Srishti Gureja, Lester James Validad Miranda, Shayekh Bin Islam, Rishabh Maheshwary, Drishti Sharma, Gusti Triandi Winata, Nathan Lambert, Sebastian Ruder, Sara Hooker, Marzieh Fadaee. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

In various applications, the multivariate time series often suffers from missing data. This issue can significantly disrupt systems that rely on the data. Spatial and temporal dependencies can be leveraged to impute the missing samples. Existing imputation methods often ignore dynamic changes in spatial dependencies. We propose a Spatial Dynamic Aware Graph Recurrent Imputation Network (SDA-GRIN) which is capable of capturing dynamic changes in spatial dependencies. SDA-GRIN leverages a multi-head attention mechanism to adapt graph structures with time. SDA-GRIN models multivariate time series as a sequence of temporal graphs and uses a recurrent message-passing architecture for imputation. We evaluate SDA-GRIN on four real-world datasets: SDA-GRIN reduces MSE by 9.51% for the AQI and 9.40% for AQI-36. On the PEMS-BAY dataset, it achieves a 1.94% reduction in MSE. Detailed ablation study demonstrates the effect of window sizes and missing data on the performance of the method. Project page: https://ameskandari.github.io/sda-grin/.

The performance differential of large language models (LLM) between languages hinders their effective deployment in many regions, inhibiting the potential economic and societal value of generative AI tools in many communities. However, the development of functional LLMs in many languages (i.e., multilingual LLMs) is bottlenecked by the lack of high-quality evaluation resources in languages other than English. Moreover, current practices in multilingual benchmark construction often translate English resources, ignoring the regional and cultural knowledge of the environments in which multilingual systems would be used. In this work, we construct an evaluation suite of 197,243 QA pairs from local exam sources to measure the capabilities of multilingual LLMs in a variety of regional contexts. Our novel resource, INCLUDE, is a comprehensive knowledge- and reasoning-centric benchmark across 44 written languages that evaluates multilingual LLMs for performance in the actual language environments where they would be deployed.

During the COVID-19 pandemic, the traditional emergency healthcare systems faced unprecedented strain due to the sharp rise in demands for urgent care, scarcity of resources, and increased risks of people getting infected while waiting at the emergency care facility. We present Triage-Bot, an online medical triage provisioning service, that can revolutionize emergency care by decreasing the load on emergency departments (ED), reducing healthcare expenses, and improving the quality of care. Empowered by artificial intelligence and natural language processing, the Triage-Bot service assesses and prioritizes patients' needs based on symptoms, medical history, and perceived conditions from multimodal video, audio, and text data captured during patients' interactions. The captured summarized information with a severity ranking is sent to a human expert to suggest the next action on the user's part. The diverse data types used by the Triage-Bot in communication, authentication, data collection, storage, and analytics requires a robust and scalable system architecture for online service provisioning. In this paper, we specifically focus on the system design and architecture of the Triage-Bot for emergency healthcare settings. With integrated electronic medical records (EMR) and online platforms, the bot fosters collaboration among healthcare professionals and enables swift and informed decision-making even in the face of crises. By partially automating and offering a hybrid triage process, the Triage-Bot improves resource allocation, reduces healthcare management costs for emergency care, minimizes patient waiting times, and improves wellbeing. To address the complexities and demands of healthcare data management, our proposed system incorporates MongoDB database for flexibility, scalability, and versatility in supporting different types of data. Additionally, we implement a data linking and analytics pipeline utilizing a data Lakehouse system to effectively ingest, manage, process, and generate knowledge from heterogeneous data sources.

The Healthcare system of India provides outreach services to the rural population with a key focus on the maternal and child health through its flagship program of Community Health Workers (CHWs). The program since its launch has reached a scale of over 900000 health workers across the country and observed significant benefits on the health indicators. However, traditional face to face training mechanisms face persistent challenge in providing adequate training and capacity building opportunities to CHWs which leads to their sub-optimal knowledge and skill sets. In this paper, we propose Sangoshthi, a low-cost mobile based training and learning platform that fits well into the environment of low-Internet access. Sangoshthi leverages the architecture that combines Internet and IVR technology to host real time training sessions with the CHWs having access to basic phones only. We present our findings of a four week long field deployment with 40 CHWs using both qualitative and quantitative methods. Sangoshthi offers a lively environment of peer learning that was well received by the CHW community and resulted into their knowledge gains (16%) and increased confidence levels to handle the cases. Our study highlights the potential of complementary training platforms that can empower CHWs in-situ without the need of additional infrastructure.