bert_descriptors's Introduction

bert_descriptors

BERT's MLM head model exposed as a HTTP service

Example below uses BERT large cased

This service is used by Unsupervised NER

Installation

Setup pytorch environment with/without GPU support using link Multi GPU test

*Make sure to follow conda environment activation instructions. Also tensorflow is not required. So we just neeed to run first.sh, second.sh and third.sh (from 1-4) performing the instructions displayed at end of each step *

Usage

Download a BERT model from Hugginface.

e.g. BERT large cased - pytorch_model.bin,vocab.txt,config.json into current directory

e.g. BERT base cased - pytorch_model.bin,vocab.txt,config.json into current directory

Start server

$ ./run_server.sh

Confirm installation works by

$ wget -O DES "http://127.0.0.1:8087/dummy/0/John flew from entity to Rio De Janiro"

The output DES file should contain

Revision notes

17 Sept 2021

Two kinds of descriptor sets are returned now.
1. Return just the descriptors of the masked position as before (option 0 in GET call)
2. Return descriptors of masked position as well as for [CLS] position (option 1 in GET call). This is useful for detecting entity type of a term/phrase. For example: Imatinib is a entity
3. Common descriptors are now used to filter terms that do not convey any entity information in results. This is to maximize entity information in generated output
Adaptive casing of input tokens to maximize hit with underlying vocab. This is meaningful for cased models only
Full distribution output is turned off by default and needs to be enabled explicitly. The output above is with full distribution option turned on. Current default output