RRReza Ravanmehr

Reza Ravanmehr

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article

Previous Organizations / Institutes

The organizations / institutes from the scholar's other articles in the past ten years


Reza Ravanmehr has a total of 59 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Flexible-TADA: Mitigating Representation Collapse in Task-Agnostic Domain Adaptation

    by Sepehr Mostafavi, Reza Ravanmehr, Hossein Saberi

    Full fine-tuning (FFT) of large pre-trained language models (PLMs) is computationally expensive, motivating the development of parameter-efficient fine-tuning (PEFT) methods. However, PEFT approaches that update only the input representations may suffer from degraded intermediate representations when the transformer backbone remains entirely frozen, limiting performance on complex language understanding tasks. To address this limitation, this paper proposes Flexible-TADA, a PEFT framework that updates the input embeddings and the final transformer layer while keeping the remaining backbone frozen. The proposed dual-unfreezing strategy enables the final layer to adapt task-specific representations before classification without substantially increasing the number of trainable parameters or inference cost. Flexible-TADA is evaluated on the GLUE benchmark using four encoder architectures (BERT, RoBERTa, DeBERTa, and ELECTRA) and is compared with FFT, LoRA, and the original TADA framework. Experimental results show that Flexible-TADA consistently improves performance over the baseline TADA and achieves performance comparable to LoRA while maintaining zero additional inference latency. Furthermore, explainable AI (XAI) analyses indicate improved semantic alignment of the learned representations. These results suggest that Flexible-TADA provides an effective trade-off between parameter efficiency, predictive performance, and model interpretability.

  • Open Access

    An Adaptive Neuro Symbolic Framework for Explainable Sentiment Analysis

    by Sepehr Mostafavi, Reza Ravanmehr

    Sentiment analysis serves as a pivotal tool for opinion mining; however, state-of-the-art Large Language Models (LLMs) often lack the transparency required for sensitive applications due to their inherent black-box nature. While Neuro-symbolic AI bridges the gap between neural learning capabilities and symbolic logical reasoning to enhance trustworthiness, adapting such models to new domains remains challenged by prohibitive computational costs and the difficulty of balancing accuracy with interpretability. To address these limitations, this study proposes a novel adaptive Neuro-symbolic framework that leverages Parameter-Efficient Fine-Tuning (PEFT) and the proposed Flexible TADA architecture to inject structured information from knowledge graphs via a specialized knowledge-aware adapter. Remarkably, this approach facilitates efficient domain adaptation by updating fewer than 20% of the total model parameters. Comprehensive empirical evaluations across diverse backbones (BERT, GPT-2, and Qwen2) on standard benchmarks (SST-2 and IMDB) demonstrate that the proposed method yields significant gains in Accuracy, Precision, and Recall. Specifically, it achieves a notable F1-score of 93.43%on long-text tasks. Beyond performance metrics, results in the interpretability dimension indicate a marked improvement in Sufficiency and a substantial reduction in Infidelity, substantiating the method's robustness and generalizability across different knowledge sources, such as SenticNet and ConceptNet.

  • Improving the Reliability of Interconnection Networks Using Replicated 4-Disjoint GIN

    by Mahsa Esmaeili, Reza Ravanmehr

    Today, high-performance computing is recognized as a necessity in various industries. The stable and secure connection between system components, such as CPUs and memories, requires efficient, reliable, and fault-tolerant networks. Multistage Interconnection Networks (MIN) are fundamental approaches to create such an interconnection facility among researchers and engineers. Meanwhile, the Gamma Interconnection Network (GIN) provides one of the highest levels of fault tolerance networks among MINs by establishing multiple paths. Considering the importance of disjoint paths, 4-Disjoint GIN (4DGIN) has been developed, which is an important type of GIN and provides four disjoint paths for all source–destination pairs. On the other hand, the replication approach improves network fault tolerance, which directly affects network reliability. In this paper, a replicated 4DGIN is proposed, enhancing the reliability and improving the fault-tolerance factor by creating disjoint and redundant paths. The reliability of the proposed network is evaluated using the Reliability Block Diagram (RBD) approach. The reliability evaluation results show that the proposed network outperforms several well-known MINs, such as different variations of SEN and Benes networks in terms of the terminal, broadcast, and network reliability factors. Moreover, it improves the cost per unit factor in comparison with the baselines.

  • SSTSA: A Self-Supervised Topic Sentiment Analysis Using Semantic Similarity Measures and Transformers

    by Azam Seilsepour, Reza Ravanmehr, Ramin Nassiri

    The exponentially increasing amount of data generated by the public on social media platforms is a precious source of information. It can be used to find the topics and analyze the comments. Some researchers have extended the Latent Dirichlet Allocation (LDA) method by adding a sentiment layer to simultaneously find the topics and their related sentiments. However, most of these approaches do not achieve admirable accuracy in Topic Sentiment Analysis (TSA), particularly when there is insufficient training data or the texts are complex, ambiguous, and short. In this paper, a self-supervised novel approach called SSTSA is proposed for TSA that extracts the hidden topics and analyzes the total sentiment related to each topic. The SSTSA proposes a new method called Pseudo-label Generator. For this purpose, first, it employs semantic similarity and Word Mover’s Distance (WMD) measures. Then, the document embedding technique is employed to semantically estimate the sentiment orientation of samples and generate the pseudo-labels (positive or negative). Afterward, a hybrid classifier composed of a pre-trained Robustly Optimized BERT (RoBERTa) and a Long Short-Term Memory (LSTM) model is trained to predict the sentiment of unseen data. The evaluation results on different datasets of various domains demonstrate that the SSTSA outperforms similar unsupervised/self-supervised methods.


  • Topic sentiment analysis based on deep neural network using document embedding technique

    by Azam Seilsepour, Reza Ravanmehr, Ramin Nassiri

    Sentiment Analysis (SA) is a domain- or topic-dependent task since polarity terms convey different sentiments in various domains. Hence, machine learning models trained on a specific domain cannot be employed in other domains, and existing domain-independent lexicons cannot correctly recognize the polarity of domain-specific polarity terms. Conventional approaches of Topic Sentiment Analysis perform Topic Modeling (TM) and SA sequentially, utilizing the previously trained models on irrelevant datasets for classifying sentiments that cannot provide acceptable accuracy. However, some researchers perform TM and SA simultaneously using topic-sentiment joint models, which require a list of seeds and their sentiments from widely used domain-independent lexicons. As a result, these methods cannot find the polarity of domain-specific terms correctly. This paper proposes a novel supervised hybrid TSA approach, called Embedding Topic Sentiment Analysis using Deep Neural Networks (ETSANet), that extracts the semantic relationships between the hidden topics and the training dataset using Semantically Topic-Related Documents Finder (STRDF). STRDF discovers those training documents in the same context as the topic based on the semantic relationships between the Semantic Topic Vector, a newly introduced concept that encompasses the semantic aspects of a topic, and the training dataset. Then, a hybrid CNN–GRU model is trained by these semantically topic-related documents. Moreover, a hybrid metaheuristic method utilizing Grey Wolf Optimization and Whale Optimization Algorithm is employed to fine-tune the hyperparameters of the CNN–GRU network. The evaluation results demonstrate that ETSANet increases the accuracy of the state-of-the-art methods by 1.92%.


  • A service-oriented framework for remote sensing big data processing

    by Roohollah Enayati, Reza Ravanmehr, Vahe Aghazarian