Soumyabrata Chaudhuri, Pranav Purkar, Ritwik Raghav, Shubhojit Mallick, Manish Gupta, Abhik Jana, Shreya Ghosh. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

In this paper, we investigate Extractive Question Answering (EQA) with Large Language Models (LLMs) under domain drift, i.e., can LLMs generalize to domains that require specific knowledge such as medicine and law in a zero-shot fashion without additional in-domain training? To this end, we devise a series of experiments to explain the performance gap empirically. Our findings suggest that: (a) LLMs struggle with dataset demands of closed do- mains such as retrieving long answer spans; (b) Certain LLMs, despite showing strong overall performance, display weaknesses in meeting basic requirements as discriminating between domain-specific senses of words which we link to pre-processing decisions; (c) Scaling model parameters is not always effective for cross-domain generalization; and (d) Closed-domain datasets are quantitatively much different than open-domain EQA datasets and current LLMs struggle to deal with them. Our findings point out important directions for improving existing LLMs.

We study extractive question-answering in the medical domain (Medical-EQA). This problem has two main challenges: (i) domain specificity, as most AI models lack necessary domain knowledge, and (ii) extraction-based answering style, which restricts most autoregressive LLMs due to potential hallucinations. To handle those challenges, we propose TOP-Training, a target-oriented pre-training paradigm that stands out among all domain adaptation techniques with two desirable features: (i) TOP-Training moves one step further than popular domain-oriented fine-tuning since it not only moves closer to the target domain, but also familiarizes itself with the target dataset, and (ii) it does not assume the existence of a large set of unlabeled instances from the target domain. Specifically, for a target Medical-EQA dataset, we extract its entities and leverage large language models (LLMs) to generate synthetic texts containing those entities; we then demonstrate that pretraining on this synthetic text data yields better performance on the target Medical-EQA benchmarks. Overall, our contributions are threefold: (i) TOP-Training, a new pretraining technique to effectively adapt LLMs to better solve a target problem, (ii) TOP-Training has a wide application scope because it does not require the target problem to have a large set of unlabeled data, and (iii) our experiments highlight the limitations of autoregressive LLMs, emphasizing TOP-Training as a means to unlock the true potential of bidirectional LLMs.

As social networks become the primary sources of information, the rise of misinformation poses a significant threat to the information ecosystem. Here, we address this challenge by proposing a dynamic system for real-time evaluation and assignment of misinformation scores to tweets, which can support the ongoing efforts to counteract the impact of misinformation public health, public opinion, and society. We use a unique combination of Temporal Graph Network (TGN) and Recurrent Neural Networks (RNNs) to capture both structural and temporal characteristics of misinformation propagation. We further use active learning to refine the understanding of misinformation, and a dual model system to ensure the accurate grading of tweets. Our system also incorporates a temporal embargo strategy based on belief scores, allowing for comprehensive assessment of information over time. We further outline a retraining strategy to keep the model current and robust in the dynamic misinformation landscape. The evaluation results across five social media misinformation datasets show promising accuracy in identifying false information and reducing propagation by a significant margin.

This paper introduces MoveInsight, a novel framework, leveraging a Mobility Knowledge Graph and deep learning architecture to analyze individuals' GPS traces from sensor-equipped smartphones for extracting trip purposes and understanding spatio-temporal mobility patterns. Unlike traditional information retrieval methods, MoveInsight deciphers the motivations behind travels by examining relations among individuals' movement behaviors, locations, and semantic contexts. The framework employs a multi-task learning approach for annotating trajectories and a transfer learning method for extending analysis to different regions, utilizing insights from comparable areas. Through real-world dataset testing, MoveInsight outperformed baseline methods in trip-purpose extraction and Point-of-Interest annotations by around 18% to 30%, showcasing its promise in enhancing location-centric services by providing deeper insights into human mobility dynamics.

The proliferation of sensor-equipped smartphones has led to the generation of vast amounts of GPS data, such as timestamped location points, enabling a range of location-based services. However, deciphering the spatio-temporal dynamics of mobility to understand the underlying motivations behind travel patterns presents a significant challenge. This paper focuses on how individuals’ GPS traces (latitude, longitude, timestamp) interpret the connection and correlations among different entities such as people, locations or point-of-interests (POIs), and semantic contexts (trip-purpose). We introduce a mobility analytics framework, namedMobilyticsdesigned to identify trip purposes from individual GPS traces by leveraging a “mobility knowledge graph” (MKG) and a deep learning architecture that automatically annotates the GPS log. Additionally, we propose a novel “transfer learning” approach to explore movement dynamics in a geographically distant area by leveraging knowledge obtained from a comparable region, such as an academic campus. In terms of major contributions and novelty, this is the first work to present end-to-end daily mobility trip purpose extraction and mobility knowledge transfer for trip annotation and POI-tagging where the labeled data are insufficient. Experimental results on real-life datasets of five different regions demonstrate the efficacy of our proposed Mobilytics framework which outperforms the baselines for trip-purpose extraction and POI annotations by a significant margin ($\approx$18% to$\approx$30%). Moreover, the analysis on huge volume of simulated traces (10,000 users) illustrates the scalability and robustness of the framework.

