FLD Corpus

This repository includes the released FLD corpora.

See the entry-point repository about the whole FLD project.

Available Corpora

• The English corpora introduced in the ICML paper are:
  • FLD (FLD.3)
  • FLD★(FLD.4)

Note that these corpora are version 2.0, which is detailed in the Appendix.H of our paper.

• The Japanese corpora, or JFLD, are described here.

How to use the corpora

First, install the datasets library:

pip install datasets

Then, you can load the FLD corpora as follows:

from datasets import load_dataset
FLD = load_dataset('hitachi-nlp/FLD.v2', name='default')
FLD_star = load_dataset('hitachi-nlp/FLD.v2', name='star')

What does the dataset example look like?

Concept

An example of deduction example in our dataset is conceptually illustrated in the figure below:

That is, given a set of facts and a hypothesis, a model must generate a proof sequence and determine an answer marker (proved, disproved, or unknown).

Schema

The actual schema can be viewed on the huggingface hub. The most important fields are:

• context (or facts in the later version of corpora): A set of facts.
• hypothesis: A hypothesis.
• proofs: Gold proofs. Each proof consists of a series of logical steps derived from the facts leading towards the hypothesis. Currently, for each example, we have at most one proof.
• world_assump_label: An answer, which is either PROVED, DISPROVED, or UNKNOWN.

Additionally, we have preprocessed fields as follows:

• prompt_serial: A serialized representation of the facts and the hypothesis.
• proof_serial: A serialized representation of the proof and answer.

To train or evaluate a Language Model (LM), one can take one of two approaches:

• Use prompt_serial as input and proof_serial as output. This will make the LM to generate both the proof and the answer.
• Use prompt_serial as input and world_assump_label as output. This will make the LM to generate only the answer.

Further, we have "logical formula" versions of the fields, such as prompt_serial_formula, which can be used to evaluate LLMs' pure logical reasoning capabilities within the domain of logical formulas, rather than natural language.