Resilient Conversational AI: A Framework for Connectivity-Adaptive Inference with Graceful Degradation

Conversational AI is one of the most active research areas in AI, and it has gained more attention from academia as well as industry. Given recent advancements in several conversational AI systems in addition to the availability of several datasets, the aim of this study is to explore the landscape of Arabic text-based conversational AI systems. In this work, we provide a thorough review of recent Arabic conversational AI systems. We group them into three categories based on their functionality: (1) question-answering (QA) systems, (2) task-oriented dialogue systems (DS), and (3) chatbots. Furthermore, we describe the common datasets used in building and evaluating conversational AI systems in Arabic. Few surveys have targeted the conversational AI field for the Arabic language, and we aim to cover this gap with this study. Our contribution focuses on reviewing and analyzing the literature in the field and highlighting future research directions towards human-like conversational AI systems in Arabic.

Resilience is an increasingly popular concept in research and practice, but quantitative resilience analyses are often disconnected from resilience theory. For example, previous studies argue that diversity, a key attribute for building resilience, and agency, essential for understanding local adaptation and transformation, are critical to understanding resilience. Despite significant progress in integrating them into qualitative frameworks, diversity and agency are rarely incorporated into quantitative social-ecological resilience metrics. This omission is concerning, given the critical role of quantitative resilience metrics in informing resilience-oriented decision-making. This study examines how diversity and agency are represented in quantitative resilience metrics across disciplines, with the goals of (a) assessing how research on social-ecological resilience currently integrates these concepts into quantitative metrics and (b) identifying future opportunities to enhance their inclusion using insights from other fields. Using topic modelling to identify different research fields and facilitate the screening process, we performed a multidisciplinary systematic meta-review of resilience metrics. To understand what types of resilience metrics are used across disciplines and where diversity and agency are more commonly included, we identified six categories of resilience metrics, with “performance under disruption” being the most used category (35%). We found that a limited number of quantitative resilience metrics include diversity and agency, with “system structure” and “compound indicators” being the main sources of diversity and agency, respectively. We further reviewed simulation models applying resilience metrics. The prevalence of performance under disruption metrics is stronger than in reviews (67%) and a similar quantity of metrics including diversity (14%) and agency (5%) was found. Drawing on insights from multiple disciplines, we outline five potential pathways to improve the inclusion of diversity and agency in social-ecological resilience metrics: using network-based metrics, using response and pathway diversity, including diversity and agency in compound indicators, integrating quantitative methodologies outside resilience theory, and improving the application of resilience in simulation models.

Conversational AI has revolutionized the way humans interact with machines, with applications spanning customer service, virtual assistants, and healthcare. This paper explores the advancements in conversational AI systems, focusing on the role of deep learning models such as Transformers, BERT, and GPT-3 in improving language understanding and response generation. The study outlines how these models enable AI systems to generate contextually relevant, coherent, and human-like responses in various conversation settings. Additionally, the paper delves into the architecture of neural networks used in Conversational AI, highlighting the progression from traditional rule-based systems to more sophisticated deep learning frameworks. The paper further discusses the challenges faced in conversational AI, such as natural language ambiguity, context retention, and ethical considerations surrounding bias in language models. Moreover, the integration of conversational AI into business processes, healthcare, and customer support is analyzed, showcasing real-world case studies where AI-driven chatbots have improved operational efficiency. The paper also explores the future of conversational AI, including multimodal systems that combine text, voice, and visual inputs for more dynamic interactions. Lastly, it considers the ethical implications of conversational AI, particularly in terms of privacy concerns and data security, offering recommendations for creating more transparent and accountable AI systems.

The integration of Virtual Reality (VR)and Conversational AI in education has modernized how students engage with educational and career guidance. Despite the growing integration of VR and Conversational AI in education, empirical studies on their combined impact on students’ intentions to pursue computer science remain limited. This study proposes the Conversational AI-empowered Virtual Reality System (CAI-VR system)as an educational guidance tool designed to enhance high school students' intentions to pursue a Bachelor of Science in Computer Science at RMUTT. We employed a developmental research method to develop and evaluate the system. The findings indicate that 1) the CAI-VR system demonstrates the convergence of VR, conversational AI, and multimedia elements in delivering an immersive 360-degree virtual experience. Users can access the system in two modes, i.e., Immersive VR Mode for full interactivity via VR devices and Web-Based Display Mode for browser access; 2) Usability was assessed with 30 high school students, while five experts evaluated system quality. Both assessments confirmed the system’s effectiveness; 3) the system significantly increased students’ intention scores, reinforcing its role in guiding educational decision-making; 4) further evaluation with 400 students demonstrated high awareness; however, behavioral commitment—especially in planning, preparation, and application—was lower than cognitive or affective aspects. Overall, students reported a high level of satisfaction with the system (M = 3.99, SD = 0.88). These findings suggest that the CAI-VR system can effectively support digital-native students’ educational and career decision-making. Further development should deepen learner engagement via gamified, collaborative, and personalized experiences.

