Elysian is a modular architecture designed to generate emotionally expressive and contextually coherent dialogue for non-playable characters (NPCs) in interactive games. As the first contribution in a larger research program, this paper focuses on the system's dialogue generation module, which leverages large language models (LLMs) under strict runtime and hardware constraints. We investigate two practical approaches for enabling emotion-aware dialogue on consumer-grade devices: parameter-efficient LoRA fine-tuning and structured prompt orchestration. Targeting deployment on 6 GB VRAM hardware with sub-5-second response latency, we evaluate whether small, 4-bit-quantized 1B-parameter models can reliably produce emotionally aligned NPC dialogue. Across 1,000 generated samples, LoRA fine-tuning yields substantial performance gains in lexical alignment (BLEU: 0.0587 vs. 0.0006), semantic similarity (BERTScore: 0.8645 vs. 0.2712), and emotion-tag accuracy (100% vs. 0%) compared to prompting alone. Meanwhile, prompt orchestration demonstrates strong structural reliability, achieving 78% valid JSON generation for branching narrative formats and reaching 100% accuracy using a lightweight two-pass correction scheme. This work establishes the technical foundation for future research on multimodal NPC behavior, including integration with speech synthesis, behavioral animation, and visually grounded interaction.

Multi-dimensional sparse channel state information (CSI) acquisition is conceived for Orthogonal time frequency space (OTFS) modulation-based millimetre wave (mmWave) multiple input and multiple output (MIMO) systems. A comprehensive end-to-end relationship is derived in the delay-Doppler (DDA) domain by additionally considering the angular parameters and a hybrid beamforming (HB) architecture. A time-domain pilot model tailored for CSI estimation (CE) in the DDA-domain is proposed, which exploits the inherent multi-dimensional (4D) sparsity that emerges in the DDA-domain during the CE process. An efficient low-complexity Bayesian learning (LC-BL) technique is conceived to fulfil the objective of CSI estimation in such systems. Subsequently, a comprehensive examination of the complexity of the algorithm under consideration is also provided. It is worth noting that the complexity of the BL scheme designed is similar to that of popular orthogonal matching pursuit (OMP), but significantly lower than that of the traditional expectation-maximization (EM) based BL technique. Moreover, a single-stage transmit precoder (TPC) and receiver combiner (RC) design is proposed. This procedure aims for maximizing the directional gain of the RF TPC/RC pair by optimizing their weights. Additionally, a series of comprehensive simulations are conducted which incorporate the use of a practical channel model and fractional Doppler shifts. In light of the inherent trade-offs between complexity and estimation algorithm performance, our proposed scheme, LC-BL, appears suitable, especially considering the substantial enhancement in the performance of CE compared to the existing benchmarks.

Summary Cross‐domain sentiment classification is a significant task of sentiment analysis that objectives to predict the opinion orientation of text documents in the target domain by using the source domain's learned classifier. Most of the existing approaches of domain‐adaptation in sentiment classification focus on sharing low‐dimensional features across the domain using domain independent and specific features to mitigate the gap between domains. Earlier cross‐domain sentiment classification approaches mainly focused on document level and sentence level, they cannot consider the full impact of aspect words, position of the words, and long‐term dependencies. To address this concern, we propose a model for cross‐domain sentiment classification, which is based on decoding‐enhanced BERT with disentangled attention (DeBERTa). DeBERTa is a pretrained language model based on transformer architecture. In this article, we perform sentence and aspect embedding to mine wordpiece information from text document. DeBERTa language‐model utilize disentangled attention mechanism and an enhanced mask decoder to understand the expression features. Disentangled attention mechanism is used to encode each word into two vectors (i.e., content and position vector). In order to predict the masked tokens during model pretraining, an enhanced mask decoder is employed, which incorporates absolute positions in the decoding layer. Finally, experiments are conducted on the benchmark dataset that demonstrates the superiority of fine‐tuned DeBERTa model for cross‐domain sentiment classification tasks.

