Apoorva Rao Balevalachilu
Here's a summary of the work I did as a software developer at Amazon:
-
Developed natural language understanding model building services.
-
Measured intent classification and named entity recognition and implemented lifecycle management.
-
Benchmarked deep neural network model release workows.
-
Improved utterance interpretation ranking for various business verticals.
-
Launched key features for voice-based search like 'Drop in on Spaces'and query expansion.
-
Load tested search on developer catalogs used in Alexa Skills.
-
Migrated Alexa entity catalogs from one search engine to another.
A presentation on the work done by my team.
An article on the usefulness of entity resolution.
A blog article on how to use Entity resolution for Alexa Skills.
Research from Alexa.
As an intern at Amazon, I implemented the 'Find a Baby Registry' feature for gifters.
Semantic Web-Based Intelligent Technical Helpdesk
Software and IT Engineers spend a considerable amount of time finding nuggets of relevant information in lengthy manuals, googling errors, and searching repositories to find ways to resolve dependencies. The application we created, accepts a user-specified keyword query, performs disambiguation, and returns relevant facts in RDF (Resource Description Framework) N-Quads format. We assimilated the factual knowledge base for this application into a Linked Open Dataset, named LOaD-IT . We created this repository by crawling and ‘semanticizing’ unstructured technical manuals and filtering existing datasets. I co-authored the paper ‘Curating Linked Open Datasets for Software Engineering ,’ published at the International Semantic Web Conference (ISWC 2013). Future releases of the application can be greatly enhanced by implementing techniques from NLP. For example, factual answers to questions, instead of RDF Quads, can be provided to the user in response to free form queries.
Visit http://54.213.4.161:8000 <moved> to try the application and http://www.iswc2013.semanticweb.org/content/demos/16 to read the paper. The linked open dataset is available at http://datahub.io/dataset/load-it
A Sarcasm Detector
The project attempts to develop a system that would determine if a given tweet has a hint of sarcasm.
The features or (not so obvious) traits of sarcastic and non-sarcastic sentences are extracted by training the system with a large set of tweets. An element of sarcasm in a new sentence is determined by finding its k-nearest neighbors. The overall sentiment and common patterns and usage of punctuation, capital letters, etc. in a sarcastic sentence further help in determining the level of sarcasm. A combination of the above two approaches gives a fairly accurate (~71%) result.
Visit http://hashsarcasm.appspot.com/ for a demo.
A Student Mini Project Workflow Management Application
A hosted web application to simplify the administrative workflow of projects within a Computer Science department created using PL/SQL and Oracle Application Express 4.1 using a backend database populated with data from the current year.