XAI human-in-the-loop information extraction framework
Project description
POTATO: exPlainable infOrmation exTrAcTion framewOrk
POTATO is a human-in-the-loop XAI framework for extracting and evaluating interpretable graph features for any classification problem
Built systems
To get started with rule-systems we provide rule-based features prebuilt with POTATO on different datasets (e.g. our paper Offensive text detection on English Twitter with deep learning models and rule-based systems for the HASOC2021 shared task). If you are interested in that, you can go under features/ for more info!
Install and Quick Start
Setup
The tool is heavily dependent upon the tuw-nlp repository. You can install tuw-nlp with pip:
pip install tuw-nlp
Then follow the instructions to setup the package.
Then install POTATO from pip:
pip install xpotato
Or you can install it from source:
pip install -e .
Usage
First import packages from potato:
from xpotato.dataset.dataset import Dataset
from xpotato.models.trainer import GraphTrainer
Initialize the dataset you want to classify:
sentences = [("fuck absolutely everything about today.", "HOF"),
("I just made food and I'm making myself sick to my stomach. Lol, wtf is this shit", "HOF"),
("RT [USER]: America is the most fucked up country [URL]", "HOF"),
("you'd be blind to not see the heart eyes i have for you.", "NOT"),
("It's hard for me to give a fuck now", "HOF"),
("tell me everything", "NOT"),
("Bitch YES [URL]", "HOF"),
("Eight people a minute....", "NOT"),
("RT [USER]: im not fine, i need you", "NOT"),
("Holy shit.. 3 months and I'll be in Italy", "HOF"),
("Now I do what I want 🤪", "NOT"),
("[USER] you'd immediately stop", "NOT"),
("Just... shut the fuck up", "HOF"),
("RT [USER]: ohhhh shit a [USER] [URL]", "HOF"),
("all i want is for yara to survive tonight", "NOT"),
("fuck them", "HOF")]
Initialize the dataset and also provide a label encoding. Then parse the sentences into graphs. Currently we provide three types of graphs: ud, fourlang, amr.
dataset = Dataset(sentences, label_vocab={"NOT":0, "HOF": 1})
dataset.set_graphs(dataset.parse_graphs(graph_format="ud"))
Rules
If the dataset is prepared and the graphs are parsed, we can write rules to match labels. We can write rules either manually or extract them automatically (POTATO also provides a frontend that tries to do both).
The simplest rule would be just a node in the graph:
#the syntax of the rules is List[List[rules that we want to match], List[rules that shouldn't be in the matched graphs], Label of the rule]
rule_to_match = [[["(u_1 / fuck)"], [], "HOF"]]
Init the rule matcher:
from xpotato.graph_extractor.extract import FeatureEvaluator
evaluator = FeatureEvaluator()
Match the rules in the dataset:
#match single feature
df = dataset.to_dataframe()
evaluator.match_features(df, rule_to_match)
The function will return a dataframe with the matched instances:
Sentence | Predicted label | Matched rule | |
---|---|---|---|
0 | fuck absolutely everything about today. | HOF | [['(u_1 / fuck)'], [], 'HOF'] |
1 | I just made food and I'm making myself sick to my stomach. Lol, wtf is this shit | ||
2 | RT [USER]: America is the most fucked up country [URL] | ||
3 | you'd be blind to not see the heart eyes i have for you. | ||
4 | It's hard for me to give a fuck now | HOF | [['(u_1 / fuck)'], [], 'HOF'] |
5 | tell me everything | ||
6 | Bitch YES [URL] | ||
7 | Eight people a minute.... | ||
8 | RT [USER]: im not fine, i need you | ||
9 | Holy shit.. 3 months and I'll be in Italy | ||
10 | Now I do what I want 🤪 | ||
11 | [USER] you'd immediately stop | ||
12 | Just... shut the fuck up | HOF | [['(u_1 / fuck)'], [], 'HOF'] |
13 | RT [USER]: ohhhh shit a [USER] [URL] | ||
