Jingshu Liu

• Built from scratch a bilingual neural network based word and phrase embedding mapping framework in Java with Deeplearning4j-0.91.
• Implemented topic model pipelines using clustering on pre-trained unified phrase embeddings.
• Provisioned sparse matrix support and other mathematical optimizations in Nd4j-0.91.
• Designed and built an encoder-decoder framework for sequence modeling with Pytorch-1.2. Fully campatible in CPU and GPU mode which runned in OVH cloud server using the manage tools openstack and nvidia gpu cloud.
• Incoporated pre-trained Transformer based language models into our neural networks for real life scenarios.
• Fine-tuned pre-trained language models for diologues.

Achievements:

• Improved the bilingual multi-word and single-word lexicon induction by an average of 22 points in MAP on client data.
• The new topic modeling system replaced the existing rule-based topic modeling system.

Environments:

• Java
• Python
• Pytorch
• Deeplearning4j
• Keras
• Scikit Learn
• OpenStack-Docker

• Implemented a term extraction and Aspect Based Sentiment Analysis pipeline for simplified and traditional Chinese language in Java with UIMA architecture and ElasticSearch storage.
• Improved Chinese language preprocessing (POS-tagging) for FNLP. Meanwhile, added an innovative Chinese lemmatizer for reduplicated words.
• Data cleaning and visulization using Pandas and R.

Achievements:

• Achieved state-of-the-art results on Aspect Based Sentiment Analysis on Semeval2016 challenge.
• Improved the term extraction accuracy by 50%.

Environments:

• Java
• UIMA
• Deeplearning4j
• Python
• R
• ElasticSearch

• Collaborated with researchers in Duel Project on humain dialogue classification.
• Annotated sentiment analysis corpora with Brat.

Environments:

• Java-Corenlp
• Perl
• Python
• Numpy
• Brat