Corporate entities increasingly embed environmental claims in their digital communication to project sustainability awareness. Detecting such claims is critical for regulatory monitoring, corporate accountability, and mitigation of greenwashing practices. Traditional neural network architectures including large language models however, struggle to capture both the complex linguistic structures and the subtle stylistic cues that characterize environmental assertions. In this work, we propose a novel Graph-Augmented Liquid Neural Network (GLNN) architecture for automatic detection of environmental claims in corporate web content. Our approach first models the syntactic and semantic dependencies of text using a Graph Convolutional Network (GCN), while concurrently encoding stylistic features derived from linguistic markers (e.g., LIWC categories) into vector representations. These representations are concatenated and passed into a Liquid Time-Constant (LTC) Network, which provides dynamic adaptability and low-power efficiency by leveraging continuous-time recurrent dynamics. The integration of GCN-based stylistic encoding with LTC networks enables the model to robustly capture both structural dependencies and temporal signal variations inherent in corporate claims, while remaining energy efficient. Extensive experiments on multiple open datasets demonstrate that our model outperforms baseline neural architectures in both accuracy and computational efficiency, highlighting the potential of graph-augmented liquid networks as a foundation for sustainable AI in sustainability monitoring.

Accurate prediction of a patient’s Length of Stay (LoS) in the hospital plays a pivotal role in effective healthcare management, particularly in supporting administrative planning, improving care delivery, optimizing resource allocation, and reducing operational costs. In this paper, we present StayLTC, a multimodal deep neural framework designed to forecast real-time LoS using Liquid Time-Constant Networks (LTCs). LTCs, with their continuous-time recurrent dynamics, are evaluated against conventional time series models by leveraging both structured Electronic Health Record (EHR) data and unstructured clinical notes. Experiments conducted on the MIMIC-III dataset demonstrate that LTCs substantially outperform several state-of-the-art time series baselines, offering superior predictive accuracy, robustness, and computational efficiency. Notably, LTCs exhibit comparable performance to large language-based time series models in LoS prediction, while requiring substantially lower computational resources and memory, highlighting their potential to advance resource-efficient Natural Language Processing (NLP) applications in healthcare.

As sustainable practices grow, so does greenwashing, where companies overstate their environmental efforts. To tackle this, we propose EcoClaim, a lightweight Liquid Time-Constant (LTC) network for accurate environmental claim detection. LTC models effectively capture temporal dependencies and adapt to diverse linguistic structures, ensuring strong performance across datasets. Evaluated on formal reports (Dataset-ECD) and informal social media data (Dataset-GCC), EcoClaim demonstrates adaptability while using only 0.612 MB of memory—far less than transformers’ 20.3 GB. Though BERT embeddings with stylistic features achieve similar accuracy, our LTC-based approach with style-infused GloVe embeddings offers superior efficiency. Benchmarked against LLAMA-3.1 8B and Mistral 7B, EcoClaim highlights the potential of liquid neural networks as scalable, resource-efficient solutions for environmental claim detection and broader text classification.

This paper proposes the development of a WCAG-compliant chatbot capable of generating multimodal content to enhance usability for all users. While LLM-based chatbots excel in generating varied responses, they often struggle with ambiguous or incomplete queries, leading to misaligned outputs. We introduce a framework that formulates domain-specific, persona-driven follow-up questions to clarify ambiguities, utilizing knowledge graphs and human feedback. The system refines queries before generating responses by employing a domain-Specific Multilayer Hierarchical Relational Graph (MHRG) to model user intent. Our preliminary evaluations indicate that the Accessibility Bot improves response relevance and quality as compared to existing techniques.

