This repo contains tools and utilities to:
- Generate datasets of theories and assertions meant to test the logical reasoning capabilities of a model. For details see the paper Transformers as Soft Reasoners over Language.
- Run existing theories through a theorem proving engine to obtain labels.
It uses Problog as the theorem-proving engine to generate labels for theory-assertion examples.
For any features you might want, or issues, we welcome your pull requests! You can also file issues in this GitHub repository.
The latest dataset is available for download from the allenai dataset page.
The original RuleTaker theory generator and inference engine were written in Lisp. This repository is a more stable and re-engineered Python version of that original software. The theory generator here uses a declarative grammar that users can modify to generate different theories with different distributions. For the inference engine, we here use Problog, with probabilities 1 and 0 to denote T and F. The Problog reasoner generates identical results to the original Lisp inference engine, with one exception: The original RuleTaker datasets accidentally included a few non-stratified, inconsistent theories (for theories with negation). For these, the Problog reasoner (correctly) throws an error. We will be releasing the Problog-syntax version of the original RuleTaker theories shortly.
In addition, when Problog proves a queried fact, it also returns an SDD tree containing the proof. However, this initial release does not use those SDD trees, thus this software does not return a proof nor proof depth for the query assertions that Problog is able to prove.
We provide theory_generator.py to generate a dataset of theories and assertions. This helps generate a dataset with gold labels derived by running each theory through a theorem proving engine. The only supported engine currently is Problog. Anyone who wishes to use a different theorem prover needs to make appropriate code changes and plug the new engine as a choice to the theorem_prover command line argument which currently defaults to problog.
The inputs and output of this tool and a description of how to use it is described in this section.
- Grammar: A grammar to generate theories, in PCFG format. E.g.:
Statement -> Fact | Rule
Fact -> Positive '(' Attribute Entity ')' [0.3] | Positive '(' Relation Entity Entity ')' [0.7]
VFact -> VFactAttr [0.3] | VFactRel [0.7]
VFactAttr -> Polarity '(' Attribute Entity ')' [0.2] | Positive '(' Attribute Variable ')' [0.8]
VFactRel -> Polarity '(' Relation Entity Entity ')' [0.2] | Positive '(' Relation Variable Entity ')' [0.8]
VFacts -> '[' VFact ',' VFact ']' [0.5] | '[' VFact ']' [0.5]
Rule -> VFacts '->' VFact [1.0]
Entity -> 'cat' | 'dog' | 'bald eagle' | 'rabbit' | 'mouse' | 'tiger' | 'lion' | 'bear' | 'squirrel' | 'cow'
Variable -> 'X'
Attribute -> 'red' | 'blue' | 'green' | 'kind' | 'nice' | 'big' | 'cold' | 'young' | 'round' | 'rough'
Relation -> 'likes' | 'chases' | 'eats' | 'sees' | 'visits' | 'needs'
Polarity -> Positive [0.5] | Negative [0.5]
Positive -> '+'
Negative -> '-'
In order for the theory generation to work properly, there are some key assumptions with respect to what the grammar generates:
- A top-level statement can be a Fact or a Rule.
- A Fact contains a predicate with one or more arguments in prefix notation, along with an associated polarity: + / - . Sample Facts:
- Cat is blue:
+ ( blue 'cat' ) - Dog does not chase cat:
- ( chases dog cat)
- Cat is blue:
- A Rule contains an antecedant and a consequent and takes the form: -> . The antecedant is one or more Facts. The consequent is a single fact. Sample Rule:
- If something is blue then it is sad:
+ (blue X) -> + (sad X)
- If something is blue then it is sad:
- The entities can represent constants or variables in the grammar.
The theory generator needs to meaningfully associate the grammar nonterminals with the entities (constants and variables) and predicates. This mapping is specified via a config file which is described below.
- Config: A config file in JSON format that contains some hyperparameter values for the theory generator. Here's an example, decorated with comments not present in the JSON file:
{
"theory": {
"num_examples": 200, # The total number of examples (theories) that we want in the dataset.
# Tells the theory generator to attempt to create at least the specified no. of positive examples. No. of -ve examples will be num_examples - min_num_positive examples.
