant performance degradation on previously learned tasks.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The replay-based learning strategy consolidates prior knowledge while enabling adaptation to new tasks.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationreasoning
    § Performance comparison

    The Laptop and Restaurant domains' more severe forgetting without sufficient replay support is likely due to their reliance on more domain-specific linguistic patterns.

    “likely due to their reliance on more domainspecific linguistic patterns.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning
    § Performance comparison

    Memory replay has a dual benefit: not only preventing forgetting, but also enhancing cross-domain reinforcement.

    “This reflects a dual benefit of memory replay-not only preventing forgetting, but also enhancing cross-domain reinforcement.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data
    § Part 4 Few-Shot Domain Adaptation Setup

    While slight drops were observed in more challenging transfer scenarios such as Finance to Laptop, the overall trend supports the model's resilience to supervision constraints.

    “While slight drops were observed in more challenging transfer scenarios-such as Finance to Laptop-the overall trend supports the model's resilience to supervision constraints.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning
    § SemEval-2014 Task 4

    The replay + ANFIS-based method provides a lightweight, scalable solution without requiring model architecture changes or heavy pre-training.

    “In contrast, our replay + ANFIS-based method provides a lightweight, scalable solution without requiring model architecture changes or heavy pre-training.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § Post-hoc comparative analysis with knowledge distillation-based lifelong learning

    The Tweets dataset was introduced in the paper 'Adaptive Recursive Neural Network for Target-dependent Twitter Sentiment Classification' by Li Dong et al., presented at ACL 2014.

    “The Tweets dataset was introduced in the paper”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Li Dong et al.

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Interpretationown data
    § Post-hoc comparative analysis with knowledge distillation-based lifelong learning

    The additional evaluation confirmed that the replay + ANFIS framework achieved superior accuracy and lower forgetting, reinforcing its suitability for scalable, cross-domain ABSA.

    “This additional evaluation confirmed that our replay + ANFIS framework achieved superior accuracy and lower forgetting, reinforcing its suitability for scalable, cross-domain ABSA.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Interpretationnone given
    § Architecture of model

    The architecture is designed to incrementally acquire new domain-specific knowledge while preserving previously learned information, thereby addressing catastrophic forgetting in sequential domain adaptation.

    “The architecture is designed to incrementally acquire new domain-specific knowledge while preserving previously learned information, thereby addressing the issue of catastrophic forgetting typically encountered in sequential domain adaptation tasks.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Catastrophic forgetting is typically encountered in sequential domain adaptation tasks.

  • Classified as probably not a substantial claim.

    Study design and conductreasoning
    § Part 4 Few-Shot Domain Adaptation Setup

    The evaluation's reliability and stability were ensured by repeating the 30% sampling process independently three times and reporting the average performance across runs.

    “To ensure the reliability and stability of the evaluation, this sampling process was repeated independently three times, and the final performance was reported as the average across these runs.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data
    § Part 5 Testing Generalization Under Extended Domain Sequences

    The model successfully captures domain-invariant features and adapts effectively to novel but semantically aligned domain combinations.

    “These findings indicate that the model successfully captures domain-invariant features and adapts effectively to novel but semantically aligned domain combinations.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Study design and conductown data
    § Conclusions

    The achievements were systematically validated across multiple rigorous experimental setups, including few-shot domain adaptation and extended domain sequence generalization, and further illuminated by an ablation analysis of replay memory ratios.

    “These achievements were systematically validated across multiple rigorous experimental setups, including fewshot domain adaptation and extended domain sequence generalization, and further illuminated by an ablation analysis of replay memory ratios, which provided empirical foundations for adaptive memory management.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data
    § Part 5 Testing Generalization Under Extended Domain Sequences

    The experiment validates the robustness and scalability of the approach compared to the tested baseline and confirms its applicability to dynamic and evolving domain settings typical of real-world lifelong learning scenarios.

    “Overall, this experiment validates the robustness and scalability of our approach compared to the tested baseline and confirms its applicability to dynamic and evolving domain settings typical of real-world lifelong learning scenarios.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisereasoning
    § SemEval-2014 Task 4

    Replay-based lifelong learning prevents catastrophic forgetting by storing and replaying data from previous tasks during training on new tasks.