Claim identification is a vital task in natural language processing, aiming to identify assertive statements within text. With the growing influence of generative AI, automated fact-checking solutions are becoming increasingly important to counter misinformation. Although traditional neural networks have shown potential, they often struggle to capture the temporal dynamics inherent in language due to its sequential nature. In this paper, we introduce StyleLTC, a model that leverages liquid neural networks, distinguished by their continuous-time recurrent properties, to address these limitations. Our model also incorporates various stylistic features in conjunction with language models to predict whether a statement qualifies as a claim. We conducted extensive evaluations using multiple open datasets. The results reveal that liquid neural networks outperform static models, delivering higher accuracy, robustness, and resource efficiency. In comparison to open-source large language models, liquid neural networks excel in claim identification, offering superior performance with significantly lower computational and memory demands. Notably, StyleLTC achieves comparable accuracy using 0.612 MB of memory, whereas traditional transformer models require 20.3 GB of GPU memory, highlighting their potential to advance this NLP task. Our findings highlight the potential of LTCs for scalable, effective claim detection in real-world scenarios, contributing to the broader fight against misinformation.

Clinical notes that describe details about diseases, symptoms, treatments, and observed reactions of patients to them, are valuable resources to generate insights about the effectiveness of treatments.Their role in designing better clinical decision making systems is being increasingly acknowledged.However, the availability of clinical notes is still an issue due to privacy violation concerns.Hence most of the work done are on small datasets and neither the power of machine learning is fully utilized, nor is it possible to validate the models properly.With the availability of the Medical Information Mart for Intensive Care (MIMIC-III v1.4) dataset for researchers though, the problem has been somewhat eased.In this paper we have presented an overview of our earlier work on designing deep neural models for prediction of outcomes and hospital stay for patients using MIMIC data.We have also presented new work on patient stratification and explanation generation for patient cohorts.This is early work targeted towards studying trajectories for treatment for different cohorts of patients, which can ultimately lead to discovery of low-risk models for individual patients to ensure better outcomes.

In this paper, we present a grammar and style aware transformer-based neural network for computing the quality of a text in an automatic essay-scoring task. The proposed model takes into consideration different grammatical error categories and discourse writing styles like, concreteness, uncertainty, conviction and commitment in text along with the pre-trained language models of a text document. We have evaluated the proposed model with the automated student assessment dataset. Our preliminary investigation shows that incorporating such stylistic vectors and grammatical error categories with the BERT based language model can give us a better understanding of improving the overall evaluation of the input essays.

The consideration of accessibility has become a crucial aspect in system design to eliminate barriers for people with disabilities. This paper presents the android application which transliterates Indian languages like Odia, Bengali, Hindi, Telugu and English to their respective braille script. It includes the prospect of voice-based search apart from the text-based search for people with visual disabilities to access reading materials in Indian regional braille scripts. The goal of this initiative is to foster an inclusive learning environment and assist special educators in acquiring, sharing, and delivering educational resources that are accessible to all.

This study aims to identify the critical factors influencing the user experience of contact tracing apps and the sentiments around them. For this purpose, we used Google play reviews of Aarogya Setu, a contact tracing app developed in India. First, we establish the relationship between review sentiment and review rating using regression between sentiment polarity and review rating. Then, we used a hybrid aspect-based sentiment analysis approach that uses unsupervised linguistic techniques to determine statistically significant concepts present in the review texts and cluster them into representative aspects that were then tagged under human supervision. Finally, supervised deep learning methods were applied for exhaustive extraction of the aspects and associated sentiments from the reviews. The final exercise of determining the key influencing factors was done by grouping these aspects under factors identified by marketing experts. A total of nine factors were identified, with the usefulness of the app being the most important factor. The findings of this study are essential for the development team and government to improve the application and increase adoption.

Motivated by the problem of optimization of force-field systems in physics using large-scale computer simulations, we consider exploration of a deterministic complex multivariate response surface. The objective is to find input combinations that generate output close to some desired or “target” vector. Despite reducing the problem to exploration of the input space with respect to a 1-D loss function, the search is nontrivial and challenging due to infeasible input combinations, high dimensionalities of the input and output space and multiple “desirable” regions in the input space, and the difficulty of emulating the objective function well with a surrogate model. We propose an approach that is based on combining machine learning techniques with smart experimental design ideas to locate multiple good regions in the input space. Note to Practitioners—ReaxFF is a force field that incorporates complex functions with associated inputs in order to describe the inter- and intra-atomic interactions in materials systems. A typical ReaxFF force field consists of hundreds of parameters (inputs) per element type. During the development of a force field for a molecular system of interest, using computer simulations, these parameters are optimized to reproduce hundreds of material properties close to some benchmark reference values. Finding “good” combinations of hundreds of parameters that produce hundreds of reference values close to their gold standards is a challenging problem because there may be several parameter combinations that may be “almost equally good” or “equally desirable.” To add to the complication, several input combinations simply lead to a system crash, not producing any output at all. Standard global optimization methods do not address such a problem. We propose a novel framework that can address this problem. Beyond the ReaxFF optimization, it can be applied to multiobjective optimization in engineering and the physical sciences, where there are unknown constraints and the focus is on obtaining several good points that can serve as alternatives to a single global optimum.