"min_num_positive_examples": 100,
# Config for each type of top-level statement that will be in the theory.
"statement_types_per_example": [ {
# This means that in each example (theory) there will be between 1 and 16 statements that are Facts.
# (A number randomly chosen in the interval [1, 16].
"start_symbol": "Fact",
"num_statements_range": [1, 16]
},
{
# This means that in each example (theory) there will be between 1 and 8 statements that are Rules.
# (A number randomly chosen in the interval [1, 8].
"start_symbol": "Rule",
"num_statements_range": [1, 8]
}
],
# Specifies the grammar nonterminals that map to predicates, variables and constants, respectively.
# If any of these is not applicable, it can be left out of the config.
"theorem_prover": {
# Nonterminals that generate the facts in the theory.
# If the grammar conforms to expected format, these nonterminals should generate statements that have
# the format: <+|-> ( <predicate> <argument1> [<argument2> <argument3> ...] )
# where predicate maps to one of the predicate_nonterminals noted below, and
# arguments are one of the variable_nonterminals or constant_nonterminals noted below.
"fact_nonterminals": ["Fact"],
# Nonterminals that generate the rules in the theory.
# If the grammar conforms to expected format, these nonterminals should generate statements that have
# the format: [ <fact1> , <fact2>, <fact3>, ... ] -> <consequent_fact>
# where fact1, fact2, fact3 and consequent_fact have the expected format for fact_nonterminals
# as noted above.
"rule_nonterminals": ["Rule"],
# Nonterminals that map to predicates in the theory's facts.
"predicate_nonterminals": ["Attribute", "Relation"],
# Nonterminals that map to arguments to the predicates (above). These are either constants or variables.
"variable_nonterminals": ["Variable"],
"constant_nonterminals": ["Entity"]
}
},
"assertion": {
# Each example in the dataset contains: a theory, an assertion to prove given the theory,
# and a label obtained by running the theory through a theorem proving engine. This specifies
# the grammar nonterminal that maps to an assertion statement.
"start_symbol": "Fact"
},
# Prefix to use while generating ids for examples in the dataset.
"example_id_prefix": "ruletaker-problog"
}
The tool provides an output JSONL file with one labeled theory-assertion example as JSON object per line. The object includes representations in three forms: the prefix notation logical forms generated by the grammar, the theory and assertion in natural languages (english sentences), and the logic program that was input to the theorem prover to generate gold label. The field names corresponding to these representations are logical_forms,english and logic_program respectively. Besides the gold label generated by Problog, also included are fields indicating the depth of the proof tree, and for examples where Problog comes up with an SDD representation, a formatted version of that structure as proof for the theory-assertion pair.
Sample output:
{
"json_class": "Example",
"theory_assertion_instance": {
"json_class": "TheoryAssertionInstance",
"theory": {
"json_class": "Theory",
"facts": [
{
"json_class": "Fact",
"polarity": "+",
"predicate": "likes",
"arguments": ["lion", "cow"],
"probability": 1.0
},
{
"json_class": "Fact",
"polarity": "+",
"predicate": "sees",
"arguments": ["mouse", "squirrel"],
"probability": 1.0
},
{
"json_class": "Fact",
"polarity": "+",
"predicate": "eats",
"arguments": ["tiger", "cat"],
"probability": 1.0
},
{
"json_class": "Fact",
"polarity": "+",
"predicate": "needs",
"arguments": ["cow", "dog"],
"probability": 1.0
},
{
"json_class": "Fact",
"polarity": "+",
"predicate": "eats",
"arguments": ["squirrel", "dog"],
"probability": 1.0
},
{
"json_class": "Fact",
"polarity": "+",
"predicate": "visits",
"arguments": ["bear", "dog"],
"probability": 1.0
}
],
"rules": [
{
"json_class": "Rule",