The buzz around Transformer-based language models (TLM) such as BERT, RoBERTa, etc. is well-founded owing to their impressive results on an array of tasks. However, when applied to areas needing specialized knowledge (closed-domain), such as medical, finance, etc. their performance takes drastic hits, sometimes more than their older recurrent/convolutional counterparts. In this paper, we explore zero-shot capabilities of large LMs for extractive QA. Our objective is to examine performance change in the face of domain drift i.e. when the target domain data is vastly different in semantic and statistical properties from the source domain and attempt to explain the subsequent behavior. To this end, we present two studies in this paper while planning further experiments later down the road. Our findings indicate flaws in the current generation of TLM limiting their performance on closed-domain tasks.

In this paper, we develop an end-to-end knowledge extraction and management framework for COVID-19 vaccination misinformation. This framework automatically extracts information consistent and inconsistent with scientific evidence regarding vaccination. Additionally, using novel natural language processing methods (including triple-attention based sarcasm detection and utilizing topic-based similarity scoring, agglomerative clustering, and word embedding vectors for misinformation category identification and counter-fact summarization in a semi-supervised way from web-based sources), we explore public opinion towards vaccination resistance. Our knowledge extraction pipeline constructs knowledge-bases automatically, categorizes vaccine dissenting tweets into 15 misinformation categories automatically, and effectively analyzes discourses in those tweets. Our contributions are as follows: (i) the proposed knowledge extraction framework does not require huge amounts of labelled tweets of different categories (our method uses only 50-labelled tweets for each of 15 misinformation categories, in stark contrast to existing approaches that typically rely on 10,000 or more labelled tweets), and (ii) our module outperformed baselines by a significant margin of ≈ 8% to ≈ 14% (F1 score) in the classification tasks using Twitter dataset.

Our primary objective is to explore and enhance AI's role for wildlife conservation, in brief, Nature Through the Lens of AI. It seeks to address crucial challenges related to data heterogeneity, scale integration, data privacy, mitigating biases, and decision-making under uncertainty. This workshop is centred around leveraging AI's prowess in deciphering complex spatio-temporal data patterns for wildlife conservation, thereby contributing significantly to the broader canvas of AI for social good. The workshop intends to create an interdisciplinary platform bringing together computer scientists, data scientists, geospatial experts, ecologists, and conservation practitioners, fostering collaboration and driving real-world impact. The program will include keynote speeches, panel discussions, and interactive sessions focusing on efficient knowledge extraction and management, remote sensing technologies, predictive modeling, species distribution modeling, habitat quality assessment, and human-wildlife conflict mitigation. With an em- phasis on CIKM's primary interests, our aim is not only to enrich understanding of AI's symbiotic potential with ecology but also to utilize it to address pressing societal and environmental challenges.

The interaction between elephants and their environment has profound implications for both ecology and conservation strategies. This study presents an analytical approach to decipher the intricate patterns of elephant movement in Sub-Saharan Africa, concentrating on key ecological drivers such as seasonal variations and rainfall patterns. Despite the complexities surrounding these influential factors, our analysis provides a holistic view of elephant migratory behavior in the context of the dynamic African landscape. Our comprehensive approach enables us to predict the potential impact of these ecological determinants on elephant migration, a critical step in establishing informed conservation strategies. This projection is particularly crucial given the impacts of global climate change on seasonal and rainfall patterns, which could substantially influence elephant movements in the future. The findings of our work aim to not only advance the understanding of movement ecology but also foster a sustainable coexistence of humans and elephants in Sub-Saharan Africa. By predicting potential elephant routes, our work can inform strategies to minimize human-elephant conflict, effectively manage land use, and enhance anti-poaching efforts. This research underscores the importance of integrating movement ecology and climatic variables for effective wildlife management and conservation planning. Codebase available at https://github.com/shreyaghosh-2016/Elephant-movement

Exploring human mobility dynamics is a challenging semantic data analysis task as conventional information retrieval techniques fail to detect "why people travel". We propose a mobility analytics framework to discover such trip-purposes from individual’s GPS traces using a mobility knowledge graph (MKG) and a deep-learning architecture that automatically annotates the GPS log. Further, a novel transfer learning technique is proposed to analyze the movement dynamics in another geographically dispersed region with the help of the knowledge gained from a region of similar type (say, academic campus).

We present a demonstration of nighttime light pattern (NTL) analysis system. Our tool named NightVIEW is powered by an efficient system architecture to easily export and analyse a huge volume of spatial data (NTL), image segmentation and clustering algorithms to find unusual NTL patterns and identify hotspots of excess night light usage as well as finding semantics of cities.

