Skip to content

Implementation of Z-BERT-A: a zero-shot pipeline for unknown intent detection.

License

Notifications You must be signed in to change notification settings

GT4SD/zero-shot-bert-adapters

Repository files navigation

Zero-shot-BERT-adapters

Repository containing code for Zero-shot-BERT-adapters (Z-BERT-A), a two-stage method for multilingual intent discovery relying on a Transformer architecture, fine-tuned with Adapters, which is initially trained for Natural Language Inference (NLI) and later applied for unknown intent classification in a zero-shot setting.

Deployment example of an use-case of Zero-shot-BERT-adapters pipeline

image

Important information

This repository makes use of the module Adapter Transformers.
The wget tool is required for this package, for Windows users please make sure to install it accordingly before running this module.

Installation

In order to install Zero-shot-BERT-adapters it is just needed to execute the following pip command which will make sure everything is accordingly installed. Zero-shot-BERT-adapters uses spaCy with 'en_core_web_trf', this package will try to install it automatically itself if not present but make sure you have it installed through the suggested way of spaCy.

pip install git+https://github.com/GT4SD/zero-shot-bert-adapters.git

Usage information

In order to reproduce the results here there is the sample code which can also be found in an example Jupyter Notebook.

import torch
from zberta.data.data_snli import DataSNLI
from zberta.data.data_banking import DataBanking
import zberta.model.model as model
from zberta.train.trainer import Trainer
from zberta.intent_discovery.unknown_intents import unknown_intents_set
from zberta.intent_discovery.zberta import ZBERTA

model_name = "bert-base-uncased"
training = False
testing = False
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
if training:
    snli = DataSNLI(model_name, device)
    train_iterator, valid_iterator, test_iterator = snli.iterators()
    berta = model.instantiate_model(snli.labels(), snli.output_dim(), device, model_name, snli.nli_labels())
    trainer = Trainer(berta, train_iterator, valid_iterator, test_iterator, device)
    trainer.start_training()
    if testing:
        trainer.start_testing()
z_banking = DataBanking(model_name, device)
z_dataset = z_banking.z_iterator()
z_intents = unknown_intents_set(z_dataset['test']['text'])
berta = model.instantiate_model(z_banking.labels(), z_banking.output_dim(), device, model_name,
                                z_banking.nli_labels(), path="model.pt", dict=True)
zberta = ZBERTA(berta, model_name, z_dataset['test']['text'], z_intents, z_dataset['test']['category'])
z_acc = zberta.zero_shot_intents()
print(z_acc)

For simple usage of the Zero-shot-BERT-adapters pipeline instead it's just needed to load the pipeline and model through this simple code which encapsulate all the complexity behind.

import torch
from zberta.intent_discovery.zberta import ZBERTA
import zberta.model.model as model

model_name = "bert-base-uncased"
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
berta = model.instantiate_model(2, 2, device, model_name,
                                {"entailment": 0, "contradiction": 1}, path="model.pt", dict=True)
zberta = ZBERTA(berta, model_name)

As example run:

zberta.find_new_intents(["I want to buy a book but I lost all my money, where can I make a withdrawal?"])
Output: ['make withdrawal']