symmetry_2022_code

This work is published in MDPI's Symmetry journal: Utilizing Language Models to Expand Vision-Based Commonsense Knowledge Graphs.

Tutorial

Expansion of a commonsense KG

Open the expansion script.
Adjust the varibales under name == "main" if needed.
The variable context_free_wpkg_path stores the path of the initial commonsense KG with the following format: "subj:pred:obj": weight.

NOTE: OpenAI's API key should be added as an environment variable.

After running, we should have a CSV file with this format in the output: "*_with_answers_*.csv". Each row contains a predicate with its query weight, find_subject and find_object questions, and also the results from GPT-3. Also a first try to extract concepts from the response, which will be refined in the next step.
Open the human evaluation preparation script. Set the workspace path to the workspace of the previous step. The outputs are two files. One a CSV file called triples_to_evaluate.csv that shows the triples going to be evaluated in CSV format. The other file is a JSONL format with generated sentences, called sentences_to_evaluate.jsonl , ready to be fed to Amazon mTurk as an input.

NOTE: As there are two error-prone processes of answer extraction from natural language and sentence generation base on them; it is necessary to vet the results to make sure there are no structurally-incorrect triples to evaluate. There is a grammar check done automatically to remove some of these issues.

For evaluation, we used Amazon mTurk through the AWS SageMaker's Ground Truth module. We created a labeling job with the following specifications and procedures:

6a. Creating a specific bucket under S3 and a specific folder underneath for each experiment. Placed the JSONL file under this folder. An empty subdirectory created for the evaluation results as well. 6b. Starting a new labeling job with the following specifications:

 - Manual data setup
 - "Input dataset location" pointing to the JSONL file and "Output dataset location" to the results subfolder, both from the 6a step.
 - Task: Text Classification (Single Label)
 - Worker types: Amazon Mechanical Turk
 - Appropriate timeout and task expiration time
 - Uncheck automated data labeling
 - Under additional configuration, select 3 workers.
 - Brief description:
     Using your own commonsense, decide if a sentence is correct or incorrect. Assume sentences with visual context, e.g. in the case of “It is likely to see cloud behind cow.”, imagine that you are in a field and you see cows. Then it makes sense to see clouds behind cows. Please ignore grammatical errors. Grammatical errors do not make a sentence incorrect.
 - Instructions:
     Using your own commonsense, decide if a sentence is correct or incorrect. 

     Assume sentences with visual context, e.g. in the case of “It is likely to see cloud behind cow.”, imagine that you are in a field and you see cows. Then it makes sense to see clouds behind cows.

     Please ignore grammatical errors. Grammatical errors do not make a sentence incorrect.

     More examples:

     It is NOT likely to see elephant on table. —> Correct

     It is likely to see food in plate. —> Correct
 - Options:
     - Correct
     - Incorrect

Once the labeling is done on Amazon mTurk, a manifest file is generated and placed under AWS S3's results folder specified in step 6a. Output manifest file is under manifests/output subfolder. Download the manifest file and place it under the same folder as the JSONL file generated in step 5. Then, open the manifest processing script. This method calculates the acuracy of the results and writes the manifest under .
To plot the output manifest results, you can use the plotting script. The output is a PDF file saved under the specified work path including the manifest output file.

Citation

If you find this project helpful, please consider citing our work.

@Article{rezaei2022expansion,
AUTHOR = {Rezaei, Navid and Reformat, Marek Z.},
TITLE = {Utilizing Language Models to Expand Vision-Based Commonsense Knowledge Graphs},
JOURNAL = {Symmetry},
VOLUME = {14},
YEAR = {2022},
NUMBER = {8},
ARTICLE-NUMBER = {1715},
URL = {https://www.mdpi.com/2073-8994/14/8/1715},
ISSN = {2073-8994},
DOI = {10.3390/sym14081715}
}

@InProceedings{rezaei2022contextual,
author="Rezaei, Navid
and Reformat, Marek Z.
and Yager, Ronald R.",
title="Generating Contextual Weighted Commonsense Knowledge Graphs",
booktitle="Information Processing and Management of Uncertainty in Knowledge-Based Systems",
year="2022",
publisher="Springer International Publishing",
address="Cham",
pages="593--606",
isbn="978-3-031-08971-8"
}

@InProceedings{rezaei2020wpkg,
author="Rezaei, Navid
and Reformat, Marek Z.
and Yager, Ronald R.",
title="Image-Based World-perceiving Knowledge Graph (WpKG) with Imprecision",
booktitle="Information Processing and Management of Uncertainty in Knowledge-Based Systems",
year="2020",
publisher="Springer International Publishing",
address="Cham",
pages="415--428",
isbn="978-3-030-50146-4"
}

navidre/lm_vision_kg_expansion

symmetry_2022_code

Tutorial

Expansion of a commonsense KG

Citation