Recent advancements in artificial intelligence (AI) and IoT technology have revolutionized the healthcare industry by providing effective remote healthcare. Furthermore, with the aging of the world’s population, remote health monitoring and recommendations are becoming imperative to provide cost-effective healthcare solutions for improving the quality of life of our senior citizens. The explosive growth of wearable sensors (IoT sensors) and health bands has facilitated the interconnection among patients and caregivers to enable assisted living by leveraging AI techniques. This work proposes an end-to-end connected smart home healthcare system (FEEL) for elderly people. Our proposed framework addresses the main challenges of the Internet of Medical Things (IoMT) system namely, the scarcity of labeled data and user’s diverse needs. The major contributions of the work are: 1) few-shot learning-enabled novel federated learning (FL) framework for health data and context information analysis and recommendation; 2) user and context-based knowledge graph (UKG) to represent and model health parameters and environmental impacts on recommendations; 3) deep learning architecture for activity monitoring and location estimation of the users; and 4) edge-fog-IoMT collaborative framework to collect, store, and share medical recommendations while protecting the privacy of the users. FEEL is specifically beneficial for elderly homes where several aged people stay together and require constant care. We aim to develop a novel AI module where along with the health parameters, the social context of the home can be augmented to provide an accurate and improved healthcare service. FEEL has been evaluated for three tasks, namely: 1) activity monitoring and location estimation; 2) fall detection; and 3) medical recommendations for unusual health conditions. A customized wearable device has been used to collect, store, and send health-related parameters. The experimental evaluation demonstrates promising accuracy (F1 score 0.86–0.94 range) for the tasks and outperforms the baselines by a significant margin ($\approx 10$%–16%).

Abstract Internet of Things (IoT) has a pivotal role in developing intelligent and computational solutions to facilitate varied real‐life applications. To execute high‐end computations and data analytics, IoT and cloud‐based solutions play the most significant role. However, frequent communication with long distant cloud servers is not a delay‐aware and energy‐efficient solution while providing time‐critical applications such as healthcare. This article explores the possibilities and opportunities of integrating cloud technology with fog and edge‐based computing to provide healthcare services to users in exigency. Here, we propose an end‐to‐end framework namedRESCUE(enabling green healthcare services using integrated iot‐edge‐fog‐cloud computing environments), consisting efficient spatio‐temporal data analytics module for efficient information sharing, spatio‐temporal data analysis to predict the path for users to reach the destination (healthcare center or relief camps) with minimum delay in the time of exigency (say, natural disaster). This module analyzes the collected information through crowd‐sourcing and assists the user by extracting optimal path postdisaster when many regions are nonreachable. Our work is different from the existing literature in varied aspects: it analyses the context and semantics by augmenting real‐time volunteered geographical information (VGI) and refines it. Furthermore, the novel path prediction module incorporates such VGI instances and predicts routes in emergencies avoiding all possible risks. Also, the design of development of a latency‐aware, power‐aware data‐driven analytics system helps to resolve any spatio‐temporal query more efficiently compared to the existing works for any time‐critical application. The experimental and simulation results outperform the baselines in terms of accuracy, delay, and power consumption.

Abstract Several global health incidents and evidences show the increasing likelihood of pandemics (large‐scale outbreaks of infectious disease), which has adversely affected all aspects of human lives. It is essential to develop an analytics framework by extracting and incorporating the knowledge of heterogeneous data‐sources to deliver insights for enhancing preparedness to combat the pandemic. Specifically, human mobility, travel history, and other transport statistics have significantly impact on the spread of any infectious disease. This article proposes a spatio‐temporal knowledge mining framework, named STOPPAGE, to model the impact of human mobility and other contextual information over the large geographic areas in different temporal scales. The framework has two key modules: (i) spatio‐temporal data and computing infrastructure using fog/edge based architecture; and (ii) spatio‐temporal data analytics module to efficiently extract knowledge from heterogeneous data sources. We created a pandemic‐knowledge graph to discover correlations among mobility information and disease spread, a deep learning architecture to predict the next hotspot zones. Further, we provide necessary support in home‐health monitoring utilizing Femtolet and fog/edge based solutions. The experimental evaluations on real‐life datasets related to COVID‐19 in India illustrate the efficacy of the proposed methods. STOPPAGE outperforms the existing works and baseline methods in terms of accuracy by (18–21)% in predicting hotspots and reduces the power consumption of the smartphone significantly. The scalability study yields that the STOPPAGE framework is flexible enough to analyze a huge amount of spatio‐temporal datasets and reduces the delay in predicting health status compared to the existing studies.