    “Replay-based lifelong learning is a subset of DBLL methods that prevents catastrophic forgetting by storing and replaying data from previous tasks during training on new tasks.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisereasoning
    § Experiment II: dynamic memory based experiment

    A transformer layer has time complexity O(L^2 • d) due to computing pairwise interactions between all tokens in the input sequence.

    “Transformer layer has a time complexity of O ( L 2 • d ) , due to the necessity of computing pairwise interactions between all tokens in the input sequence.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: The complexity analysis assumes the conventional quadratic-cost attention mechanism.

  • Classified as probably not a substantial claim.

    Study design and conductreasoning
    § Part 5: Testing Generalization Under Extended Domain Sequences

    A simulated fifth domain constructed by combining samples from the four domains is controlled to ensure semantic diversity and label balance, so that it reflects a realistic but unseen distribution shift.

    “To assess model behavior when trained on more than four domains, a simulated fifth domain is constructed by combining samples from the four domains.The sampling process is controlled to ensure semantic diversity and label balance, allowing the simulated domain to reflect a realistic but unseen distribution shift.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given
    § Computational complexity

    The observed performance degradation over time on earlier domains provides evidence of catastrophic forgetting.

    “The self-attention operation in each performance degrades over time on earlier domains, providing evidence of catastrophic forgetting.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisenone given
    § SemEval-2014 Task 4

    Balancing old and new knowledge remains a challenge in replay-based learning.

    “Despite the resource efficiency of replay-based learning, balancing old and new knowledge remains a challenge, which can affect learning outcomes.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning
    § Part 4 Few-Shot Domain Adaptation Setup

    This design helps mitigate the potential variance introduced by a single random split and offers a more robust estimate of the model's behavior when training data is scarce.

    “This design helps mitigate the potential variance introduced by a single random split and offers a more robust estimate of the model's behavior when training data is scarce.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data
    § Post-hoc comparative analysis with knowledge distillation-based lifelong learning

    The chosen approach offers a balanced and scalable solution.

    “the results here provide empirical justification that our chosen approach offers a balanced and scalable solution”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationown data
    § Lifelong learning model accuracy

    There is a potential beneficial overlap between the domains that aids in memory retention.

    “suggesting a potential beneficial overlap between the domains that aids in memory retention.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationreasoning
    § Datasets

    The generative QA design with GPT-2 eliminates the need for task-specific architectural changes, leveraging GPT's flexibility for both aspect extraction and sentiment classification.

    “This generative design eliminates the need for task-specific architectural changes, leveraging GPT's flexibility for both aspect extraction and sentiment classification.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data
    § Results of experiment II: dynamic memory based experiment

    Higher replay enables effective knowledge retention and reduces catastrophic forgetting.

    “suggesting that higher replay enables effective knowledge retention and reduces catastrophic forgetting.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationreasoning
    § Steps 4, 5 and 6. System Integration and Application

    Once trained, the ANFIS model can be applied to a new domain dataset by feeding in its expected F1-score, Precision, Recall, and METEOR score to produce a suggested reply memory ratio.

    “Once trained, the ANFIS model can be applied to a new domain dataset by feeding in its expected F1-score, Precision, Recall, and METEOR score (e.g., from small-scale pilot runs or estimated values based on domain similarity).”
    • Also stated in: § Steps 4, 5 and 6. System Integration and Application · plain
      “In Step 5, we enter these measures to ANFIS then in step 6 the fuzzy rules will be applied to inputs and ANFIS will produce the proper reply memory ratio for new dataset as output.”
    • Also stated in: § Steps 4, 5 and 6. System Integration and Application · hedged
      “A suggested memory ratio, allowing adaptive tuning of the lifelong learning system's replay memory.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The four input performance metrics (F1-score, Precision, Recall, METEOR) are sufficient to predict the optimal reply memory ratio for a new domain.
    • Rests on an assumption: The ANFIS model trained on 15 records from three domains will generalize to new, unseen domains.

  • May combine two claims.

    Interpretationown data
    § Conclusions

    The model effectively learns new domain knowledge while minimizing catastrophic forgetting.

    “The model effectively learns new domain knowledge while minimizing catastrophic forgetting.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Interpretationreasoning
    § Post-hoc comparative analysis with knowledge distillation-based lifelong learning

    Unlike meta-learning approaches, the replay + ANFIS strategy offers practical benefits in terms of interpretability, adaptability, and computational cost.

