ArcaneQA is a generation-based KBQA model built upon the encoder-decoder framework. It features two mutually-boosting modules: dynamic program induction and dynamic contextualized encoding in a unified framework. For dynamic program induction, we unify the meaning representation in several KBQA datasets, and our annotations are also available in this repo; For dynamic contextualized encoding, we introduce a novel idea of applying pre-trained language models (PLMs) to KBQA.
This is the accompanying code & data for the paper "ArcaneQA: Dynamic Program Induction and Contextualized Encoding for Knowledge Base Question Answering" published at COLING 2022 (:trophy:Outstanding Paper Award).
2022-10-21
Numbers on WebQSP are fixed.
This repository is structured as follows:
ArcaneQA/
├─ configs/
│ ├─ train/: Configuration files for training
│ └── infer/: Configuration files for inference
├─ data/: Processed data files for different KBQA datasets
├─ ontology/: Processed Freebase ontology files
│ ├─ domain_dict: Mapping from a domain in Freebase Commons to all schema items in it
│ ├─ domain_info: Mapping from a schema item to a Freebase Commons domain it belongs to
│ ├─ fb_roles: Domain and range information for a relation (Note that here domain means a different thing from domains in Freebase Commons)
│ ├─ fb_types: Class hierarchy in Freebase
│ ├─ reverse_properties: Reverse properties in Freebase
│ └── commons/: Corresponding files of Freebase Commons subset
├─ answer_typing/: Answer typing results for GrailQA
├─ el_results/: Entity linking results
├─ vocabulary/: Preprocessed vocabulary
├─ cache/: Cached results for SPARQL queries, which are used to accelerate the experiments by caching many SPARQL query results offline
├─ saved_models/: Trained models
├─ predictions/: ArcaneQA's predictions on different KBQA benchmarks
├─ source/:
│ └── utils/:
│ ├─ bert_interface.py: Interface to BERT
│ ├─ my_pretrained_transformer_indexer.py: Class for indexer for PLMs, adapted from AllenNLP
│ ├─ my_pretrained_transformer_tokenizer.py: Class for tokenizer for PLMs, adapted from AllenNLP
│ ├─ logic_form_util.py: Tools related to logical forms, including the exact match checker for two logical forms
│ ├─ arcane_beam_search.py: Beam search for autoregressive decoding in ArcaneQA
│ ├─ sparql_executor.py: Sparql-related tools
│ ├─ kb_engine.py: Core functions for KB querying and constrained decoding
│ ├─ sparql_cache.py: Cache executions of different types of Sparql queries
│ ├─ run.py: Main function
│ ├─ evaluation.py: Evaluation script for WebQSP and GraphQuestion
│ ├─ webqsp_official_evaluation.py: Official evaluation script for WebQSP
│ ├─ arcane.py: ArcaneQA model class
│ └── arcane_reader.py: ArcaneQA dataset reader class
└── environment_arcane.yml: yml file for conda environment
We recommend you following the detailed instructions in our GrailQA Repo.
- Download cache files: https://1drv.ms/u/s!AuJiG47gLqTzoG9zmReRGBEhnqw-?e=iqNRQa
- Download saved models: https://1drv.ms/u/s!AuJiG47gLqTzoHGJ8UCnNvAx9M4g?e=lDoGRc
To train the model, run the following command
$ PYTHONHASHSEED=23 python run.py
train
[your_config_file_for_training]
--include-package
arcane
--include-package
arcane_reader
--include-package
utils.bert_interface
--include-package
utils.my_pretrained_transformer_indexer
--include-package
utils.my_pretrained_transformer_tokenizer
-s
[your_path_specified_for_training]
To make predictions using a trained model, run the following command
$ PYTHONHASHSEED=23 python run.py
predict
[path_for_trained_model]
[path_for_test_dataset]
--include-package
arcane
--include-package
arcane_reader
--include-package
utils.bert_interface
--include-package
utils.my_pretrained_transformer_indexer
--include-package
utils.my_pretrained_transformer_tokenizer
--use-dataset-reader
--predictor
seq2seq
-c
[your_config_file_for_inference]
--output-file
[output_file_name]
--cuda-device
0
We still need to process the output from AllenNLP's predict command into the data format we want. To this end, we can use process_predictions.py as follows,
$ python process_predictions.py
--dataset [grail_dev|grail_test|graphq|webq]
--prediction [output from AllenNLP's predict command]
--output [file name of the final output]
We give detailed instructions on how to get our evaluation results in Table 2.
For results on GrailQA, we followed the official instruction of GrailQA.
For results on WebQSP, we use the official evaluation script from WebQSP (i.e., eval/eval.py). Similar to RnG-KBQA, we adapt the input format using webqsp_official_evaluation.py. The official evaluation script considers all possible parses for each question (i.e., some questions in WebQSP are annotated with multiple parses). To eliminate the ambiguity in evaluation, we also implement our own evaluation script to only evaluate the top-1 parse for each question. The usage is as follows,
$ python evaluation.py
[path_to_the_dataset_file]
[path_to_the_prediction_file]
--fb_roles
ontology/fb_roles
--fb_types
ontology/fb_types
--reverse_properties
ontology/reverse_properties
The same script is also used to evaluate results on GraphQuestions.
To test the inference speed without caching, you may simply delete or rename files under /cache.
Please please consider citing the following BibTeX entry if you find our work helpful to your research.
@inproceedings{gu-su-2022-arcaneqa,
title = "{A}rcane{QA}: Dynamic Program Induction and Contextualized Encoding for Knowledge Base Question Answering",
author = "Gu, Yu and
Su, Yu",
booktitle = "Proceedings of the 29th International Conference on Computational Linguistics",
month = oct,
year = "2022",
address = "Gyeongju, Republic of Korea",
publisher = "International Committee on Computational Linguistics",
url = "https://aclanthology.org/2022.coling-1.148",
pages = "1718--1731",
abstract = "Question answering on knowledge bases (KBQA) poses a unique challenge for semantic parsing research due to two intertwined challenges: large search space and ambiguities in schema linking. Conventional ranking-based KBQA models, which rely on a candidate enumeration step to reduce the search space, struggle with flexibility in predicting complicated queries and have impractical running time. In this paper, we present ArcaneQA, a novel generation-based model that addresses both the large search space and the schema linking challenges in a unified framework with two mutually boosting ingredients: dynamic program induction for tackling the large search space and dynamic contextualized encoding for schema linking. Experimental results on multiple popular KBQA datasets demonstrate the highly competitive performance of ArcaneQA in both effectiveness and efficiency.",
}
Many thanks to the thoughtful comments from the OSU NLP Group and the anonymous reviewers.
We would also like to thank Ritam Dutt for first trying out our code in this repository and help us to identify some bugs during code cleaning at an early stage.
Please consider creating a new issue or directly sending an email to Yu Gu. Normally, we will respond to your questions within 1-2 days.