Are Pretrained Language Models Symbolic Reasoners over Knowledge?

This repository contains the code for "Are Pretrained Language Models Symbolic Reasoners over Knowledge?".

We provide a way to generate datasets that contain triples of the form "entity relation entity". These triplets follow different relational or logical rules. We then train BERT from scratch on this data and evaluate its ability to generalize and these rules. While we trained on BERT, this data generation process can in principle be used for testing any kind of language model. The possible rules are:

• equivalence
• symmetry
• inversion
• composition (and enhanced_composition)
• negation
• implication

Setup

We recommend running the following command in virtual environment:

pip install -r requirements.txt

Generating a dataset

To create your own dataset, navigate to scripts/RELATION. Here you can specify parameters (number of entities, number relations...) in datagen_config.py, which currently contain default parameters. Then run the following to create the data, here exemplified for symmetry: python3 -m scripts.symmetry.generate_data --dataset_name MY_DATASET_NAME

The dataset will be written to data/symmetry/datasets/MY_DATASET_NAME.

Training

To train the language model on a dataset, you run run_language_modeling.py as follows:

python3 -m scripts.run_language_modeling \
--relation RELATION_NAME
--dataset_name DIR_NAME_OF_SPECIFIC_DATASET
--anti SET_TRUE_IF_DATA_SHOULD_BE_EVALUATED_ON_ANTI_RULE_FACTS
--random SET_TRUE_IF_DATA_SHOULD_BE_EVALUATED_ON_RANDOM_FACTS

where

• relation is usually chosen from the above list of covered rules, e.g. "symmetry"; but you can come up with your own names as well
• DATA_DIR When creating a dataset, you will have to specify a name. This name is then needed here.

Optional parameters:

• anti tells the script to look for a json with the answers. Only include this if the data has anti-rule relations.
• random tells the script to look for a json with the answers. Only include if the rule has relations with random facts.
• numb_correct_answers indicates how many correct answers a given query of the form "subject relation [MASK] has. This influences our metric for evaluation accuracy.
• epochs: number of epochs. Default is 2000
• batch_size: we recommmend the default 1024
• learning_rate: default is 6e-5
• many more parameters that we didn't change but could be changed in a model like BERT

The resulting model is saved under outputs/model/RELATION/ and the events-file under outputs/runs/RELATION/.

Here is an exmaple command for symmetry:

python3 -m scripts.run_language_modeling \
--relation symmetry
--dataset_name StandardSym
--anti
--random

Probing BERT

We also provide our Notebooks and data for probing BERT for consistent predictions regarding symmetry & inversion. This can be found under probeBERT.

Citation

If you use this code, please cite:

@inproceedings{kassner-etal-2020-pretrained,
    title = "Are Pretrained Language Models Symbolic Reasoners over Knowledge?",
    author = {Kassner, Nora  and
      Krojer, Benno  and
      Sch{\"u}tze, Hinrich},
    booktitle = "Proceedings of the 24th Conference on Computational Natural Language Learning",
    month = nov,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.conll-1.45",
    pages = "552--564",
    abstract = "How can pretrained language models (PLMs) learn factual knowledge from the training set? We investigate the two most important mechanisms: reasoning and memorization. Prior work has attempted to quantify the number of facts PLMs learn, but we present, using synthetic data, the first study that investigates the causal relation between facts present in training and facts learned by the PLM. For reasoning, we show that PLMs seem to learn to apply some symbolic reasoning rules correctly but struggle with others, including two-hop reasoning. Further analysis suggests that even the application of learned reasoning rules is flawed. For memorization, we identify schema conformity (facts systematically supported by other facts) and frequency as key factors for its success.",
}