    “Moreover, unlike meta-learning approaches which often require task-specific adaptation strategies and episodic training, the replay + ANFIS strategy offers practical benefits in terms of interpretability, adaptability, and computational cost.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Meta-learning approaches often require task-specific adaptation strategies and episodic training.

  • May combine two claims.

    Interpretationown data
    § Post-hoc comparative analysis with knowledge distillation-based lifelong learning

    Integrating GPT-2 with a replay-based lifelong learning method, supported by adaptive memory management and data augmentation, effectively mitigates catastrophic forgetting in ABSA while enabling the model to continuously acquire and adapt to new domain knowledge.

    “Our study concludes that integrating GPT-2 with a replaybased lifelong learning method, supported by both adaptive memory management and data augmentation, effectively mitigates catastrophic forgetting in aspect-based sentiment analysis (ABSA) while enabling the model to continuously acquire and adapt to new domain knowledge.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given
    § Statistical evaluation of augmentation impact

    Finance-trained models showed robust performance across all domains.

    “Finance-trained models showed robust performance across all domains.”
    Quote in the paperNot a substantial claim

  • The quote only partly states this claim.

    Interpretationown data
    § Performance comparison of proposed model and original GPT2

    The proposed GPT2 with lifelong learning showed a statistically significant improvement over original GPT2 across four domains (Laptop, Restaurant, Tweets, Finance).

    “The results showed a statistically significant improvement”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning
    § Domain Transition

    The minor improvement following Dfinance training may be attributed to the diverse Dfinance dataset, which includes over 10,000 entries mainly from the financial news domain, possibly contributing to a small increase in the METEOR score.

    “The minor improvement following the D finance training may be attributed to the diverse dataset, which includes over 10,000 entries mainly from the financial news domain, possibly contributing to a small increase in the METEOR score.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data
    § Conclusions

    The BWT score indicates strong accuracy and effective lifelong learning capabilities.

    “indicating strong accuracy and effective lifelong learning capabilities”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning
    § Experiment II: dynamic memory based experiment

    The proposed framework demonstrates polynomial time and space complexity, making it suitable for scalable and efficient lifelong learning on aspect-based sentiment tasks across diverse domains.

    “In summary, the proposed framework demonstrates polynomial time and space complexity, making it suitable for scalable and efficient lifelong learning on aspect-based sentiment tasks across diverse domains.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: The complexity analysis assumes the conventional quadratic-cost attention mechanism.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductnone given
    § ANFIS building and applying for reply memory ratio prediction

    The integrity of the augmented data is ensured by employing two evaluation metrics to assess augmentation quality, though they are not used for filtering.

    “To ensure the integrity of the augmented data, two evaluation metrics are employed to assess augmentation quality, though they are not used for filtering.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationown data
    § Post-hoc comparative analysis with knowledge distillation-based lifelong learning

    This vertical GPT model offers a scalable, intelligent solution for researchers and practitioners seeking persistent domain adaptation in evolving environments.

    “This vertical GPT model offers a scalable, intelligent solution for researchers and practitioners seeking persistent domain adaptation in evolving environments.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data
    § Part 4 Few-Shot Domain Adaptation Setup

    These findings confirm that our method is not only effective in high-resource conditions but also robust in practical low-resource scenarios, where labeled data may be limited.

    “These findings confirm that our method is not only effective in high-resource conditions but also robust in practical low-resource scenarios, where labeled data may be limited.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: 30% of the training data constitutes a low-resource scenario.

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Interpretationreasoning
    § Data preprocessing

    This predictive memory management supports efficient knowledge retention and performance across evolving domains.

    “This predictive memory management supports efficient knowledge retention and performance across evolving domains.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Interpretationreasoning
    § Steps 4, 5 and 6. System Integration and Application

    The ANFIS-produced memory ratio enables dynamic memory control tailored to the performance profile of the incoming dataset, optimizing resource utilization while preserving learning efficacy.

    “This enables dynamic memory control tailored to the performance profile of the incoming dataset, optimizing resource utilization while preserving learning efficacy.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The four input performance metrics (F1-score, Precision, Recall, METEOR) are sufficient to predict the optimal reply memory ratio for a new domain.
    • Rests on an assumption: The ANFIS model trained on 15 records from three domains will generalize to new, unseen domains.

