Projects

TraceHub - A Platform to bridge gap between State-Of-The-Art Time-Series Analytics and Datasets

Paper PDF
(IBM Research)

  • Awarded Best Technical Demonstration, AAAI 2020
  • A platform that connects new non-trivial state-of-the-art time-series analytics with datasets from different domains
  • Analytics owners can run their insights on new datasets in an automated setting to find insight’s potential and improve it.
  • Dataset owners can find all possible types of non-trivial insights based on latest research.
  • A plug-n-play system as a set of Dataset, Transformer pipeline, and Analytics APIs for both kinds of users.
MAi : Model Acquisition Interface for Declarative Dialogue Design of Goal Oriented Multi-Turn Conversation

Paper PDF
(IBM Research)

  • Declarative specification of dialogue agent’s behavior
  • Compile the complete implicit graphs from this compact specification using a non-deterministic AI planner
  • Frontend in Bootstrap and backend as a Flask server
  • Deployed overall service infrastructure using Docker containers
  • Hosted the containerized platform on Kubernetes clusters enabling high-availability and auto-scaling. CI/CD using Travis
  • Publication in AAAI Proceedings 2019
HOVOR : Goal Oriented Dialogue Agent Executor for open domain models

Paper PDF
(IBM Research)

  • Runtime execution of a deployed conversation agent’s contingent plan
  • Action execution in the declarative environment based on its precondition & outcome determination best describing the real world change
  • A general purpose executor for open domain models including dialogue systems
  • AAAI Symposium, Intex, DEEP-DIAL 2019 publications
  • US Filed Patent : Link
Aligning English Sentences With AMR graphs using ILP

CODE

  • AMR: Semantic formalism to English natural language encoding meaning of a sentence in a rooted graph
  • Idea of approach was predicting concepts invoked by words in a sentence is same as aligning words to those concepts
  • Extracted linguistic background knowledge from sentences like lemma, part of speech, modals, named entities, question tokens
  • Concepts in AMR split in nine categories. Learnt ILP rules for each category that invoke AMR concepts from sentence-AMR graph pairs in the training data
  • Learnt ILP rules using open source system XHAIL deriving hypothesis in three steps : grounding, finding kernel, hypothesis generation
  • Dataset consisted of 13050 AMR/English sentence pairs inclusive of 200 development and test pairs
  • Performance of the aligner was measured using precision, recall and f-score measures on test dataset [P=0.971 | R=0.858 | F=0.91]
Semantic Search on Movie Database

CODE

  • Created a text based system that predicts movie names on input user query
  • Dataset of movie summaries text crawled from IMDB
  • Proposed and implemented a semantic approach to find similarity between query and movie summary texts. Created a movie graph of events with Characters(nodes) and Events(edges)
  • Used a semantic K(knowledge)-parser to extract events from query and movie summaries
  • Used multiple similarity scores to calculate similarity between input and movie graphs.} \item {Used NER similarity using Stanford CoreNLP, Term similarity using WS4J’s PATH, LIN, LESK algorithms and Tf-IDF
  • Used NLTK for NER detection and for name-co-reference unification of text
  • Evaluated results using a hand prepared test dataset of 50 movies