Recent research has utilized resumes and occupational profiles to identify and quantify discrimination within conversational AI systems. We extend this concept by proposing a novel approach to quantification which leverages Holland’s theory of occupational interests to provide a more holistic interpretation of discrimination in these systems. Building upon the seminal work of Bertrand and Mullainathan (Am Econ Rev 94(4):991–1013, 2004), which examined racial discrimination in hiring through a naturalistic experiment, we adapt their methodology to explore biases embedded in AI-driven job recommendations. By manipulating gender and race within user profiles, as well as investigating if these characteristics are inferred by name alone, we assess the extent to which conversational AI models (GPT-3.5, GPT-4, Llama 2, and Gemini) exhibit discriminatory patterns. Our findings indicate significant disparities in job recommendations based on demographic factors. These findings vary between AI models. We extend previous research in this domain which explores bias with singular metrics by transforming recommendations via Holland’s theory to a multidimensional set of characteristics allowing for a more thorough interpretation of bias. Our work underscores the importance of continuous auditing and refinement of AI systems to mitigate bias and promote fairness in automated decision-making processes.

Query rewrite (QR) is an emerging component in conversational AI systems, reducing user defect. User defect is caused by various reasons, such as errors in the spoken dialogue system, users’ slips of the tongue or their abridged language. Many of the user defects stem from personalized factors, such as user’s speech pattern, dialect, or preferences. In this work, we propose a personalized search-based QR framework, which focuses on automatic reduction of user defect. We build a personalized index for each user, which encompasses diverse affinity layers to reflect personal preferences for each user in the conversational AI. Our personalized QR system contains retrieval and ranking layers. Supported by user feedback based learning, training our models does not require hand-annotated data. Experiments on personalized test set showed that our personalized QR system is able to correct systematic and user errors by utilizing phonetic and semantic inputs.

This paper presents a novel approach to end-of-turn detection for conversational AI systems, combining traditional voice activity detection (VAD) with semantic understanding through a transformer-based model. End-of-turn detection remains one of the most challenging aspects of creating natural conversational AI interfaces, as current systems rely primarily on silence thresholds that fail to capture the semantic cues humans use to determine speaking turns. Our approach leverages a lightweight transformer model based on the Gemma-3-1b architecture that analyzes transcribed speech in real-time to predict when a user has completed their speaking turn. The model utilizes a sliding context window of recent conversation turns to make these predictions. Experimental results demonstrate that our hybrid system reduces unintentional interruptions by 85% compared to VAD-only approaches, with only a 3% rate of false negatives where turns are incorrectly determined to be incomplete. The proposed system operates effectively on consumer hardware with minimal latency (~50ms) and demonstrates robust performance across various conversational contexts including complex information gathering scenarios like address collection and customer service interactions. Keywords: Conversational AI, End-of-turn detection, Transformer models, Natural language processing, Voice activity detection, Turn-taking, Human-computer interaction

The integration of conversational artificial intelligence into enterprise data analytics and governance represents a paradigm shift in how organizations interact with their data assets. This paper presents a comprehensive framework for implementing conversational AI systems that enable natural language querying, automated compliance monitoring, and intelligent data discovery in enterprise environments. The proposed architecture leverages advanced natural language processing techniques, including large language models and context-aware dialogue systems, to bridge the gap between business users and complex data infrastructures. Through empirical evaluation across three enterprise domains— financial services, healthcare, and telecommunications—the framework demonstrates significant improvements in query response time (67% reduction), user adoption rates (89% increase), and data governance compliance (78% improvement). The system addresses critical challenges including query ambiguity resolution, multi-modal data integration, and real-time governance policy enforcement. Results indicate that conversational AI can effectively democratize data access while maintaining stringent security and compliance requirements, positioning it as a transformative technology for enterprise data management.

Conversational AI systems have emerged as key enablers of human-like interactions across diverse sectors. Nevertheless, the balance between linguistic nuance and factual accuracy has proven elusive. In this paper, we first introduce LLMXplorer, a comprehensive tool that provides an in-depth review of over 205 large language models (LLMs), elucidating their practical implications, ranging from social and ethical to regulatory, as well as their applicability across industries. Building on this foundation, we propose a novel functional architecture that seamlessly integrates the structured dynamics of knowledge graphs with the linguistic capabilities of LLMs. Validated using real-world AI news data, our architecture adeptly blends linguistic sophistication with factual rigor and further strengthens data security through role-based access control. This research provides insights into the evolving landscape of conversational AI, emphasizing the imperative for systems that are efficient, transparent, and trustworthy.