  • Classified as probably not a substantial claim. May combine two claims.

    Premisereasoning
    § Ablation study: disentangling the impact of replay memory ratios

    Higher replay ratios inherently utilize more past samples in the buffer, so replay ratio serves as a proxy for memory size.

    “This design serves as a proxy for memory size analysis, as higher replay ratios inherently utilize more past samples in the buffer.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationreasoning
    § Data collection

    Integrating ANFIS as a predictive component in the lifelong learning process makes the training method more adaptive and data-driven.

    “By integrating ANFIS as a predictive component in the lifelong learning process, the training method becomes more adaptive and data-driven.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationreasoning
    § Datasets

    By combining a pretrained GPT backbone, ANFIS-guided memory replay, and flexible prompt-based inference, the proposed architecture offers a scalable and intelligent solution for dynamic sentiment analysis in evolving multi-domain environments.

    “By combining a pretrained GPT backbone, ANFIS-guided memory replay, and flexible promptbased inference, it offers a scalable and intelligent solution for dynamic sentiment analysis in evolving multi-domain environments.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The replay-based learning strategy consolidates prior knowledge while enabling adaptation to new tasks.
    • Rests on an assumption: Prior task performance indicators (precision, recall, F1-score, METEOR) are sufficient to infer the optimal replay memory ratio for a new domain.

  • May combine two claims.

    Interpretationown data
    § Statistical evaluation of augmentation impact

    The observed gains are unlikely to be due to chance and reflect a meaningful enhancement in model performance.

    “This suggests that the observed gains are unlikely to be due to chance and reflect a meaningful enhancement in model performance.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductreasoning
    § Experiment II: dynamic memory based experiment

    The Restaurant domain was fixed as a control point while varying memory allocations for the remaining domains to isolate the influence of memory size on each domain's performance.

    “To analyze domain-specific sensitivity to replay memory, the Restaurant domain (D 2 ) was fixed as a control point while varying the memory allocations for the remaining domains.This configuration was designed to isolate the influence of memory size on each domain's performance.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning
    § Experiment II: dynamic memory based experiment

    For a model with H layers, the per-step computation is O(H • L^2 • d).

    “Since the model contains 𝐻 layers, the per-step computation becomes: O ( H • L 2 • d )”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: The complexity analysis assumes the conventional quadratic-cost attention mechanism.

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Interpretationreasoning
    § Data collection

    This dynamic replay strategy improves both learning efficiency and memory retention across domains, providing a significant advancement over static replay configurations.

    “This dynamic replay strategy improves both learning efficiency and memory retention across domains, providing a significant advancement over static replay configurations.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Interpretationreasoning
    § Lifelong learning model accuracy

    Other domain-specific factors play a significant role in influencing the model's accuracy.

    “This points to the possibility that other domain-specific factors play a significant role in influencing the model's accuracy.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data
    § Part 5 Testing Generalization Under Extended Domain Sequences

    Overall, the model demonstrates strong proficiency in each domain and retains previously acquired knowledge, effectively addressing catastrophic forgetting while maintaining good performance.

    “Overall, the model demonstrates strong proficiency in each domain and retains previously acquired knowledge, effectively addressing catastrophic forgetting while maintaining good performance.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The four domains (Laptop, Restaurant, Tweets, Finance) are representative of domain adaptation tasks for aspect-based sentiment analysis.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning
    § Performance comparison of proposed model and original GPT2

    The ANFIS recommendation is derived from the comprehensive interaction of the learned fuzzy rules and membership functions.

    “This decision is derived from the comprehensive interaction of the learned fuzzy rules and membership functions.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Interpretationreasoning
    § Research method

    The model leverages prior task performance indicators to infer appropriate memory allocation, offering a practical and interpretable approach for enhancing memory management in continual learning scenarios.

    “The model leverages prior task performance indicators to infer appropriate memory allocation, offering a practical and interpretable approach for enhancing memory management in continual learning scenarios.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Prior task performance indicators (precision, recall, F1-score, METEOR) are sufficient to infer the optimal replay memory ratio for a new domain.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning
    § Experiment II: dynamic memory based experiment

    The total time complexity across all T domains is O(T • E(B/bs) • H • L^2 • d).

