Working on speech recognition training infra and modeling

For course 11-785 Introduction to Deep Learning (http://deeplearning.cs.cmu.edu/)

Worked on query understanding service optimization

Worked on coreference resolution and character profiling for fan-fiction corpus based on entity-centric clustering method, contributed as a part of a fan-fiction processing pipeline (https://github.com/michaelmilleryoder/fanfiction-nlp).

· Text Error Correction System- Proposed a word- and character-level correction pipeline composed of candidate correction set generation, and GBDT-based candidate ranking with text similarity, language model, and multiple document-wise features- Utilized open datasets to bootstrap the model (LightGBM), and iteratively harvested new training data to be tagged- Crawled and obtained 10GB+ word- and character-level n-gram statistics, stored in LevelDB and Redis- Achieved 15%+ precision (65%→80%) and 4× recall compared to the original rule-based system- Achieved 10× speedup by caching and batching computation and network request- Launched as Thrift service· Text Classification Platform- Extended the classifier training and evaluation pipeline with HTTP communications to support multiple machines- Refactored the task queue and task manager using Celery, RabbitMQ and MySQL to support larger-scale, distributed tasks

· Multiple Documents Reading Comprehension- Utilized Wikipedia entity link and co-reference information to build a distant-supervised reading comprehension dataset composed of 80k+ cloze texts, each with multiple blanks and documents- Formulated a novel multi-level attention neural model with PyTorch to take into consideration multiple blanks and documents, reaching an accuracy of 32+% on this challenging dataset

· Implemented a text classifier based on both word- and character-level CNN with TensorFlow, achieving an accuracy of 78+% on the Movie Review dataset, outperforming the word-level CNN baseline· Experimented on open information extraction using bi-LSTM sequence-to-sequence model with attention and copy mechanism to generate relation triples from sentences