Abstract The purpose of cross‐domain opinion classification is to leverage useful information acquired from the source domain to train a classifier for opinion classification in the target domain, which has a huge amount of unlabeled data. An opinion classifier trained on a specific domain usually acts poorly, when directly employed to another domain. Annotating the data for all the domains is a laborious and costly process. The majority of available approaches are centered on identifying invariant features among domains. Unluckily, they are unable to properly capture the context within the sentences and better utilization of unlabeled data. To properly address this issue, we propose an aspect‐based attention model for cross‐domain opinion classification. By incorporating knowledge of aspects and sentences, the proposed model provides a transfer mechanism for better‐transferring opinions among domains. We introduce two learning networks, first learning network aims to recognize the shared features between domains, while the purpose of the second learning network is to extract the information from the aspects by utilizing shared words as a bridge. We benefit from BERT and bidirectional gated recurrent unit to get a deep understanding and deep level semantic information of the text. Further, the joint attention learning mechanism is performed for these two learning modules so that the aspects and sentences can impact the resulting opinion expression. In addition, we introduce a gradient reversal layer to obtain invariance features. The comprehensive experiments are performed on Amazon multidomain product datasets and show the effectiveness and significance of the proposed model over state‐of‐the‐art techniques.

A sparse channel state information (CSI) estimation model is proposed for reducing the pilot overhead of orthogonal time frequency space (OTFS) modulation aided multiple-input multiple-output (MIMO) systems. Explicitly, the pilots are directly transmitted over the time-frequency (TF)-domain grid for estimating the delay-Doppler (DD)-domain CSI that leads to a reduction of the pilot overhead, training duration and pre-processing complexity. Furthermore, it completely avoids placing multiple DD-domain guard intervals corresponding to each transmit antenna within the same OTFS frame, while keeping the training duration flexible, hence increasing the bandwidth efficiency. A unique benefit of the proposed CSI estimation model is that it can efficiently handle fractional Dopplers also. The resultant DD-domain CSI becomes simultaneously row and group (RG)-sparse. To exploit this compelling property, an orthogonal matching pursuit (OMP)-based RG-OMP technique is developed, conveniently complemented by an enhanced Bayesian learning (BL)-based RG-BL framework, both of which substantially outperform the state-of-the-art methods. Furthermore, low-complexity linear detectors are designed for the ensuing data detection phase, which directly employ the estimated DD-domain sparse CSI, without assuming any further knowledge concerning the number of dominant multipath components. Finally, simulation results are provided to demonstrate performance improvement of the proposed BL-based schemes over the OMP and the state-of-the-art schemes.

Orthogonal time frequency space (OTFS) waveform based millimeter wave (mmWave) MIMO systems are capable of achieving high data rates in high-mobility scenarios. Hence, transceivers are designed for both analog beamforming (AB) and hybrid beamforming (HB), where we commence by deriving the delay-Doppler (DD)-domain input-output relationship considering a delay-Doppler-angular domain channel model. Subsequently, a novel two-stage procedure is developed for transmit beamformer (TBF)/ precoder (TPC) and receiver combiner (RC) design, and for estimating the DD-domain’s equivalent channel state information (CSI). The key feature of the proposed framework is that the RF TBF/ TPC and RC design maximizes the directional beamforming gains. It is also demonstrated that the low-dimensional baseband CSI of the DD-domain becomes sparse for mmWave-AB MIMO OTFS systems, and block-sparse for mmWave-HB MIMO OTFS systems. Subsequently, Bayesian learning (BL) and block-sparse BL (BS-BL) solutions are developed for improved CSI estimation. We also derive the Bayesian Cramer-Rao lower bounds (BCRLB) for benchmarking the mean-squared-error (MSE) of the CSI estimates. Finally, our simulation results demonstrate the improved efficacy of the proposed transceiver designs and confirm the enhanced CSI estimation performance of the BL-based schemes over other competing sparse signal recovery schemes.

In this paper, a system for oral cancer screening using Laser Induced Fluorescence(LIF) has been developed. A hybrid approach of classification using Support Vector Machine (SVM) and Random Forest (RF) classifier's is proposed. Performance of the classifier is evaluated using several features types such as Wavelet, DFT, LDFT, ILDFT, DCT, LDCT and Slopes features. The most discriminating features are selected using Recursive Feature Elimination(RFE). Analysis of the problem of subset selection from SVM-RFE ranked list is also performed. The hybrid approach has been compared with stand-alone SVM, SVM-RFE and RF classifiers. The proposed technique improves the performance of the classification system significantly. The novelty of the approach lies in the way the most significant features are exstracted in separate modules to arrive at a decision and how the decision are then fused in an intelligent fashion to arrive at a final classification.