    “Therefore, the total time complexity across all 𝑇 domains is given by: O ( T • E (B/bs) • H • L 2 • d )”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: The complexity analysis assumes the conventional quadratic-cost attention mechanism.
95 of 95

  • The paper relies on this without support.

    The four selected ABSA domains (Laptops, Restaurants, Tweets, Finance) are representative of diverse multi-domain ABSA and support generalization to continuous, multi-domain sentiment analysis.

    “The model is sequentially fine-tuned using real-data replay across four diverse ABSA domains: Laptops, Restaurants, Tweets, and Finance.”

  • The paper relies on this without support.

    A simulated fifth domain constructed from a mixture of all original domains adequately tests generalization beyond trained distributions.

    “another part evaluates generalization under a simulated fifth domain constructed from a mixture of all original domains.”

  • The paper relies on this without support.

    The t-test assumptions are met for the performance improvements.

    “We further conducted a t-test to validate the statistical significance of the improvements.”

  • The paper relies on this without support.

    The ANFIS trained on the paper's experimental results can predict the proper memory ratio for new datasets.

    “we design an ANFIS-based model trained on experimental results to predict the proper memory ratio for new datasets, enabling more effective and adaptive lifelong learning in GPT-based architectures.”

  • The paper relies on this without support.

    Removing a component and observing a performance drop indicates that component contributes meaningfully to overall effectiveness.

    “The results demonstrate that removing any single component leads to a measurable drop in performance, confirming that each part of the framework contributes meaningfully to overall effectiveness.”

  • The paper relies on this without support.

    The replay-based learning strategy consolidates prior knowledge while enabling adaptation to new tasks.

    “This technique consolidates prior knowledge while enabling adaptation to new tasks.”

  • The paper relies on this without support.

    Adaptive decision-making in dynamic environments requires models that are both interpretable and capable of handling uncertainty.

    “Adaptive decision-making in dynamic environments requires models that are both interpretable and capable of handling uncertainty.”

  • The paper relies on this without support.

    Prior task performance indicators (precision, recall, F1-score, METEOR) are sufficient to infer the optimal replay memory ratio for a new domain.

    “The model leverages prior task performance indicators to infer appropriate memory allocation”

  • The paper relies on this without support.

    Catastrophic forgetting is typically encountered in sequential domain adaptation tasks.

    “catastrophic forgetting typically encountered in sequential domain adaptation tasks”

  • The paper relies on this without support.

    Each domain presents unique aspect-sentiment pairs and linguistic characteristics.

    “Each domain presents unique aspect-sentiment pairs and linguistic characteristics”

  • The paper relies on this without support.

    Incorporating an entirely external domain for testing generalization brings confounding factors such as annotation inconsistency, mismatched label distributions, and domain-specific terminology, making fair comparison with the original benchmark difficult.

    “While incorporating an entirely external domain might appear ideal for testing generalization, it often brings confounding factors such as annotation inconsistency, mismatched label distributions, and domain-specific terminology.These differences make fair comparison with the original benchmark difficult.”

  • The authors state this as an assumption.

    The complexity analysis assumes the conventional quadratic-cost attention mechanism.

    “It should be noted that our analysis assumes the conventional quadratic-cost attention mechanism, which provides a clear baseline for comparison.”

  • The paper relies on this without support.

    The four ANFIS inputs (F1-score, Precision, Recall, METEOR) collectively represent the model's generalization capability, semantic understanding, and classification performance, and are influenced by memory size in lifelong learning.

    “These four inputs were selected because they collectively represent the model's generalization capability, semantic understanding, and classification performance, all of which are influenced by memory size in lifelong learning.”

  • The paper relies on this without support.

    The four input performance metrics (F1-score, Precision, Recall, METEOR) are sufficient to predict the optimal reply memory ratio for a new domain.

    “Once trained, the ANFIS model can be applied to a new domain dataset by feeding in its expected F1-score, Precision, Recall, and METEOR score (e.g., from small-scale pilot runs or estimated values based on domain similarity).”
    “In Step 5, we enter these measures to ANFIS then in step 6 the fuzzy rules will be applied to inputs and ANFIS will produce the proper reply memory ratio for new dataset as output.”

  • The paper relies on this without support.

