Explainable Zero-Shot Topic Extraction

D2K Lab

Last update: Dec 14, 2022

Related tags

Deep Learning ZeSTE

Overview

Zero-Shot Topic Extraction with Common-Sense Knowledge Graph

This repository contains the code for reproducing the results reported in the paper "Explainable Zero-Shot Topic Extraction with Common-Sense Knowledge Graph" (pdf) at the LDK 2021 Conference.

A user-friendly demo is available at: http://zeste.tools.eurecom.fr/

ZeSTE

Based on ConceptNet's common sense knowledge graph and embeddings, ZeSTE generates explainable predictions for a document topical category (e.g. politics, sports, video_games ..) without reliance on training data. The following is a high-level illustration of the approach:

API

ZeSTE can also be accessed via a RESTful API for easy deployment and use. For further information, please refer to the documentation: https://zeste.tools.eurecom.fr/doc

Dependencies

Before running any code in this repo, please install the following dependencies:

numpy
pandas
matplotlib
nltk
sklearn
tqdm
gensim

Code Overview

This repo is organized as follows:

generate_cache.py: this script processes the raw ConceptNet dump to produce cached files for each node in ConceptNet to accelerate the label neighborhood generation. It also transforms ConceptNet Numberbatch text file into a Gensim word embedding that we pickle for quick loading.
zeste.py: this is the main script for evaluation. It takes as argument the dataset to process as well as model configuration parameters such as neighborhood depth (see below). The results (classification report, confusion matrix, and classification metrics) are persisted into text files.
util.py: contains the functions that are used in zeste.py
label_mappings: contains the tab-separated mappings for the studied datasets.

Reproducing Results

1. Downloads

The two following files need to be downloaded to bypass the use of ConceptNet's web API: the dump of ConceptNet triplets, and the ConceptNet Numberbatch pre-computed word embeddings. You can download them from ConceptNet's and Numberbatch's repos, respectively.

# wget https://s3.amazonaws.com/conceptnet/downloads/2019/edges/conceptnet-assertions-5.7.0.csv.gz
# wget https://conceptnet.s3.amazonaws.com/downloads/2019/numberbatch/numberbatch-19.08.txt.gz
# gzip -d conceptnet-assertions-5.7.0.csv.gz
# gzip -d numberbatch-19.08.txt.gz

2. generate_cache.py

This script takes as input the two just-downloaded files and the cache path to where precomputed 1-hop label neighborhoods will be saved. This can take up to 7.2G of storage space.

usage: generate_cache.py [-h] [-cnp CONCEPTNET_ASSERTIONS_PATH] [-nbp CONCEPTNET_NUMBERBATCH_PATH] [-zcp ZESTE_CACHE_PATH]

Zero-Shot Topic Extraction

optional arguments:
  -h, --help            show this help message and exit
  -cnp CONCEPTNET_ASSERTIONS_PATH, --conceptnet_assertions_path CONCEPTNET_ASSERTIONS_PATH
                        Path to CSV file containing ConceptNet assertions dump
  -nbp CONCEPTNET_NUMBERBATCH_PATH, --conceptnet_numberbatch_path CONCEPTNET_NUMBERBATCH_PATH
                        Path to W2V file for ConceptNet Numberbatch
  -zcp ZESTE_CACHE_PATH, --zeste_cache_path ZESTE_CACHE_PATH
                        Path to the repository where the generated files will be saved

3. zeste.py

This script uses the precomputed 1-hop label neighborhoods to recursively generate label neighborhoods of any given depth (-d). It takes also as parameters the path to the dataset CSV file (which should have two columns: text and label). The rest of the arguments are for model experimentation.

usage: zeste.py [-h] [-cp CACHE_PATH] [-pp PREFETCH_PATH] [-nb NUMBERBATCH_PATH] [-dp DATASET_PATH] [-lm LABELS_MAPPING] [-rp RESULTS_PATH]
                [-d DEPTH] [-f FILTER] [-s {simple,compound,depth,harmonized}] [-ar ALLOWED_RELS]

Zero-Shot Topic Extraction

optional arguments:
  -h, --help            show this help message and exit
  -cp CACHE_PATH, --cache_path CACHE_PATH
                        Path to where the 1-hop word neighborhoods are cached
  -pp PREFETCH_PATH, --prefetch_path PREFETCH_PATH
                        Path to where the precomputed n-hop neighborhoods are cached
  -nb NUMBERBATCH_PATH, --numberbatch_path NUMBERBATCH_PATH
                        Path to the pickled Numberbatch
  -dp DATASET_PATH, --dataset_path DATASET_PATH
                        Path to the dataset to process
  -lm LABELS_MAPPING, --labels_mapping LABELS_MAPPING
                        Path to the mapping between the dataset labels and ZeSTE labels (multiword labels are comma-separated)
  -rp RESULTS_PATH, --results_path RESULTS_PATH
                        Path to the directory where to store the results
  -d DEPTH, --depth DEPTH
                        How many hops to generate the neighborhoods
  -f FILTER, --filter FILTER
                        Filtering method: top[N], top[P]%, thresh[T], all
  -s {simple,compound,depth,harmonized}, --similarity {simple,compound,depth,harmonized}
  -ar ALLOWED_RELS, --allowed_rels ALLOWED_RELS
                        Which relationships to use (comma-separated or all)

