rfernand/basic_sentence_transforms

Dataset Card for Active/Passive/Logical Transforms

Dataset Summary

This dataset is a synthetic dataset containing structure-to-structure transformation tasks between English sentences in 3 forms: active, passive, and logical. The dataset also includes several tree-transformation diagnostic/warm-up tasks.

Supported Tasks and Leaderboards

[TBD]

Languages

All data is in English.

Dataset Structure

The dataset consists of several subsets, or tasks. Each task contains a train split, a validation split, and a test split, with most tasks also containing two out-of-distruction splits (one for new adjectives and one for longer adjective phrases).

Each sample in a split contains a source string, a target string, and 0-2 annotation strings.

Dataset Subsets (Tasks)

The dataset consists of diagnostic/warm-up tasks and core tasks. The core tasks represent the translation of English sentences between the active, passive, and logical forms.

The 12 diagnostic/warm-up tasks are:

- car_cdr_cons              (small phrase translation tasks that require only: CAR, CDR, or CAR+CDR+CONS operations)
- car_cdr_cons_tuc          (same task as car_cdr_cons, but requires mapping lowercase fillers to their uppercase tokens)
- car_cdr_rcons             (same task as car_cdr_cons, but the CONS samples have their left/right children swapped)
- car_cdr_rcons_tuc         (same task as car_cdr_rcons, but requires mapping lowercase fillers to their uppercase tokens)
- car_cdr_seq               (each samples requires 1-4 combinations of CAR and CDR, as identified by the root filler oken)
- car_cdr_seq_40k           (same task as car_cdr_seq, but train samples increased from 10K to 40K)
- car_cdr_seq_tuc           (same task as car_cdr_seq, but requires mapping lowercase fillers to their uppercase tokens)
- car_cdr_seq_40k_tuc       (same task as car_cdr_seq_tuc, but train samples increased from 10K to 40K)
- car_cdr_seq_path          (similiar to car_cdr_seq, but each needed operation in represented as a node in the left child of the root)
- car_cdr_seq_path_40k      (same task as car_cdr_seq_path, but train samples increased from 10K to 40K)
- car_cdr_seq_path_40k_tuc  (same task as car_cdr_seq_path_40k, but requires mapping lowercase fillers to their uppercase tokens)
- car_cdr_seq_path_tuc      (same task as car_cdr_seq_path, but requires mapping lowercase fillers to their uppercase tokens)

There are 22 core tasks are:

- active_active_stb         (active sentence translation, from sentence to parenthesized tree form, both directions)
- active_active_stb_40k     (same task as active_active_stb, but train samples increased from 10K to 40K)

- active_logical_ssb        (active to logical sentence translation, in both directions)
- active_logical_ssb_40k    (same task as active_logical_ssb, but train samples increased from 10K to 40K)
- active_logical_ttb        (active to logical tree translation, in both directions)
- active_logical_ttb_40k    (same task as active_logical_ttb, but train samples increased from 10K to 40K)

- active_passive_ssb        (active to passive sentence translation, in both directions)
- active_passive_ssb_40k    (same task as active_passive_ssb, but train samples increased from 10K to 40K)
- active_passive_ttb        (active to passive tree translation, in both directions)
- active_passive_ttb_40k    (same task as active_passive_ttb, but train samples increased from 10K to 40K)

- actpass_logical_ss        (mixture of active to logical and passive to logical sentence translations, single direction)
- actpass_logical_ss_40k    (same task as actpass_logical_ss, but train samples increased from 10K to 40K)
- actpass_logical_tt        (mixture of active to logical and passive to logical tree translations, single direction)
- actpass_logical_tt_40k    (same task as actpass_logical_tt, but train samples increased from 10K to 40K)

- logical_logical_stb       (logical form sentence translation, from sentence to parenthesized tree form, both directions)
- logical_logical_stb_40k   (same task as logical_logical_stb, but train samples increased from 10K to 40K)

- passive_logical_ssb       (passive to logical sentence translation, in both directions)
- passive_logical_ssb_40k   (same task as passive_logical_ssb, but train samples increased from 10K to 40K)
- passive_logical_ttb       (passive to logical tree translation, in both directions)
- passive_logical_ttb_40k   (same task as passive_logical_ttb, but train samples increased from 10K to 40K)

- passive_passive_stb       (passive sentence translation, from sentence to parenthesized tree form, both directions)
- passive_passive_stb_40k   (same task as passive_passive_stb, but train samples increased from 10K to 40K)

Data Splits

Most tasks have the following splits:

• train
• validation
• test
• ood_new
• ood_long
• ood_all

Here is a table showing how the number of examples varies by split (for most tasks):

Dataset Split	Number of Instances in Split
train	10,000
validation	1,250
test	1,250
ood_new	1,250
ood_long	1,250
ood_all	1,250

Data Instances

For each sample, there is source and target string. Source and target string are either plain text, or a parenthesized version of a tree, depending on the task.

Here is an example from the train split of the active_passive_ttb task:

{
  'source': '( S ( NP ( DET his ) ( AP ( N cat ) ) ) ( VP ( V discovered ) ( NP ( DET the ) ( AP ( ADJ blue ) ( AP ( N priest ) ) ) ) ) )',
  'target': '( S ( NP ( DET the ) ( AP ( ADJ blue ) ( AP ( N priest ) ) ) ) ( VP ( AUXPS was ) ( VPPS ( V discovered ) ( PPPS ( PPS by ) ( NP ( DET his ) ( AP ( N cat ) ) ) ) ) ) )',
  'direction': 'forward'
}

Data Fields

• source: the string denoting the sequence or tree structure to be translated
• target: the string denoting the gold (aka label) sequence or tree structure

Optional annotation fields (their presence varies by task):

• direction: describes the direction of the translation (forward, backward), relative to the task name
• count : a string denoting the count of symbolic operations needed (e.g., "s3") to translate the source to the target
• class : a string denoting the type of translation needed

Dataset Creation

Curation Rationale

We wanted a dataset comprised of relatively simple English active/passive/logical form translations, where we could focus on two types of out of distribution generalization: longer source sequences and new adjectives.

Source Data

[N/A]

Initial Data Collection and Normalization

[N/A]

Who are the source language producers?

The dataset by generated from templates designed by Paul Smolensky and Roland Fernandez.

Annotations

Besides the source and target structured sequences, some of the subsets (tasks) contain 1-2 additional columns that describe the category and tree depth of each sample.

Annotation process

The annotation columns were generated from the each sample template and source sequence.

Who are the annotators?

[N/A]

Personal and Sensitive Information

No names or other sensitive information are included in the data.

Considerations for Using the Data

Social Impact of Dataset

The purpose of this dataset is to help develop models that can translated structured data from one form to another, in a way that generalizes to out of distribution adjective values and lengths.

Discussion of Biases

[TBD]

Other Known Limitations

[TBD]

Additional Information

The internal name of this dataset is nc_pat.

Dataset Curators

The dataset by generated from templates designed by Paul Smolensky and Roland Fernandez.

Licensing Information

This dataset is released under the Permissive 2.0 license.

Citation Information

[TBD]

Contributions

Thanks to The Neurocompositional AI group at Microsoft Research for creating and adding this dataset.