The predominance of biased articles and its consumption by the readers is becoming a considerable issue. Researchers across domains have made efforts to mitigate biases in language. However, due to the subjective nature of the problem, it is not trivial to detect bias embedded in a text. In this paper, we propose a deep linguistically informed multi-task transformer-based model to automatically detect bias in written text. The model is fine-tuned with a domain-specific corpus and further trained for learning the objectives. We evaluate the performance of the proposed model with respect to baseline systems across multiple datasets. We observed that augmenting linguistic features along with contextual embedding improves the performance of the neural network model to automatically detect bias in text.

Persuasive conversation leverages conversational strategies by the persuader to change the attitude or behavior of a persuadee towards achieving a specific goal. It involves understanding the linguistic and cognitive principles underlying the organization of strategic disclosures and appeals employed in human persuasion. One of the main challenges of such conversation is the inability of a persuader to detect the outcome of their conversation during the interaction. Such prior knowledge can help a persuader to change their conversation strategy and pre-empt possible conversation failures. In this paper, we propose a technique that analyses conversations to predict whether the persuader is going to successfully persuade the persuadee. We propose a joint model of latent utterance categorization to predict the success or the failure of a persuasive conversation. This latent categorization allows the model to identify high-level conversational contexts that influence patterns of language in a persuasive conversation. We evaluate the performance of our model on an openly available dataset. Our preliminary results demonstrate that the proposed model outperforms competitive baselines.

Automatically generating or ranking distractors for multiple-choice questions (MCQs) is still a challenging problem. In this work, we have focused towards automatic ranking of distractors for MCQs. Accordingly, we have proposed an semantically aware CNN-BiLSTM model. We evaluate our model with different word level embeddings as input over two different openly available datasets. Experimental results demonstrate our proposed model surpasses the performance of the existing baseline models. Furthermore, we have observed that intelligently incorporating word level semantic information along with context specific word embeddings boost up the predictive performance of distractors, which is a promising direction for further research.

In this paper we have proposed a linguistically informed recursive neural network architecture for automatic extraction of cause-effect relations from text.These relations can be expressed in arbitrarily complex ways.The architecture uses word level embeddings and other linguistic features to detect causal events and their effects mentioned within a sentence.The extracted events and their relations are used to build a causal-graph after clustering and appropriate generalization, which is then used for predictive purposes.We have evaluated the performance of the proposed extraction model with respect to two baseline systems,one a rule-based classifier, and the other a conditional random field (CRF) based supervised model.We have also compared our results with related work reported in the past by other authors on SEMEVAL data set, and found that the proposed bidirectional LSTM model enhanced with an additional linguistic layer performs better.We have also worked extensively on creating new annotated datasets from publicly available data, which we are willing to share with the community.

In this paper, we have proposed natural language processing and deep learning based techniques for the automatic extraction and curation of occupational health and safety related information from safety-related articles. Such articles typically contain details of the organizations that have been cited for violating the health and safety regulations, safety-related issues and incidents, the location of the incident, and finally details of the penalties incurred. We have done experiments with a collection of 5400 related articles. The end-product of our work is an occupational risk-register that contains details of safety incidents across geographies and time. This register can be further utilized for analytical and reporting purposes. Such information is extremely valuable to industries which see a high occurrence of occupational injuries.