Cite this work

@InProceedings{harrando_et_al_zeste_2021,
  author ={Harrando, Ismail and Troncy, Rapha\"{e}l},
  title ={{Explainable Zero-Shot Topic Extraction Using a Common-Sense Knowledge Graph}},
  booktitle ={3rd Conference on Language, Data and Knowledge (LDK 2021)},
  pages ={17:1--17:15},
  year ={2021},
  volume ={93},
  publisher ={Schloss Dagstuhl -- Leibniz-Zentrum f{\"u}r Informatik},
  URL ={https://drops.dagstuhl.de/opus/volltexte/2021/14553},
  URN ={urn:nbn:de:0030-drops-145532},
  doi ={10.4230/OASIcs.LDK.2021.17},
}

Comments

Code is broken

Hi,

I'm trying to reproduce your results, I can generate the cache without issues. Then when I run the zeste.py script, it seems you're trying to match the labels in the dataset(I'm using the 20 news group) against the number batch, and they just don't match. The keys in the embeddings are prefixed with the language like in /c/en/potato, but the labels are not. So they never match, and you cannot retrieve any neighbour. Has this code ever worked?

opened by albertoandreottiATgmail 8
Code not working

As per your documentation i run the generate_cache.py file but not generate the output files in the data/zeste/ folder. Help me to solve this issues or share the latest documentation.

opened by radhakrishr36 9

[ICCV 2021] Official Pytorch implementation for Discriminative Region-based Multi-Label Zero-Shot Learning SOTA results on NUS-WIDE and OpenImages

Discriminative Region-based Multi-Label Zero-Shot Learning (ICCV 2021) [arXiv][Project page coming soon] Sanath Narayan*, Akshita Gupta*, Salman Kh

54 Nov 21, 2022

EMNLP 2021 Adapting Language Models for Zero-shot Learning by Meta-tuning on Dataset and Prompt Collections

Adapting Language Models for Zero-shot Learning by Meta-tuning on Dataset and Prompt Collections Ruiqi Zhong, Kristy Lee*, Zheng Zhang*, Dan Klein EMN

42 Nov 3, 2022

ZSL-KG is a general-purpose zero-shot learning framework with a novel transformer graph convolutional network (TrGCN) to learn class representation from common sense knowledge graphs.

ZSL-KG is a general-purpose zero-shot learning framework with a novel transformer graph convolutional network (TrGCN) to learn class representa

94 Nov 21, 2022

Original code for "Zero-Shot Domain Adaptation with a Physics Prior"

Zero-Shot Domain Adaptation with a Physics Prior [arXiv] [sup. material] - ICCV 2021 Oral paper, by Attila Lengyel, Sourav Garg, Michael Milford and J

40 Dec 21, 2022

The code for our paper "NSP-BERT: A Prompt-based Zero-Shot Learner Through an Original Pre-training Task —— Next Sentence Prediction"

201 Nov 21, 2022

Explainable Zero-Shot Topic Extraction

Related tags

Overview

Zero-Shot Topic Extraction with Common-Sense Knowledge Graph

ZeSTE

API

Dependencies

Code Overview

Reproducing Results

1. Downloads

2. generate_cache.py

3. zeste.py

Cite this work

You might also like...

[ICCV 2021] Official Pytorch implementation for Discriminative Region-based Multi-Label Zero-Shot Learning SOTA results on NUS-WIDE and OpenImages

[ICCV 2021] Official Pytorch implementation for Discriminative Region-based Multi-Label Zero-Shot Learning SOTA results on NUS-WIDE and OpenImages

EMNLP 2021 Adapting Language Models for Zero-shot Learning by Meta-tuning on Dataset and Prompt Collections

ZSL-KG is a general-purpose zero-shot learning framework with a novel transformer graph convolutional network (TrGCN) to learn class representation from common sense knowledge graphs.

Original code for "Zero-Shot Domain Adaptation with a Physics Prior"

The code for our paper "NSP-BERT: A Prompt-based Zero-Shot Learner Through an Original Pre-training Task —— Next Sentence Prediction"

GPU-accelerated PyTorch implementation of Zero-shot User Intent Detection via Capsule Neural Networks

IntraQ: Learning Synthetic Images with Intra-Class Heterogeneity for Zero-Shot Network Quantization

Rethinking of Pedestrian Attribute Recognition: A Reliable Evaluation under Zero-Shot Pedestrian Identity Setting

Comments

Code is broken

Code not working

Owner

D2K Lab

[CVPR 2021] Released code for Counterfactual Zero-Shot and Open-Set Visual Recognition

SC-GlowTTS: an Efficient Zero-Shot Multi-Speaker Text-To-Speech Model

code for CVPR paper Zero-shot Instance Segmentation

Zero-shot Synthesis with Group-Supervised Learning (ICLR 2021 paper)

Codes for ACL-IJCNLP 2021 Paper "Zero-shot Fact Verification by Claim Generation"

Official Pytorch Implementation of: "Semantic Diversity Learning for Zero-Shot Multi-label Classification"(2021) paper

Shared Attention for Multi-label Zero-shot Learning

PyTorch implementation of 1712.06087 "Zero-Shot" Super-Resolution using Deep Internal Learning

Zero-Shot Text-to-Image Generation VQGAN+CLIP Dockerized

An official implementation of "Exploiting a Joint Embedding Space for Generalized Zero-Shot Semantic Segmentation" (ICCV 2021) in PyTorch.