With the enormous growth of wireless technology, and location acquisition techniques, a huge amount of spatio-temporal traces are being accumulated. This dataset facilitates varied location-aware services and helps to take real-life decisions. Efficiently handling and processing spatio-temporal queries are necessary to respond in real-time. Processing the vast spatio-temporal data requires scalable computing infrastructure. In this regard, an efficient query resolution system can be deployed if we predict the infrastructure requirement of the user query apriori along with the identification of the geospatial service chain. In this work, we propose a framework, namelyLYRIC(deadLine and budget aware spatio-temporal querYpRocessingInCloud), where the spatio-temporal queries are resolved efficiently considering user-defined deadline and budget constraint. Our framework shows high deadline completion accuracy in the range of 1.0 - 0.937, which is more accurate than SparkGIS, GeoSpark, GeoMesa and JUST. This also reduces the resource prediction error by 11 percent, considering the geospatial service chain than without it. The cost of the spatio-temporal query is reduced by$\approx$23% in LYRIC, further, the simulation study (using CloudSim) illustrates the efficacy and scalability of LYRIC in terms of optimal budget usage and execution time compared to four baseline approaches.

With the major development of sensor technologies and advancements of communication network infrastructures, there is a growing interest to add more intelligence in the e-health monitoring for facilitating an effective healthcare system. While IoT devices are capable of continuous health-parameter sensing and providing notifications to the user, an effective business process management (BPM) facilitates effective system integration and data processing workflow. This paper proposes an efficient framework for managing emergency situations (specifically, health-related) through the analysis of heterogeneous data sources. The proposed framework, named CLAWER (CLoud-Fog bAsed Workflow for Emergency seRvice) aims to bridge the gap between process management and data analytics by providing an automated workflow for personalized health-monitoring and efficient recommendation system. Here, the IoT devices are used for collecting the movement and health data. The smart phone can act as an edge device to acquire data with user movement information. The accumulated data is initially processed inside the fog device, and finally the analysis and recommendations are generated by the cloud. In this paper the indoor health-status of the users are analysed in small cell cloud enhanced eNode B, which is used as fog device. The generated recommendations are stored in the fog device to provide the recommendations to the users with low latency and in timely manner. The experimental analysis of CLAWER yields better precision and recall values than the existing methods.

Both analyzing mobility traces and understanding a user's movement semantics from mobile sensor data are challenging issues in ubiquitous computing systems. With the pervasiveness of sensor technologies, wireless networks and GPS-equipped devices, a huge volume of location information is being accumulated. Several techniques have been proposed to analyze the mobility traces and extract informative knowledge for varied location-aware applications. However, all of these applications necessitate an effective mobility-analysis framework to capture the movement behavior of individuals in minimum delay. This paper aims to develop a cloud-based mobility analytics framework to model peoples' mobility behaviour in varied granular scale and extract usable knowledge to provision location-aware services. The preliminary experimental results on real-life dataset depict the effectiveness of our proposed framework.

Understanding human interests and intents from movement data are fundamental challenges for any location-based service. With the pervasiveness of sensor embedded smartphones and wireless networks and communication, the availability of spatio-temporal mobility trace (timestamped location information) is increasingly growing. Analysing these huge amount of mobility data is another major concern. This paper proposes a cloud-based framework named MovCloud to efficiently manage and analyse mobility data. Specifically, the framework presents a hierarchical indexing schema to store trajectory data in different spatio-temporal resolution, clusters the trajectories based on semantic movement behaviour instead of only raw latitude, longitude point and resolves mobility queries using MapReduce paradigm. MovCloud is implemented over Google Cloud Platform (GCP) and an extensive set of experiments on real-life data yield the effectiveness of the proposed framework. MovCloud has achieved ~ 28% better clustering accuracy and also executed three times faster than the baseline methods.

Computing optimal routes in a road network is both space and time-consuming. This paper proposes an innovative way of finding a route, given the traffic conditions of every edge present in the map. This article aims to bring about a convergence between methods used in routing data over a network and path planning used in AI technique. It makes use of certain concepts of network routing tables in order to develop path planners suitable for urban conditions. It also has a linear space complexity.

Human mobility prediction is a challenging and crucial task for fostering wide spectrum of location based applications, namely, traffic planning, travel-plan recommendation etc. However, modeling human movement behaviour and predicting next location are non-trivial tasks. We consider the human mobility pattern modeling and predicting next location problem for semantic trajectory data, wherein sparse GPS records are associated with other contextual information. Intuitively, human moves with an intent and the confluence of movement history (GPS traces) and contextual information may help in capturing people's intents at a fine granularity and has potential to improve the efficacy of location prediction problem. In this paper, a hierarchical and layered hidden markov model (HMM) framework for mobility prediction from sparse trajectories has been proposed. Experiment on real-life mobility dataset of Nokia Mobile Data Challenge (MDC) demonstrates the efficacy of the proposed framework even when with sparse GPS traces.