End-To-End Crowdsourcing

Last update: Mar 06, 2022

Related tags

Overview

End-To-End Crowdsourcing

Comparison of traditional crowdsourcing approaches to a state-of-the-art end-to-end crowdsourcing approach LTNet on sentiment analysis. LTNet is adapted from "Facial Expression Recognition with Inconsistently Annotated Datasets" to text data. It encompasses a simple attention based neural network and utilizes confusion matrices as a noise reduction technique. For comparison, the traditional ground truth estimators "Fast-Dawid-Skene" and "MACE" are applied.

This codebase was used in both "End-to-End Annotator Bias Approximation on Crowdsourced Single-Label Sentiment Analysis" and "Deep End-to-End Learning for Noisy Annotations and Crowdsourcing in Natural Language Processing".

Training

This is an example training procedure for the TripAdvisor dataset. The dataset and solver objects are initialized before a standard LTNet model is trained for 300 epochs.

import torch
import pytz
import datetime

from datasets.tripadvisor import TripAdvisorDataset
from solver import Solver
from utils import *

# gpu
DEVICE = torch.device('cuda')

# cpu
# DEVICE = torch.device('cpu')

label_dim = 2
annotator_dim = 2
loss = 'nll'
one_dataset_one_annotator = False
dataset = TripAdvisorDataset(device=DEVICE, one_dataset_one_annotator=one_dataset_one_annotator)

lr = 1e-5
batch_size = 64
current_time = datetime.datetime.now(pytz.timezone('Europe/Berlin')).strftime("%Y%m%d-%H%M%S")
hyperparams = {'batch': batch_size, 'lr': lr}
writer = get_writer(path=f'../logs/test',
                    current_time=current_time, params=hyperparams)

solver = Solver(dataset, lr, batch_size, 
                writer=writer,
                device=DEVICE,
                label_dim=label_dim,
                annotator_dim=annotator_dim)

model, f1 = solver.fit(epochs=300, return_f1=True,
                       deep_randomization=True)

These initialization and training steps of a network are abstracted away into src/training. Scripts with many more details on training procedures and different configurations can be found in src/scripts. All are best loaded into an ipython terminal with the %load command.

Databases

How to use them from outside the src folder?

It makes us able to refer to the classes properly.

import sys
sys.path.append("src/")

Pass the root folders of the embeddings and the data.

from datasets.emotion import EmotionDataset

dataset = EmotionDataset(
        text_processor='word2vec', 
        text_processor_filters=['lowercase', 'stopwordsfilter'],
        embedding_path='data/embeddings/word2vec/glove.6B.50d.txt',
        data_path='data/'
        )

Datasets are available at "TripAdvisor", "Emotion" and "Organic".

TripAdvisor Dataset

code

from datasets.tripadvisor import TripAdvisorDataset

dataset = TripAdvisorDataset(text_processor='word2vec', text_processor_filters=['lowercase', 'stopwordsfilter'])

print(f'Dataset is in {dataset.mode} mode')
print(f'Train-Validation split is {dataset.train_val_split}')
print(f'1st train datapoint: {dataset[0]}')

output

Dataset is in train mode
Train-Validation split is 0.8
1st train datapoint: {'label': 0, 'annotator':'f', 'rating': 4, 'text': 'I realise ...', 'embedding': array}

Emotion Dataset

Every headline has been annotated on each emotion. One can select one emotion as the label by the set_emotion method.

code

from datasets.emotion import EmotionDataset

dataset = TripAdvisorDataset(text_processor='word2vec', text_processor_filters=['lowercase', 'stopwordsfilter'])

print(f'Dataset is in {dataset.mode} mode')
print(f'Train-Validation split is {dataset.train_val_split}')
dataset.set_emotion('anger')
print(f'1st train datapoint: {dataset[0]}') # select anger_label as label
dataset.set_emotion('disgust')
print(f'1st train datapoint: {dataset[0]}') # select disgust_label as label

output

Dataset is in train mode
Train-Validation split is 0.8
1st train datapoint: {'label': 0, 'annotator':'xxx1', 'anger_response':0, 'anger_label':0, 'anger_gold'=1, 'disgust_response':0 ... 'text': 'I realise ...', ... 'embedding': array}
1st train datapoint: {'label': 1, 'annotator':'xxx1', 'anger_response':0, 'anger_label':0, 'anger_gold'=1, 'disgust_response':0 ... 'text': 'I realise ...', ... 'embedding': array}

End-To-End Crowdsourcing

Related tags

Overview

End-To-End Crowdsourcing

Training

Databases

How to use them from outside the src folder?

TripAdvisor Dataset

Emotion Dataset

Owner

Andreas Koch

UCSD Oasis platform

Attentional Focus Modulates Automatic Finger‑tapping Movements

Learning to trade under the reinforcement learning framework

DGL-TreeSearch and the Gurobi-MWIS interface

Unofficial implementation of Proxy Anchor Loss for Deep Metric Learning

DenseNet Implementation in Keras with ImageNet Pretrained Models

WHENet: Real-time Fine-Grained Estimation for Wide Range Head Pose

Topic Discovery via Latent Space Clustering of Pretrained Language Model Representations

Implementation of Kaneko et al.'s MaskCycleGAN-VC model for non-parallel voice conversion.

Official PyTorch implementation for paper "Efficient Two-Stage Detection of Human–Object Interactions with a Novel Unary–Pairwise Transformer"

Simple Tensorflow implementation of Toward Spatially Unbiased Generative Models (ICCV 2021)

Code for GNMR in ICDE 2021

Local-Global Stratified Transformer for Efficient Video Recognition

A simple, fast, and efficient object detector without FPN

Code for "LoFTR: Detector-Free Local Feature Matching with Transformers", CVPR 2021

Does Oversizing Improve Prosumer Profitability in a Flexibility Market? - A Sensitivity Analysis using PV-battery System

Official implementation of Monocular Quasi-Dense 3D Object Tracking

Evaluating saliency methods on artificial data with different background types

FactSeg: Foreground Activation Driven Small Object Semantic Segmentation in Large-Scale Remote Sensing Imagery (TGRS)

Course about deep learning for computer vision and graphics co-developed by YSDA and Skoltech.