Dmytro Babych

By utilizing advanced natural language processing techniques, the application extracted relevant information from the text, established relationships between entities, and classified sentiment toward extracted entities and pre-defined terms.

The project aimed to enhance researchers' understanding of biomedical concepts by constructing a knowledge graph representing interrelationships among entities and providing valuable insights into opinions and attitudes expressed in the literature. Ultimately, this automated approach facilitated efficient knowledge discovery and contributed to advancements in the field of biomedical research.

As the lead developer on this project, I implemented key components, including biomedical named-entity recognition, relationship classification, targeted sentiment classification, and sub-graph visualization.

In order to process text “on device” without sending the data to external servers (due to privacy issues), we needed to adopt neural networks (NN) for smartphone usage. Using different techniques like pruning, quantization, and word2bits we were able to achieve such results:

Text Classification:
Model size on disk: 80 MB => 12 MB, 85 % reduced
RAM usage: 85 MB => 25 MB, 71% reduced
Accuracy drop: None

Named Entity Recognition:
Model size on disk: 323 MB => 43 MB, 87 % reduced
RAM usage: 400 MB => 40 MB, 90% reduced
Accuracy drop: None

Workflow included:
1. Training the model is the same as with normal embeddings;
2. Quantization was implemented in C++ TensorFlow source code:
2.1 Finding relevant (biggest) tensor in the model's graph;
2.2 Implementing quantization of -0.33 +0.33 values to 1s and 0s;
2.3 Packing ones and zeros on int8 data type so it’ll take less space;
2.4 Implementation of unpacking operation for inference and replacing old operation with it;
2.5 Saving the transformed graph.
3. Java package implementation for model inference and performance measuring.