14 | all i want is for yara to survive tonight | ||
15 | fuck them | HOF | [['(u_1 / fuck)'], [], 'HOF'] |
One of the core features of our tool is that we are also able to match subgraphs:
#match a simple graph feature
evaluator.match_features(df, [[["(u_1 / fuck :obj (u_2 / everything))"], [], "HOF"]])
This will only return one match instead of three:
Sentence | Predicted label | Matched rule | |
---|---|---|---|
0 | fuck absolutely everything about today. | HOF | [['(u_1 / fuck :obj (u_2 / everything))'], [], 'HOF'] |
1 | I just made food and I'm making myself sick to my stomach. Lol, wtf is this shit | ||
2 | RT [USER]: America is the most fucked up country [URL] | ||
3 | you'd be blind to not see the heart eyes i have for you. | ||
4 | It's hard for me to give a fuck now | ||
5 | tell me everything | ||
6 | Bitch YES [URL] | ||
7 | Eight people a minute.... | ||
8 | RT [USER]: im not fine, i need you | ||
9 | Holy shit.. 3 months and I'll be in Italy | ||
10 | Now I do what I want 🤪 | ||
11 | [USER] you'd immediately stop | ||
12 | Just... shut the fuck up | ||
13 | RT [USER]: ohhhh shit a [USER] [URL] | ||
14 | all i want is for yara to survive tonight | ||
15 | fuck them |
We can also add negated features that we don't want to match (this won't match the first row where 'absolutely' is present):
#match a simple graph feature
evaluator.match_features(df, [[["(u_1 / fuck)"], ["(u_2 / absolutely)"], "HOF"]])
Sentence | Predicted label | Matched rule | |
---|---|---|---|
0 | fuck absolutely everything about today. | ||
1 | I just made food and I'm making myself sick to my stomach. Lol, wtf is this shit | ||
2 | RT [USER]: America is the most fucked up country [URL] | ||
3 | you'd be blind to not see the heart eyes i have for you. | ||
4 | It's hard for me to give a fuck now | HOF | [['(u_1 / fuck)'], ['(u_2 / absolutely)'], 'HOF'] |
5 | tell me everything | ||
6 | Bitch YES [URL] | ||
7 | Eight people a minute.... | ||
8 | RT [USER]: im not fine, i need you | ||
9 | Holy shit.. 3 months and I'll be in Italy | ||
10 | Now I do what I want 🤪 | ||
11 | [USER] you'd immediately stop | ||
12 | Just... shut the fuck up | HOF | [['(u_1 / fuck)'], ['(u_2 / absolutely)'], 'HOF'] |
13 | RT [USER]: ohhhh shit a [USER] [URL] | ||
14 | all i want is for yara to survive tonight | ||
15 | fuck them | HOF | [['(u_1 / fuck)'], ['(u_2 / absolutely)'], 'HOF'] |
If we don't want to specify nodes, regex can also be used in place of the node and edge-names:
#regex can be used to match any node (this will match instances where 'fuck' is connected to any node with 'obj' edge)
evaluator.match_features(df, [[["(u_1 / fuck :obj (u_2 / .*))"], [], "HOF"]])
Sentence | Predicted label | Matched rule | |
---|---|---|---|
0 | fuck absolutely everything about today. | HOF | [['(u_1 / fuck :obj (u_2 / .*))'], [], 'HOF'] |
1 | I just made food and I'm making myself sick to my stomach. Lol, wtf is this shit | ||
2 | RT [USER]: America is the most fucked up country [URL] | ||
3 | you'd be blind to not see the heart eyes i have for you. | ||
4 | It's hard for me to give a fuck now | ||
5 | tell me everything | ||
6 | Bitch YES [URL] | ||
7 | Eight people a minute.... | ||
8 | RT [USER]: im not fine, i need you | ||
9 | Holy shit.. 3 months and I'll be in Italy | ||
10 | Now I do what I want 🤪 | ||
11 | [USER] you'd immediately stop | ||
12 | Just... shut the fuck up | ||
13 | RT [USER]: ohhhh shit a [USER] [URL] | ||
14 | all i want is for yara to survive tonight | ||
15 | fuck them | HOF | [['(u_1 / fuck :obj (u_2 / .*))'], [], 'HOF'] |
We can also train regex rules from a training data, this will automatically replace regex '.*' with nodes that are 'good enough' statistically based on the provided dataframe.