Conversational AI has become popular in modern society. This paper discusses one such conversational framework, Continuous Question Answering (CQA), which is designed to support a goal-oriented process. In this paper, an end-to-end intelligent dialogue framework is presented to automate any process in which there is a need to converse by asking questions and evaluating the answers between two entities. The paper discusses the techniques and strategies used to tackle challenges in the areas of Natural Language Understanding, Contextual Modeling, Dialogue Planning, Question Generation, and Information Retrieval. To demonstrate the efficiency of the CQA framework, an end-to-end recruitment system is developed to interview candidates and grade them accordingly. The system understands the user’s prior knowledge and intelligently sets up an open-ended interview process. The goal of the system is to assess the user’s technical proficiency through a Question-and-Answer framework and recommend the best candidates to the employer for a specific job role. A grading system is used to evaluate each candidate and provide an in-depth analysis of their performance. The system received a human satisfaction score of 4.3 out of 5 for the recruitment system implemented using the CQA framework, which outperforms other state-of-the-art conversational systems.

In this position paper, we present five key principles, namely interpretability, inherent capability to explain, independent data, interactive learning, and inquisitiveness, for the development of conversational AI that, unlike the currently popular black box approaches, is transparent and accountable. At present, there is a growing concern with the use of black box statistical language models: While displaying impressive average performance, such systems are also prone to occasional spectacular failures, for which there is no clear remedy. In an effort to initiate a discussion on possible alternatives, we outline and exemplify how our five principles enable the development of conversational AI systems that are transparent and thus safer for use. We also present some of the challenges inherent in the implementation of those principles.

It is becoming feasible to stream user generated video directly to the Internet, which enables a new class of exciting applications and services. However, energy consumption of smartphones that transmit the content over mobile network is a limiting factor. We study the potential of frame bundling to save energy while streaming video through a LTE uplink. Our simulation results show that if real-time streaming is not strictly necessary, bundling video frames for a few seconds before transmitting them is an effective mechanism to save energy. We analyze the energy savings achievable for different video bit rates with varying network conditions, namely round-trip time (RTT), packet loss, UE mobility, and background traffic. The results confirm that bundling remains effective also in non-ideal network conditions.

A fundamental problem that needs to be addressed when designing survivable optical transport networks is the capability of these networks to quickly recover from element failures. Current restoration signaling is a two-phase messaging procedure between the sender and the receiver and depends heavily on the control message propagation delays during the recovery process and the optical cross-connect switching times. Offset time based restoration has been proposed to address [6] the impact of these network parameters and upper bound expressions have been derived. However, a closer investigation showed that as the network conditions change (e.g., increasing the number of wavelength channels per link) the benefits offered by the proposed framework would rapidly be depleted, and thereby rendering the argument of avoiding conventional restoration signaling inept. In this paper, we intend to re-use the same framework, however instead of using upper bound expressions to estimate the restoration times, we propose a more accurate model to estimate the offset time of each failed connection using a timedriven scheduling procedure. We show that propagation delays have very minimal impact on the network recovery times and switching delays are minimized whereas their accumulation along restoration routes is eliminated. We evaluate our proposal through simulation experiments and we show that by deploying a scheduling process, substantial restoration gain can be achieved under varying network conditions.

The use of traditional TCP, in its present form, for reliable transport over Ad hoc Wireless Networks (AWNs) leads to a significant degradation in the network performance. This is primarily due to the congestion window (cwnd) updation and congestion control mechanisms employed by TCP and its inability to distinguish congestion losses from wireless losses. In order to provide an efficient reliable transport over AWNs, we propose Learning-TCP, a novel learning automata based reliable transport protocol, which efficiently adjusts the cwnd size and thus reduces the packet losses. The key idea behind Learning-TCP is that, it dynamically adapts to the changing network conditions and appropriately updates the cwnd size by observing the arrival of acknowledgment (ACK) and duplicate ACK (DUPACK) packets. Learning-TCP, unlike other existing proposals for reliable transport over AWNs, does not require any explicit feedback, such as congestion and link failure notifications, from the network. We provide extensive simulation studies of Learning-TCP under varying network conditions, that show increased throughput (9-18%) and reduced packet loss (42-55%) compared to that of TCP.

In the future, applications will need to execute in a ubiquitous environment with varying network conditions (connectivity, bandwidth, etc.) and system constraints (e.g., power and storage). The distributed object paradigm is often used to facilitate the development of large-scale distributed applications. However, the traditional object messaging layer operates with limited awareness of underlying system and network conditions, whereas current system and network monitoring tools operate at the network layer with little awareness of application-level object communication requirements. This article explores the possibility, mechanisms and benefits of filling the gap between object messaging and system monitoring. We introduce connection abstraction as the mechanism for these two layers to communicate and exchange information. Through this integration, object messaging can proactively adapt to changing system conditions; system monitoring policies and parameters can be optimized based on interobject communication properties.