"lhs": [
{"json_class": "Fact", "polarity": "+", "predicate": "likes",
"arguments": ["X", "'bear'"], "probability": 1.0}
],
"rhs": {"json_class": "Fact", "polarity": "+", "predicate": "round",
"arguments": ["X"], "probability": 1.0},
"probability": 1.0
},
{
"json_class": "Rule",
"lhs": [
{"json_class": "Fact", "polarity": "+", "predicate": "blue", "arguments": ["X"], "probability": 1.0}],
"rhs": {"json_class": "Fact", "polarity": "+", "predicate": "young", "arguments": ["X"], "probability": 1.0},
"probability": 1.0}
,
{
"json_class": "Rule",
"lhs": [
{"json_class": "Fact", "polarity": "+", "predicate": "chases", "arguments": ["X", "'mouse'"], "probability": 1.0},
{"json_class": "Fact", "polarity": "+", "predicate": "kind", "arguments": ["X"], "probability": 1.0}
],
"rhs": {"json_class": "Fact", "polarity": "+", "predicate": "blue", "arguments": ["X"], "probability": 1.0},
"probability": 1.0
},
{
"json_class": "Rule",
"lhs": [
{"json_class": "Fact", "polarity": "+", "predicate": "cold", "arguments": ["X"], "probability": 1.0}
],
"rhs": {"json_class": "Fact", "polarity": "+", "predicate": "nice", "arguments": ["X"], "probability": 1.0},
"probability": 1.0
},
{
"json_class": "Rule",
"lhs": [
{"json_class": "Fact", "polarity": "+", "predicate": "nice", "arguments": ["X"], "probability": 1.0},
{"json_class": "Fact", "polarity": "+", "predicate": "cold", "arguments": ["X"], "probability": 1.0}
],
"rhs": {"json_class": "Fact", "polarity": "+", "predicate": "green", "arguments": ["X"], "probability": 1.0},
"probability": 1.0
},
{
"json_class": "Rule",
"lhs": [
{"json_class": "Fact", "polarity": "+", "predicate": "likes", "arguments": ["X", "'cow'"], "probability": 1.0}
],
"rhs": {"json_class": "Fact", "polarity": "+", "predicate": "cold", "arguments": ["X"], "probability": 1.0},
"probability": 1.0
},
{
"json_class": "Rule",
"lhs": [
{"json_class": "Fact", "polarity": "+", "predicate": "sees", "arguments": ["X", "'rabbit'"], "probability": 1.0}
],
"rhs": {"json_class": "Fact", "polarity": "+", "predicate": "red", "arguments": ["X"], "probability": 1.0},
"probability": 1.0
}
]
},
"assertion": {
"json_class": "Fact",
"polarity": "+",
"predicate": "nice",
"arguments": ["'lion'"],
"probability": 1.0
},
"label": true,
"exception": null,
"min_proof_depth": 2,
"proof": "1: atom(identifier=0, probability=1.0, group=None, name=likes('lion','cow')
source=None) |
2: atom(identifier=(63, ('lion',) {{}}, 0), probability=1.0, group=(63, ('lion',) {{}}),
name=choice(63,0,cold('lion'),'lion'), source=None) |
3: conj(children=(1, 2), name=None) |
4: atom(identifier=(42, ('lion',) {{}}, 0), probability=1.0, group=(42, ('lion',) {{}}),
name=choice(42,0,nice('lion'),'lion'), source=None) |
5: conj(children=(3, 4), name=nice('lion')) |
Queries : nice('lion') : 5 [query]"
},
"logical_forms": {
"json_class": "TheoryAssertionRepresentation",
"theory_statements": [
"+ ( visits 'bear' 'dog' )",
"+ ( needs 'cow' 'dog' )",
"+ ( eats 'tiger' 'cat' )",
"+ ( eats 'squirrel' 'dog' )",
"+ ( sees 'mouse' 'squirrel' )",
"[ + ( cold X ) ] -> + ( nice X )",
"[ + ( nice X ) , + ( cold X ) ] -> + ( green X )",
"[ + ( likes X 'cow' ) ] -> + ( cold X )",
"+ ( likes 'lion' 'cow' )",
"[ + ( likes X 'bear' ) ] -> + ( round X )",
"[ + ( blue X ) ] -> + ( young X )",
"[ + ( sees X 'rabbit' ) ] -> + ( red X )",
"[ + ( chases X 'mouse' ) , + ( kind X ) ] -> + ( blue X )"
],
"assertion_statement": "+ ( nice 'lion' )"
},
"english": {
"json_class": "TheoryAssertionRepresentation",
"theory_statements": [
"Lion likes cow.",
"Mouse sees squirrel.",
"Tiger eats cat.",
"Cow needs dog.",
"Squirrel eats dog.",
"Bear visits dog.",
"If X likes bear then X is round.",
"If X is blue then X is young.",
"If X chases mouse and X is kind then X is blue.",
"If X is cold then X is nice.",
"If X is nice and X is cold then X is green.",
"If X likes cow then X is cold.",
"If X sees rabbit then X is red."