#regex can be used to match any node (this will match instances where 'fuck' is connected to any node with 'obj' edge)
evaluator.train_feature("HOF", "(u_1 / fuck :obj (u_2 / .*))", df)
This will return '(u_1 / fuck :obj (u_2 / everything|they))'] (replaced '.*' with everything and they)
Learning rules
To extract rules automatically, train the dataset with graph features and rank them based on relevancy:
df = dataset.to_dataframe()
trainer = GraphTrainer(df)
#extract features
features = trainer.prepare_and_train()
from sklearn.model_selection import train_test_split
train, val = train_test_split(df, test_size=0.2, random_state=1234)
#save train and validation, this is important for the frontend to work
train.to_pickle("train_dataset")
val.to_pickle("val_dataset")
import json
#also save the ranked features
with open("features.json", "w+") as f:
json.dump(features, f)
You can also save the parsed graphs for evaluation or for caching:
import pickle
with open("graphs.pickle", "wb") as f:
pickle.dump(val.graph, f)
To see the code you can check the jupyter notebook under notebooks/examples.ipynb
Frontend
If the DataFrame is ready with the parsed graphs, the UI can be started to inspect the extracted rules and modify them. The frontend is a streamlit app, the simplest way of starting it is (the training and the validation dataset must be provided):
streamlit run frontend/app.py -- -t notebooks/train_dataset -v notebooks/val_dataset -g ud
it can be also started with the extracted features:
streamlit run frontend/app.py -- -t notebooks/train_dataset -v notebooks/val_dataset -g ud -sr notebooks/features.json
if you already used the UI and extracted the features manually and you want to load it, you can run:
streamlit run frontend/app.py -- -t notebooks/train_dataset -v notebooks/val_dataset -g ud -sr notebooks/features.json -hr notebooks/manual_features.json
Unsupervised mode
If labels are not or just partially provided, the frontend can be started also in unsupervised mode, where the user can annotate a few examples at the start, then the system gradually offers rules based on the provided examples.
Dataset without labels can be initialized with:
sentences = [("fuck absolutely everything about today.", ""),
("I just made food and I'm making myself sick to my stomach. Lol, wtf is this shit", ""),
("RT [USER]: America is the most fucked up country [URL]", ""),
("you'd be blind to not see the heart eyes i have for you.", ""),
("It's hard for me to give a fuck now", ""),
("tell me everything", ""),
("Bitch YES [URL]", ""),
("Eight people a minute....", ""),
("RT [USER]: im not fine, i need you", ""),
("Holy shit.. 3 months and I'll be in Italy", ""),
("Now I do what I want 🤪", ""),
("[USER] you'd immediately stop", ""),
("Just... shut the fuck up", ""),
("RT [USER]: ohhhh shit a [USER] [URL]", ""),
("all i want is for yara to survive tonight", ""),
("fuck them", "")]
Then, the frontend can be started:
streamlit run frontend/app.py -- -t notebooks/unsupervised_dataset -g ud -m unsupervised
Evaluate
If you have the features ready and you want to evaluate them on a test set, you can run:
python scripts/evaluate.py -t ud -f notebooks/features.json -d notebooks/val_dataset
The result will be a csv file with the labels and the matched rules:
Sentence | Predicted label | Matched rule | ||
---|---|---|---|---|
0 | RT [USER]: ohhhh shit a [USER] [URL] | HOF | ['(u_48 / shit)'] | |
1 | [USER] you'd immediately stop | HOF | ['(u_40 / user :punct (u_42 / LSB))'] | |
2 | fuck absolutely everything about today. | HOF | ['(u_1 / fuck)'] |
Contributing
We welcome all contributions! Please fork this repository and create a branch for your modifications. We suggest getting in touch with us first, by opening an issue or by writing an email to Adam Kovacs or Gabor Recski at firstname.lastname@tuwien.ac.at
Citing
License
MIT license
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.