Research
Projects
Funded and self-directed work, from real-time broadcast analytics to counter-terrorism network analysis.
Completed · 5
Social Security Threat Assessment through Opinion Mining and Sentiment Analysis
Design and development of a framework for assessing social security threats by mining opinion and sentiment from social media content.
Multimodal Broadcast Analytics
A multi-modal broadcast analytics system that detects and tracks events of security concern from continuous, heterogeneous data streams — TV news channels and news websites. Because the source data is multi-modal, the system is built from three sub-systems: video analysis, audio analysis and speech recognition, and text analytics.
Input: multi-modal data from TV news channels and web news articles — video, audio, images, and text. Broadcast news video is pre-processed to segment streams into news stories and extract overlay text, speech transcripts, and metadata.
Output: multi-modal event retrieval against free-form text queries, and timeline creation for profiling entities of interest such as terror groups and organizations.
Text-to-Speech Synthesis of Manipuri Language
Speech synthesis is the artificial production of human speech; text-to-speech (TTS) converts linguistic information stored as text into speech. The objective here is to convert arbitrary Manipuri text into its corresponding spoken waveform, mainly through concatenative unit-selection and parametric synthesis.
For concatenative unit-selection synthesis, the group used the open Festival framework, built on large databases of recorded speech segmented into phones, bi-phones, and syllables — the last chosen as the primary unit since Indian languages are largely syllabic. Segmentation used hybrid HMM-with-Group-Delay methods, indexed by acoustic parameters (pitch, duration, syllable position) and neighboring-syllable context, with unit selection performed via classification and regression (CART) trees.
Flite served as an alternative synthesis engine designed for small embedded machines and large servers. For parametric synthesis, the group used HMM-based synthesis (HTS): spectral and excitation parameters are extracted during training to build context-independent monophone HMMs, then context-dependent HMMs are concatenated at synthesis time and rendered to waveform via an MLSA filter.
Aakash Application Development
Dedicated to developing useful applications and content for the Aakash tablet, aiming to empower teachers through a blend of technology, e-content, and pedagogy.
Projects built under this initiative:
- Examination Conducting System
- Interactive C
- Note App
- eDiscussion
VISHLESHAKEE 2: Unified Platform for Social Media Content Analytics
A unified platform for social media content analytics.
Ongoing · 1
AI Assisted Legal Translation of Judgements from English to Assamese
Machine translation of court judgements from English to Assamese, developed for the Gauhati High Court.
Other Research · 3
Counter-Terrorism and Homeland Security using Social Network Analysis
Open information sharing on social media poses new challenges for national security and public safety, and one of the core challenges for counter-terrorism agencies is turning a huge real-time information stream into usable intelligence. Just as groups have used social media for propaganda, recruitment, and coordination, social and communicative media can also be used to combat and track terrorism — recognized as important well before 2001, but drawing far more research attention since.
Most counter-terrorism datasets used in the past are homogeneous, but information from social media is heterogeneous — different attribute types describing the same event. This work proposes a social network analysis framework that captures that heterogeneity, and predicts relationships between attributes and organizations involved in an attack, including:
- The likelihood of a terrorist organization attacking a particular country or city in future
- Given a future attack, the likelihood of a given target type (government building, crowded place, etc.)
- Similarities between different terrorist organizations
The method generalizes to analyzing the network from different perspectives without modifying the underlying model, using the Global Terrorism Database (GTD), which covers terrorist attack information from 1970–2014.
Analyzing Political Integrity of Indian Politicians using Twitter
Political parties split, governments become unstable, politicians change parties — this study aims to predict integrity of candidates within a political party using their tweets. The group identifies the topics politicians comment on and classifies each as positive, negative, or neutral using sentiment analysis, then uses statistical inference over the results to measure how integrated candidates are within their party. The study aims to extend to other parties and to journalists.
Event Detection
Ongoing work on detecting events from open, continuous data streams.