Archive for category: Publications

They developed PACE a Python AI companion for Enhanced Engagement. For this work they generated synthetic data from GPT 3.5 turbo for scaffolding and conversion fine-tuning. The LORA fine tuned Gemma 2B model was used for making the system relatively lightweight. This trains the LLM model to breakdown complex problems into subproblems and generate hints and structured steps for the students. On the other hand the conversation data allows the LLM to engage with users using natural, human-like dialogue, to avoid hallucinations, to supports error correction and detailed feedback, and to enhances motivation and interest through interaction with users with different learning styles, and pace. Evaluation of the system was also performed.

A wider evaluation of the system is underway and we plan to make the version 2 of the system available to our intro to python programming students. Interaction datasets collected from students (with their consent of course) will be valuable for making such a system more reliable.

We hope to see more exciting work from Ibteda and Ashfaq in the near future.

A paper titled “Capturing the spatiotemporal inequality in electricity consumption at the subnational level of Bangladesh using Nighttime Lights” has been accepted for publication in the Journal of the Asia Pacific Economy (SJR Q2, H-index 38, (Scopus) CiteScore 3.7 in 2023).

The research work was led by Dr Amin Masud Ali, Professor, Dept of Economics, JU and a co-director / supervisor of Data Science wing, CCDS. The paper is co-authored by Muntasir Wahed (then RA of DnDLab and Data Science wing, currently PhD student at UIUC), Dr Amin Ahsan Ali (Dept of CSE, IUB, and Director, AI & ML Wing, CCDS), and Dr Moinul I Zaber (Dept of CSE, DU and a collaborator of the Data Science wing, CCDS).

This paper examines the spatiotemporal inequality in electricity consumption at the subnational level (Zila and Upazila/Thana) of Bangladesh using nighttime light (NTL) data. The NTL data, sourced from the Visible Infrared Imaging Radiometer Suite (VIIRS) day/night band (DNB) for the period from 2013 to 2020, reveals persistent variability in electricity consumption among the districts. Notably, the gap between urban and non-urban areas has widened. While within district inequality (measured by NTL Gini) has declined over time, it remains high in several districts. Convergence analysis confirms that while lagging districts are showing a catching up effect, the sub-districts are diverging among themselves (in terms of mean NTL per capita). Interestingly, the rural sub-districts are converging among themselves despite urban sub-district divergence. The study also identifies regions with significant imbalance between NTL, population, and built-up area density values.

These findings have implications for policymakers aiming to ensure electricity for all and reduce inequality. First, these findings provide a clear picture of the NTL inequality pattern at the subnational level of the country. The findings should contribute to the process of ensuring electricity for all (SDG, Goal 7) producing and monitoring the evolution of inequalities within the country (SDG, Goal 10) to achieve the sustainable development goal of reducing inequalities. Secondly, this investigation also captures the inequality in regional economic development as NTL is a recognized proxy indicator of poverty, public service coverage, and economic activity.

Congratulations to our senior project student Fahim Ahmed and research assistant Md Fahim for getting their paper accepted into the core rank A conference, European Conference on AI (ECAI) https://www.ecai2024.eu/ . The acceptance rate was very competitive (24%) this time for ECAI 2024. The title of the paper is, “Improving the Performance of Transformer-based Models Over Classical Baselines in Multiple Transliterated Languages”.

Here is a short description of the paper:

Online discourse, by its very nature, is rife with transliterated text along with code-mixing and code-switching. Transliteration is heavily featured due to the ease of inputting romanized text with standard keyboards over native scripts. Due to its ubiquity, it is a critical area of study to ensure NLP models perform well in real-world scenarios.

In this paper, we analyze the performance of various language model’s performance on classification of romanized/transliterated social media text. We chose the tasks of sentiment analysis and offensive language identification. We carried out experiments for three different languages, namely Bangla, Hindi, and Arabic (for six datasets). To our surprise, we discovered across multiple datasets that the classical machine learning methods (Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF), and XGBoost) perform very competitively with fine-tuned transformer-based mono / multilingual language models (BanglishBERT, HingBERT, and DarijaBERT, XLM-RoBERTa, mBERT, and mDeBERTa), tiny LLMs (Gemma-2B, and TinyLLaMa) and ChatGPT for classification tasks in transliterated text. Additionally, we investigated various mitigation strategies such as translation and augmentation via the use of ChatGPT, as well as Masked Language Modelling to dataset-specific pretraining for language models. Depending on the dataset and language, employing those mitigation techniques yields a 2-3% further improvement in accuracy and macro-F1 above baseline.

We demonstrate TF-IDF and BoW-based classifiers achieve performance within around 3% of fine-tuned LMs and thus could thus be considered as a strong baseline for transliterated text-based NLP tasks.