],
"assertion_statement": "Lion is nice."
},
"logic_program": {
"problog": {
"json_class": "TheoryAssertionRepresentation",
"theory_statements": [
"1.0::likes('lion', 'cow').",
"1.0::sees('mouse', 'squirrel').",
"1.0::eats('tiger', 'cat').",
"1.0::needs('cow', 'dog').",
"1.0::eats('squirrel', 'dog').",
"1.0::visits('bear', 'dog').",
"1.0::round(X) :- likes(X, 'bear').",
"1.0::young(X) :- blue(X).",
"1.0::blue(X) :- chases(X, 'mouse'), kind(X).",
"1.0::nice(X) :- cold(X).",
"1.0::green(X) :- nice(X), cold(X).",
"1.0::cold(X) :- likes(X, 'cow').",
"1.0::red(X) :- sees(X, 'rabbit')."
],
"assertion_statement": "query(nice('lion'))."
}
}
}
We provide theory_label_generator.py to run existing theories and assertions through a theorem proving engine to obtain labels. You can use this to evaluate any theory dataset as long as facts and rules in the theory can be provided in the expected format. The tool takes the given theories with Facts and Rules, and assertions and runs the theorem prover to obtain a label for the theory assertion pair. If gold labels have already been provided, it computes accuracy metrics against them.
Input can be provided in one of two formats.
File in JSONL format, i.e., a JSON object per line with structure as defined by the TheoryAssertionRepresentationWithLabel class:
{
"theory_statements": List of theory statement logical forms in prefix notation.
"assertion_statement": String containing the assertion in prefix notation.
"label": Optional Boolean label field. If set with value true, then the input object has a gold label specified, to compare accuracy against. Output object will contain a boolean value as returned from the theorem prover or None if the theorem prover threw an exception.
"exception": Optional String field that is a placeholder for the output object use to report exceptions if any, from the theorem prover.
}
Sample input:
{
"json_class": "TheoryAssertionRepresentationWithLabel",
"theory_statements": [
"+ ( nice 'Harry' )",
"[ + ( red X ) ] -> + ( red X )",
"+ ( big 'Charlie' )",
"+ ( white 'Erin' )",
"+ ( green 'Bob' ), [ + ( white X ) , + ( red X ) ] -> + ( nice X )",
"[ + ( red X ) ] -> + ( cold X )",
"[ + ( rough X ) ] -> + ( nice X )",
"[ + ( young X ) ] -> + ( kind X )",
"[ + ( nice X ) ] -> + ( quiet X )",
"+ ( blue 'Fiona' )",
"[ + ( round X ) ] -> + ( young X )",
"+ ( green 'Erin' )", "+ ( round 'Harry' )",
"[ + ( nice X ) ] -> + ( green X )",
"+ ( green 'Gary' )",
"assertion_statement": "+ ( kind 'Fiona' )",
"label": false,
"exception": None
}
This is primarily for AI2's internal use. This is the RuleTaker legacy JSONL format. Sample JSONL snippet is below (this only includes fields that are relevant to the tool).
{
"id":"AttNoneg-D3-319",
...