    The ANFIS model trained on 15 records from three domains will generalize to new, unseen domains.

    “Once trained, the ANFIS model can be applied to a new domain dataset by feeding in its expected F1-score, Precision, Recall, and METEOR score (e.g., from small-scale pilot runs or estimated values based on domain similarity).”
    “This enables dynamic memory control tailored to the performance profile of the incoming dataset, optimizing resource utilization while preserving learning efficacy.”

  • The paper relies on this without support.

    A decrease in METEOR score for the initial domain after exposure to all domains indicates catastrophic forgetting.

    “confirming the presence of catastrophic forgetting.”

  • The paper relies on this without support.

    A moderately stable METEOR score indicates that the model struggled to improve.

    “indicating that the model struggled to improve before learning domain-specific knowledge from the tweets data.”

  • The authors state this as an assumption.

    30% of the training data constitutes a low-resource scenario.

    “To simulate a low-resource scenario, we randomly sampled 30% of the training data from each of the four source domains-Laptop, Restaurant, Tweets, and Finance-using a fixed seed.”

  • The paper relies on this without support.

    The four domains (Laptop, Restaurant, Tweets, Finance) are representative of domain adaptation tasks for aspect-based sentiment analysis.

    “To simulate a low-resource scenario, we randomly sampled 30% of the training data from each of the four source domains-Laptop, Restaurant, Tweets, and Finance-using a fixed seed.”

  • The paper relies on this without support.

    The original GPT2 model is a valid baseline for comparison.

    “However, when combining all data, the proposed GPT2 achieved a significantly higher mean accuracy (M = 0.104) than the original GPT2 (t = 3.77, p = 0.0009), supporting the overall effectiveness of the proposed method.”

  • The paper relies on this without support.

    Original GPT2 is an appropriate baseline for evaluating the proposed GPT2 with lifelong learning.

    “A one-tailed paired t-test was conducted to evaluate whether the proposed GPT2 with lifelong learning significantly outperforms the original GPT2 across four domains.”

  • The paper relies on this without support.

    The performance indicators (F1-score, Precision, Recall, METEOR) are sufficient to determine the optimal memory ratio for a dataset.

    “given a new dataset with F1-score of 0.52, Precision of 0.52, Recall of 0.49, and METEOR score of 0.59, the ANFIS recommends a memory ratio of 0.5.”

  • The paper relies on this without support.

    The four domains (Laptop, Restaurant, Tweets, Finance) are representative of domain-specific sentiment analysis tasks.

    “Table 8 presents the average accuracy of each model and the statistical outcomes for each of the four domains: Laptop, Restaurant, Tweets, and Finance.”

  • The paper relies on this without support.

    The observed optimal memory ratios in the experiments are the true optimal ratios.

    “Table 10 summarizes the actual best-performing memory ratios based on observed metrics and the predicted memory ratios generated by the ANFIS model:”

  • The paper relies on this without support.

    The structured augmentation pipeline enriches semantic and lexical variability without compromising task integrity.

    “The augmentation strategy comprises chained multilingual back-translation, T5-based paraphrasing, and lexical-level EDA, aiming to enrich semantic and lexical variability without compromising task integrity.”

  • The authors state this as an assumption.

    A teacher model trained on all domain data in a single batch setting represents an upper bound for performance.

    “A teacher model was trained using all domain data in a single batch setting (representing an upper bound).”

  • The paper relies on this without support.

    Meta-learning approaches often require task-specific adaptation strategies and episodic training.

    “unlike meta-learning approaches which often require task-specific adaptation strategies and episodic training”
27 of 27

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Augmentation significance is reported with two different p-values

    Major

    The abstract and conclusion report p = 0.0049 for the augmentation improvement, whereas the augmentation results report p = 0.036 for the paired test across 16 domain pairs.

  • ANFIS is both excluded from and credited for the experiments

    Major

    The introduction explicitly says ANFIS was not used during the current experiments, but the extended-domain experiment attributes its results in part to the ANFIS-based adaptive memory controller.

  • Lifelong-learning table repeats the wrong cell label

    Minor

    Table 5 labels the D1-training, D2-testing cell R1,1, although its own definition requires R1,2 for a model trained on D1 and tested on D2.

