Enhancing Tip of of the Tongue IR Systems
“Tip-of-the-Tongue (ToT) retrieval tackles the challenge of
identifying known items when users recall only vague or fragmentary de
tails. Conventional lexical methods struggle to bridge the semantic gap
inherent in ambiguous queries. In this study, we present a compact two
stage pipeline built around the all-MiniLM-L6-v2 sentence transformer.
First, we fine-tune it on the TREC ToT 2023 Q&A pair dataset, achiev
ing a notable increase in single-stage retrieval performance (nDCG@1000
rising from 0.0322 to 0.1437 and MRR from 0.0005 to 0.0690). Second, we
apply lightweight neural re-ranking—employing both a MonoT5 point
wise re-ranker and a MiniLM-based cross-encoder—and fuse their out
puts via Reciprocal Rank Fusion (RRF). While individual re-rankers
yielded mixed results, RRF consistently enhanced both early- and deep
list metrics (nDCG@1000 up to 0.1498, MRR to 0.0861). Finally, we
report a small-scale zero-shot trial of four GPT variants, observing that
“mini” models outperform their full-size counterparts in top-3 accuracy.
Ourfindings demonstrate that a resource-efficient, fine-tuned transformer,
when coupled with strategic fusion of lightweight re-rankers, can deliver
improved performance on ToT known-item retrieval tasks.”
Advancing Resume Grading Systems with NLP and Explainable AI
“Traditional resume analysing systems rely heavily on keyword matching and rigid formats, which can lead to qualified candidates being overlooked if they don’t conform to the system’s format. Additionally, these systems often function as “black boxes,” providing a final score without any transparency in the decision-making process. This lack of explainability can lead to biases in candidate evaluations, raising concerns about fairness and accuracy. To address these issues, this project introduces a more transparent, section-based resume scoring system.
A hybrid model which integrates Named Entity Recognition, transformer neural networks, and a point-based scoring system was developed. It provides granular evaluations of each resume section. Using Explainable AI (XAI) techniques, the system provides insight into how each section’s score is calculated, promoting trust and accountability. The approach also allows recruiters to adjust evaluation criteria, enhancing flexibility.
The initial Named Entity Recognition (NER) model showed significant results, with an F1 score of 86.84%, a recall of 88.24%, and a precision of 85.48%. Similarly, the transformer regression model demonstrated strong performance, with an R-squared value of 0.896428, a Mean Squared Error (MSE) of 0.194837 and a Mean Absolute Error (MAE) of 0.312830. These results indicate the effectiveness of the models and their potential for further improvement.”
Bias Lens: Systemic Bias Detection with Explainable Analysis
Artificial Intelligence, particularly Natural Language Processing, often inherits and amplifies societal biases, undermining fairness and trust. Existing bias detection tools frequently lack the granularity and transparency needed for effective mitigation, often operating as ‘black boxes’. Bias Lens is a pioneering framework designed to address this critical challenge. It leverages a fine-tuned BERT-based model for multi-label token classification, enabling the identification of six distinct, potentially overlapping, bias types (Generalization, Unfairness, Stereotype, Assumption, Exclusion, Framing) directly at the word level. A key innovation lies in its integration of advanced Explainable AI (XAI) techniques, including a novel enhanced Integrated Gap Gradients (IG2). This provides unprecedented clarity, attributing bias to specific tokens and explaining why content is flagged. Achieving 93% accuracy and significantly faster inference (~8x) than baseline research models, Bias Lens translates complex analyses into actionable insights through an interactive visualization dashboard, accessible to diverse users. By delivering deep, interpretable bias detection and contributing open-source resources (validated dataset, model, XAI library), Bias Lens empowers stakeholders to proactively build fairer, more transparent, and ethically sound AI, fostering responsible technological advancement.
AKURU: Addressing Ad hoc Back-Transliteration with Word Sense Disambiguation in Romanized Sinhala Through a Context-Aware Approach
The growing use of Romanized Sinhala in digital communication platforms brings significant challenges to natural language processing tasks, particularly in backward transliterations which is the process of converting Romanized Sinhala text into the native Sinhala script. Since Romanized Sinhala is an informal way to represent native Sinhala, it lacks standardized spelling conventions leading to ad-hoc typing variations in which users frequently omit vowels, apply inconsistent phonetic spellings and use alternative consonants. This inconsistency exacerbates the problem of lexical ambiguity making it difficult to interpret meanings from Romanized Sinhala text. Existing back-transliteration systems struggle with these ad-hoc typing variations and word sense disambiguation, leading to significant accuracy loss.
To address these challenges in existing back-transliteration systems, this research introduces a novel context-aware hybrid approach that combines an ad-hoc transliteration dictionary and rule-based approach with BERT-based language model trained on native Sinhala text.
The proposed system was evaluated for backward transliteration using Sinhala BERT and its fine-tuned variant, achieving BLEU scores of around 0.91 with remarkably low Word Error Rate and Character Error Rate, approximately 0.09 and 0.02 respectively. Additionally, the first ever Word Sense Disambiguation (WSD) dataset for Romanized Sinhala is introduced as a part of this research. The proposed transliterator achieved an overall F1 score of approximately 0.94 highlighting the effectiveness of the proposed approach in handling ambiguous words in Romanized Sinhala.
LLM based Automatic Speech Recognition for Medical Documentation
This project is dedicated on leveraging Automatic Speech Recognition (ASR) within the medical domain, which emphasizes on refining and enhancing the accuracy of the transcription through Large Language Model (LLM) based approach.
The major challenge discussed is the difficulty of manual documentation which is time consuming and laborious. ASR meets the challenge of transcribing medical conversations, but still struggles to understand the intricacies in patient-doctor consultations. These problems arise mostly because there are complexities in medical language, nuanced phrases, detailed medical terms and people speaking with different accents. Poor performance with special vocabulary and frequent transcription errors are usual for general ASR models in these domain-specific information systems. Different accents can further disrupt word understanding which adds more challenges to transcription. This issue is very serious because inaccurate information from transcription may affect how patients’ treatment and diagnosis. An illustrative example of this problem is that “”Cystic fibrosis”” being misinterpreted as “”65 Roses””.
This work aims to analyze interconnections between context and ASR results related to medical terms and accents which will help to fix parts of current technology and thereby enhance accuracy in ASR. The approach improves the problem area by creating a medical ASR system that considers the context and adapts to the accent used by both patients and healthcare providers.
For its first ASR component, the developed system recorded a Word Error Rate (WER) of 12%. Following this, a Large Language Model (LLM) helped to correct the errors made by speech recognition. This new method with LLMs makes it easier to understand sentences more completely. It depends on deep learning methods, especially neural networks and contextual understanding, for speech recognition in the medical domain.
The outcome of this project is anticipated to serve on optimally deploying ASR in healthcare settings. This research addresses the critical need for domain-specific ASR system which is adaptable to diverse accent and is contextually aware regarding the medical terminology. As a result, this contributes to improve the overall patient satisfaction and the productivity of medical documentation within clinical settings.
Sentiment Analysis and Categorization of Mobile App User Reviews
This project aims to develop an automated system for extracting meaningful insights from mobile app user feedback. It addresses the challenges of analyzing unstructured and complex language in reviews. By integrating machine learning and natural language processing techniques, the project focuses on accurately performing sentiment analysis and categorizing feedback into relevant topics, providing actionable insights for app developers to enhance app quality and user satisfaction.