"triples":{
"triple1":{
"text":"Bob is cold.",
"representation":"(\"Bob\" \"is\" \"cold\" \"+\")"
},
"triple2":{
"text":"Erin is nice.",
"representation":"(\"Erin\" \"is\" \"nice\" \"+\")"
},
"triple3":{
"text":"Gary is nice.",
"representation":"(\"Gary\" \"is\" \"nice\" \"+\")"
},
"triple4":{
"text":"Harry is blue.",
"representation":"(\"Harry\" \"is\" \"blue\" \"+\")"
}
},
"rules":{
"rule1":{
"text":"Blue people are furry.",
"representation":"(((\"someone\" \"is\" \"blue\" \"+\")) -> (\"someone\" \"is\" \"furry\" \"+\"))"
},
"rule2":{
"text":"Nice people are furry.",
"representation":"(((\"someone\" \"is\" \"nice\" \"+\")) -> (\"someone\" \"is\" \"furry\" \"+\"))"
},
"rule3":{
"text":"Blue, big people are nice.",
"representation":"(((\"someone\" \"is\" \"blue\" \"+\") (\"someone\" \"is\" \"big\" \"+\"))
-> (\"someone\" \"is\" \"nice\" \"+\"))"
},
"rule4":{
"text":"If someone is cold then they are quiet.",
"representation":"(((\"someone\" \"is\" \"cold\" \"+\"))
-> (\"someone\" \"is\" \"quiet\" \"+\"))"
}
},
"questions":{
"Q1":{
"question":"Erin is nice.",
"answer":true,
...
"representation":"(\"Erin\" \"is\" \"nice\" \"+\")"
},
"Q2":{
"question":"Gary is not nice.",
"answer":false,
...
"representation":"(\"Gary\" \"is\" \"nice\" \"-\")"
},
"Q3":{
"question":"Gary is furry.",
"answer":true,
...
"representation":"(\"Gary\" \"is\" \"furry\" \"+\")"
}
}
}
The output object will take the structure of the Example class, which is the same as the output structure of examples generated by theory_generator.py. This structure is documented in detailed along with an output snippet Theory Generator section under Output description.
To run any of the tools in the repo, first create a python environment with the necessary dependencies:
virtualenv venv
source venv/bin/activate
pip install -r requirements.txt
pip install --upgrade --force-reinstall --no-binary :all: --no-deps pysdd
Then use the following command line to run the theory generator.
Currently we are using Problog as the underlying theorem proving engine. That is the recommended engine to use.
Once you have prepared the Python environment as described, you can run the theory generator as follows-
python theory_generator.py \
--grammar <grammar_cfg_file_path> \
--config-json <theory_config_json_file_path> \
--op-theory-jsonl <theory_op_jsonl_file_path>
E.g.:
python theory_generator.py \
--grammar grammars_and_config/grammars/ruletaker_grammar_theory1.txt \
--config-json grammars_and_config/config/ruletaker_theory_generator_config_theory1.json \
--op-theory-jsonl ruletaker_theory1_examples.jsonl
This will take a minute to run. Once it completes, you should see a message that looks like:
Generated 1000 examples.
No. with True label: 432
No. with False label: 568
After preparing the Python environment as described earlier, use the following command line to run theory_label_generator.py.
Currently we are using Problog as the underlying theorem proving engine and that is the only supported engine.
python theory_label_generator.py \
--input-file <theory_assertion_representations.jsonl> | <ruletaker_theory_assertions.jsonl>
--output-file <labeled_output.jsonl>
[--input-format current|legacy]
[--report-metrics]
Note that --report-metrics is an optional flag to get accuracy and timing related metrics at the end of the run. If the flag is specified, and your input dataset includes gold labels, you will see metrics reported as follows:
Total no. of examples: 150
No. true: 75
No. correct: 73
No. of exceptions: 6
No. correct with exceptions: 4
No. incorrect without exception: 0
No. false: 75
No. correct: 72
No. of exceptions: 7
No. correct with exceptions: 4
No. incorrect without exception: 0
Total no. correct: 145
Total no. with exceptions: 13
Accuracy: 96.66666666666667
Failure Breakdown by Exception:
problog.engine_stack.NegativeCycle: 5
Average theorem proving time per example: 0.025506666666666667 secs
You can run unit tests using pytest as below:
pytest tests.py
This code has been tested and runs on Python 3.7.7 on the following platforms:
- Mac OS 10.13, 10.14
- Ubuntu 16.04.6 LTS
If you need to directly contact someone for any reason where a GitHub pull request or an issue is not appropriate, you can email oyvindt@allenai.org.
Copyright 2020 Allen Institute for AI
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.