Natural Language Processing

From year 2012 till the beginning of my PhD, I was working with several problems related to Natural Language Processing (NLP) for academic research as well as industry products. Despite major shift in the technological landscape, foundational approaches have guided development and evaluation of modern AI based systems.

Research Focus

My NLP work studies how language technologies handle noisy, multilingual, informal, and real-world text — especially where standard assumptions about clean vocabulary, grammar, or monolingual input break down.

Code-Mixed Language

Modeling text where users naturally combine Hindi, English, scripts, spellings, slang, and cultural expressions.

A group of colorful speech bubbles on a wooden wall
A group of colorful speech bubbles on a wooden wall
Opinion Mining

Analyzing how people express attitudes, emotion, evaluation, and opinion in noisy online text.

Using NLP in real-world products, including recruitment intelligence, search, compliance, and education.

Applied NLP Systems
person in white shoes standing on gray concrete road
person in white shoes standing on gray concrete road
stacks of paper documents and file folders
stacks of paper documents and file folders

Analyzing Hi-NGLISH Text

This research project developed a solution to process code mixed text, that is, words or phrases borrowed from one language written in another script. The paper about this work presents a usecase where we aim to find the sentiment of the message. This work addressed an early and still relevant challenge in multilingual NLP: real users do not write in clean, standardized language. In code-mixed settings, spelling variation, transliteration, slang, and informal grammar make word-level modeling fragile. The project showed that sub-word patterns can provide a more robust representation for noisy language.

Problem

Hindi-English code-mixed social media text is highly noisy, informal, and inconsistent. The same Hindi word may appear in many Roman-script spellings, while sentence structure remains flexible. This makes sentiment analysis difficult for models that expect clean, monolingual, or standardized text.

HIGHLIGHT

Gap

Existing sentiment analysis methods were not well-suited for this setting. Word-level models struggled with sparse and constantly changing vocabulary, lexicon-based systems failed on misspellings and transliteration errors, and many NLP tools required resources such as parsers or language-specific preprocessing that were not available for code-mixed text.

Approach

I created and annotated a Hindi-English code-mixed sentiment dataset from public social media comments, then compared several sentiment analysis methods existing at that time. The main contribution was a Subword-LSTM model that learned morpheme-like character patterns using convolutional filters and then used an LSTM to model sentiment across the sentence.

Results

The Subword-LSTM achieved the strongest performance on our dataset, reaching 69.7% accuracy, outperforming both traditional baselines and a character-level LSTM. The results suggested that sub-word representations were useful because they captured sentiment-bearing parts of noisy and misspelled words.