Generative AI has rapidly progressed from chatbots and assistants to agents across diverse applications and domains. Generative AI agents demonstrate sophisticated operation through autonomy, tool use and decision making with minimal human input. Despite these performance gains, agents are still impacted by the foundational limitations of Generative AI models. Among these, hallucinations are a major limitation that affects agent operation in real-world settings, leading to risk and loss. Several recent work aim to address hallucinations through methods such as retrieval-augmented generation and reflection prompting, however, these only provide partial improvements. An effective yet underexplored approach is in the observability data generated by an agent in its deployed and operational settings. Drawing on agent observability data, this paper proposes a dual memory knowledge graph approach that integrates Semantic and Observability Memory to address hallucinations in Generative AI agents. Semantic Memory provides organized domain knowledge for precise factual grounding. Observability Memory transforms logs, traces, and execution results into agent validated planning histories. Hallucinations are then addressed by grounded planning in verified past interactions with known, reliable outcomes. This approach is evaluated in a two-stage experimental setup aligned with its dual memory design. Observability memory is evaluated on the HotpotQA dataset to assess its impact on reasoning grounding, using metrics that capture both factual accuracy and reasoning hallucinations. The SM3-Text-to-Query benchmark and Synthea-based medical QA datasets are used to assess factual grounding of the semantic memory. Results from both experiments demonstrate reductions in hallucinations, with semantic memory for contextual grounding reducing factual hallucinations, and observability memory for reasoning grounding reducing faithfulness hallucinations.

The escalating prevalence of work-related stress has led to a notable decline in work performance and the mental well-being of the workforce. Studies suggest that personally tailored music can have a positive impact on reducing stress levels. This paper proposes a novel approach to generate personalized music by combining brain activation analysis and musical preference analysis. The proposed method involves three main stages: (1) Source Separation, (2) Brain Activation Analysis, and (3) Personalized Music Generation. In the first stage, we use a variant of the Wave-U-Net to decompose the input song into its vocals and melodies. Subsequently, we employ electroencephalography (EEG) to analyze the user’s brain activation and identify preferred segments. The third stage involves the generation of personalized music using a transformer-based music generation model, which takes into account the user’s music preference and brain activation patterns. Furthermore, this approach was assessed through an experiment with 10 participants, achieving an average satisfaction level of 3.9/5 and a 52.97% increase in Frontal Alpha Asymmetry (FAA) fluctuation percentage after analyzing brain activation with the generated melody. As future work, we intend to conduct intervention-based experiments with large sample sizes and expand into evaluation metrics that provide objective measures of work stress reduction.

Machine perception of emotions is integral to the development of human-centric Artificial Intelligence (AI) in sustainable industrial applications. Human expressions of emotions are not always direct. Word embeddings are mature techniques that can extract the semantics of such indirect expressions from text data. However, they are not primed to extract emotions. In this paper, we propose a novel approach that generates robust word embeddings for implicit and explicit expressions of emotion. This approach consists of two techniques, mask and rogue, we evaluate both techniques on two benchmark datasets for emotion classification. Our results confirm the effectiveness of the proposed approach in extracting emotions from diverse contexts. We have shared the emotion word embedding for public use.

Contemporary energy platforms are leveraging advanced data management and Artificial Intelligence (AI) capabilities in response to the increasing complexity of energy systems and grids. Despite these advances, it is a non-trivial and challenging task to support the decision-making needs of the human operators of such complex energy-related implementations. Conversational agents or chatbots are a potential emerging technology that can be utilized to address this challenge. Although there is a large body of literature on chatbots in general, they are not robust as they rely on predefined conversational pathways that are inadequate to efficiently address the complexities of dynamic data spaces in energy platforms. The capability of generating answers in real-time by communicating with the dynamic dataspace is crucial as energy management decisions are real-time and time sensitive. In this paper, we present the design and development of Cooee, a chatbot for conversational engagement with the dynamic data spaces of complex energy environments. Cooee leverages state-of-art language models along with rule-based language processing methods for a conversational interaction with dynamic data spaces, which consequently supports and enables decision-making by human experts. We have developed Cooee as a standalone application and then integrated into a real-world energy AI platform deployed within a multi-campus tertiary education institution setting. Cooee was empirically evaluated in this setting and compared with several state-of-the-art Q&A approaches.