News articles from different sources regularly report crime incidents that contain details of crime, information about accused entities, details of the investigation process and finally details of judgement. In this paper, we have proposed natural language processing techniques for extraction and curation of crime-related information from digitally published News articles. We have leveraged computational linguistics based methods to analyse crime related News documents to extract different crime related entities and events. This includes name of the criminal, name of the victim, nature of crime, geographic location, date and time, and action taken against the criminal. We have also proposed a semi-supervised learning technique to learn different categories of crime events from the News documents. This helps in continuous evolution of the crime dictionaries. Thus the proposed methods are not restricted to detecting known crimes only but contribute actively towards maintaining an updated crime dictionary. We have done experiments with a collection of 3000 crime-reporting News articles. The end-product of our experiments is a crime-register that contains details of crime committed across geographies and time. This register can be further utilized for analytical and reporting purposes.

There has been a growing need to automatically identify, extract and analyze risk related statements from textual data. In this paper, we have exploited natural language processing research to develop a risk analytics framework that processes human-reported risk statements to analyzes the enterprise risk description texts to classify them into valid and invalid risk categories, and perform analytics to extract information from the text pertaining to the different categories of risks and their possible cause and impacts. A manual annotation study from management experts using risk descriptions collected for a specific organization was conducted to evaluate the framework. The evaluation showed promising results for automated risk analysis and identification.

We present design and development of a web browser that allow easy dissemination of information through World Wide Web for people with cerebral palsy in India. Our focus user group comprises people with severe form of spastic cerebral palsy and highly restricted motor movement skills. Throughout the development process we have interacted with the target users to understand their requirements and to get design advises. The browser is augmented with an intelligent auto-scanning mechanism through which the web contents and browser GUI controls can be accessed with less time and effort. We have field tested the browser with the target users where preliminary evaluation results suggests that the proposed browser is quite effective in terms of task execution time, cognitive effort and overall usability.

In this paper we have developed a toolset that will allow people with severe spastic cerebral palsy (CP) and highly restricted motor movement skills to access popular social-networking and communication mediums like, Facebook and E-mails. To understand the requirements of the intended users we have performed a number of surveys that acted as basis of our system design. The developed tools use special access switch based scanning technique for easy navigation in different applications. We have evaluated the toolset with six target users. The preliminary results demonstrate a positive response.

The paper presents design and development of WebSanyog, an Android based web browser that helps people with severe form of spastic cerebral palsy and highly restricted motor movement skills to access web contents. The target user group has acted as our design advisors through constant interaction during the whole process. Features like, auto scanning mechanism, predictive keyboard and intelligent link parser make the system suitable for our target users. The browser is primarily developed for mobile and tablet based devices keeping in mind the portability issue.

This work intends to understand and model the kinetic aspect or the change of substrate weight over time in the selective growth of silica nanowires (NWs) catalyzed through Pd thin film. Various adsorption-induced, diffusion-induced, or unified vapor-liquid-solid (VLS) growth models have been developed to describe the NW length varying with time. Since NW length has been difficult to be measured, substrate weight change is therefore used as an alternative in this study to investigate growth kinetics of NWs. We investigate six different weight kinetics models in predicting weight changes during growth. Model estimation and comparison are conducted using both maximum-likelihood estimation (MLE) and Bayesian approaches. Owing to the embedded kinetics information in the nonlinear growth models, the Bayesian hierarchical model is shown to be more desirable when process data is limited.

In this paper we present a speech enabled bidirectional automatic Indian language text to Braille transliteration system. The system allows bridging the communication gap between a visually impaired and a sighted person. The present system can be configured to take Indian language text document as input and based on some transliteration rules, can generate the corresponding Braille output. The system is augmented by an Indian language text-to-speech (TTS) system through which a user can get instantaneous audio feedback from the input text. We further extended the system to support transliteration of Dzongkha1text to Braille. Finally we present an Audio QWERTY editor which allows a visually impaired person to read and write Indian language texts through a computer.

This paper presents a prototype Text-To-Indian Sign Language (ISL) translation system. The system will help dissemination of information to the deaf people in India. This paper also presents the SL-dictionary tool, which can be used to create bilingual ISL dictionary and can store ISL phonological information.