  • Laptop-to-Tweets few-shot example conflicts with the results table

    Minor

    The text reports 0.88 for Laptop-to-Tweets with 30% training data, but Table 6 gives F1 = 0.28 for that training/testing direction in Part 4. The reported 0.88 instead matches Tweets tested on Tweets.

  • Reported backward-transfer scores omit the stated averaging step

    Minor

    The stated BWT formula averages changes across the three earlier domains. Table 7's entries give (-0.05 − 0.03 − 0.01)/3 = -0.03 with lifelong learning and (-0.24 − 0.27 − 0.13)/3 ≈ -0.21 without it; the reported -0.09 and -0.64 are the unaveraged sums.

5 of 5

Cites

Flagged first, then by citations

Research works printed but not linked: 13

  1. Yoon J, Yang E, Lee J, Hwang SJ (2017) Lifelong learning with dynamically expandable networks. ArXiv Preprint arXiv:170801547
  2. Shin H, Lee JK, Kim J, Kim J (2017) Continual learning with deep generative replay. Adv Neural Inf Process Syst, 30
  3. an de Ven GM, Tolias AS (2018) Generative replay with feedback connections as a general strategy for continual learning. ArXiv Preprint arXiv:180910635
  4. Liu L, Zhang Y, Liu Q (2021) CLASSIC: Continual and contrastive learning for aspect sentiment classification. In: Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics (ACL) (pp. 3720–3730). https://doi.org/10.18653/v1/2021.acl-main.293 · DOI
  5. Kirkpatrick J, Pascanu R, Rabinowitz N, Veness J, Desjardins G, Rusu AA, Milan K, Quan J et al (2017) Tiago Ramalho, Agnieszka Grabska-Barwinska,. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521–3526 · DOI · Paper page
  6. Zhang J, Fu Y, Peng Z, Yao D, He K (2024) Core: mitigating catastrophic forgetting in continual learning through cognitive replay. ArXiv preprintarXiv:240201348
  7. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., … Liu, P. J.(2020). Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research 21(140):1–67.
  8. Xie Q, Dai Z, Hovy E, Luong M-T, Le QV (2020) Unsupervised data augmentation for consistency training. Adv Neural Inform Process Syst (NeurIPS) 33:6256–6268
  9. Hoang M, Bihorac OA, Rouces J (2019) Aspect-based sentiment analysis using bert. In: Proceedings of the 22nd nordic conference on computational linguistics (pp. 187–196)
  10. Zenke F, Poole B, Ganguli S (2017) Continual learning through synaptic intelligence. In: International conference on machine learning (pp. 3987–3995)
  11. Chaudhry A, Ranzato M, Rohrbach M, Elhoseiny M (2018) Efficient lifelong learning with a-gem. ArXiv Preprint arXiv:181200420
  12. Li H, Wu J, Braverman V (2023) Fixed design analysis of regularization-based continual learning. In: Conference on lifelong learning agents (pp. 513–533)
  13. Raghu A, Raghu M et al (2020) Rapid learning or feature reuse? Towards understanding the effectiveness of MAML. International Conference on Learning Representations (ICLR)

Related

Authors
Huang, H., Ye, X., Mustafa, M. B., Dong, Q., Li, Y., Asemi, A., & Asemi, A.
Affiliations
  • Wenzhou University of Technology (CN)
  • University of Malaya (MY)
  • Illinois Institute of Technology (US)
ORCID
Venue
Complex & Intelligent Systems
Publisher
Springer Science and Business Media LLC
Published
2025-10-10
Volume
11
Issue
11
ISSN
  • 2198-6053 (ISSN-L)
  • 2199-4536
DOI
10.1007/s40747-025-02086-2
OpenAlex
W4415040416
Cited by
0
References
33
FWCI
0.00
Type
journal article
Language
en
Open access
Licence
Locations
  • Complex & Intelligent Systems (publishedVersion, cc-by-nc-nd)
  • Corvinus Research Archive (Corvinus University of Budapest) (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Sentiment Analysis and Opinion Mining
  • Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
  • Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
Keywords
aspect-based sentiment analysis, lifelong learning, catastrophic forgetting, GPT-2, adaptive neuro-fuzzy inference system, backward transfer, data augmentation, domain generalization, knowledge retention, large language models, tweets

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-29.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-29.

Cited sources are checked by gpt-6-sol.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review