Idioms are figures of speech that contradict the principle of compositionality. This disposition of idioms can misdirect Natural Language Processing (NLP) techniques, which mostly focus on the literal meaning of terms. In this paper, we propose a novel idiom detection model that distinguishes between literal and idiomatic expressions. It utilizes a token classification approach to fine-tune BERT(Bidirectional Encoder Representations from Transformers). It is empirically evaluated on four idiom datasets, yielding an accuracy of more than 0.94. This model adds to the robustness and diversity of NLP techniques available to process and understand increasing magnitudes of free-form text and speech. Furthermore, the social value of this model is in enabling non-native speakers to comprehend the nuances of a foreign language.

Safe and effective vaccination is leading the recovery from the COVID-19 pandemic. Despite the urgency of full vaccination that prevents serious illness, a state of vaccine messaging augmented by disinformation campaigns in online social networks has emerged. Several studies have established a link between social media activity and vaccine messaging, and most platforms are actively removing vaccine disinformation. The objective of this study is to apply a validated Artificial Intelligence (AI) framework to extract, analyze and synthesize themes, emotions and emotion transitions associated with COVID-19 vaccine messaging in online social networks. We applied the framework on approximately 400,000 COVID-19 vaccine-related posts and conversations on two social media platforms, Twitter and Reddit, from March 2020 to September 2021. The results of this study are threefold, firstly, the discovery of a minority of implied anti-vaccine themes on infertility, microchips, gene editing and fetal cells that have remained undetected. Secondly, the discovery of six themes that capture a majority of the vaccine messaging namely, social lockdown measures, frontline healthcare providers, side effects, vaccine distribution, breakthrough infections and vaccine efficacy on variants. Thirdly, the variety and intensity of emotions expressed since the start of the pandemic, and comparatively negative emotions being expressed in recent months. We anticipate the findings of our study will contribute towards improved vaccine messaging as the world returns to a new normal from the COVID-19 pandemic.

Recent advancements in Artificial intelligence (AI) have led to its widespread adoption in healthcare applications and services. The global pandemic has further hastened the integration of AI into telehealth services, such as service quality improvement and new models of care. This paper focuses on the service quality improvement of telehealth, specifically, cancer information and telephone support services, in terms of its human system interaction. We present an AI framework for the detection of patient emotions and emotion transitions based on call recordings of telehealth services. The call recordings are typically a conversation between the caller (patient or carer) and the healthcare practitioner (nurse or counsellor). All primary emotions are expressed across diverse topics during these calls. It is anticipated that the caller emotion state improves during the call due to the information and support received. The proposed AI framework is designed to detect emotions expressed at all stages of the call and based on these expressions, formulate the emotion transition during the call. We have evaluated the proposed AI framework on a large real-world dataset of 60,000 call recordings of cancer information and telephone support services provided by Cancer Council Victoria, Australia. The results confirm the effectiveness of this AI framework in detection of emotions and emotion transitions during the provision of telehealth services.

We introduce an open dataset of high-granularity Photovoltaic (PV) solar energy generation, solar irradiance, and weather data from 42 PV sites deployed across five campuses at La Trobe University, Victoria, Australia. The dataset includes approximately two years of PV solar energy generation data collected at 15-minute intervals. Geographical placement and engineering specifications for each of the sites are also provided to aid researchers in modelling solar energy generation. Weather data is available at 1-minute intervals and is provided by the Australian Bureau of Meteorology (BOM). Apparent temperature, air temperature, dew point temperature, relative humidity, wind speed, and wind direction were provided under the weather data. The paper describes the data collection methods, cleaning, and merging with weather data. This dataset can be used to forecast, benchmark, and enhance operational outcomes in solar sites.