Paraphrase Sense-Tagged Sentences (PSTS)

This directory holds the PSTS resource, which contains sentences pertaining to paraphrase pairs from PPDB. For example, sentences containing bug in its sense as a paraphrase of error include:

• Report a **bug** .
• This **bug** code is normally caused by a parity error in the system memory .

and sentences containing bug in its sense as a paraphrase of microbe include:

• You know, maybe it's that **bug** that's been going around.
• We all picked up a little **bug** from India.

If you use this resource in your work, please cite this paper:

@article{doi:10.1162/tacl-a-00295,
  author       = {Anne Cocos and Chris Callison-Burch}, 
  title        = {Paraphrase-Sense-Tagged Sentences},
  journal      = {Transactions of the Association for Computational Linguistics},
  year         = {2019},
  volume       = {7},
  number       = {},
  pages        = {714-728},
  doi          = {10.1162/tacl\_a\_00295}
}

Contents

|-- README.md  -- This file
|-- src
    |-- psts.py  -- Code to connect to and query mongodb instance
    |-- download_db.sh  -- Command to download psts-small or psts-all dump (run this first)
    |-- build_db_from_dump.sh  -- Code to build mongodb instance (run this second)
    |-- deploy.sh  -- Code to deploy mongodb instance locally (run once built, each time you want to use `psts.py`)
    |-- stop-db.sh -- Code to stop mongodb instance locally
|-- small
    |-- psts-small.zip -- zipped dump of psts-small (to be downloaded by download_db.sh)
    |-- db/  -- mongodb instance of psts-small (filled after running build_db_from_dump.sh)
|-- all
    |-- psts-all.zip  -- zipped dump of psts-all (to be downloaded by download_db.sh)
    |-- db/  -- mongodb instance of psts-all  (filled after running build_db_from_dump.sh)

Sizes

Currently PSTS comes in two sizes:

• psts-all (252.0 GB): Contains up to 10k sentences pertaining to each of the PPDB-XXL paraphrases having PPDB2.0Score at least 2.0

• psts-small (2.2 GB): Contains up to 30 sentences pertaining to the PPDB-XXL paraphrases having PPDB2.0Score at least 2.0

Download and Installation

1. First you'll need to have MongoDB installed, which you can do by following the directions for your OS here.

2. Download one of the PSTS dump files (psts-all.zip or psts-small.zip) using the script ./src/download_db.sh:

   ./src/download_db.sh <SIZE>
   

   where <SIZE> is one of all or small. Warning the all version is 252 GB compressed. If you are just browsing, or don't need more than 30 sentences per paraphrase pair, use the small version.

3. Build the database from the downloaded dump, using the script ./src/build_db_from_dump.sh:

   ./src/build_db_from_dump.sh <SIZE>
   

   where <SIZE> is one of all or small. The whole process should take about 15-20 min for the small version...significantly longer for all.

4. Once the MongoDB instance has been setup in step 2, each time you want to query it using psts.py you'll need to deploy a local instance of the database. You can do this easily by running:

   ./src/deploy.sh <SIZE>
   

   where <SIZE> is one of all or small.

5. The script psts.py contains functions for querying the local PSTS MongoDB instance. The main query function is PSTS.find() -- here is an example of its use:

   import psts
   
   resource = psts.PSTS('localhost',27017)
   
   result = resource.find('NN', 'bug', 'microbe', n=5, weights='score')
   for sent in result:
       print '\t'.join((sent['tokens'], sent['lang'], sent['trans']))
   
   # ulcer bug strongly linked to irregular heartbeat        ar      ميكروب
   # how do you keep yourself safe from the flu bug when traveling ? ar      ميكروب
   # could have been a bug he got here .     fr      microbe
   # has he got this bug ?   fr      microbe
   # i mean , if we can get a bug that 's resistant to the virus , this might be worth it .        fr      bactérie

Database Structure

The PSTS resource is stored as a collection of MongoDB documents. For a (tgt, pp) (target, paraphrase) pair, the available sentences in PSTS are stored with the structure:

{'pos': 'NN',
 'pp1': 'bug',
 'pp2': 'microbe',
 'ppdb2score': 2.51369,
 'pp1_sents': [{'lang': "ar",
                'e_idx': 1,
                's_idx': 1,
                'log_p_f': -14.8315074951,
                'tokens': "ulcer bug strongly linked to irregular heartbeat",
                'score': -1.0,
                'log_pmi': 12.6342829177,
                'log_p_f_e': -2.19722457734,
                'trans': "ميكروب"}, ...],
  'pp2_sents': [{'lang': "fr",
                 'e_idx':  2,
                 's_idx':  2,
                 'log_p_f': -14.5358728955,
                 'tokens': "although the microbe has occasionally caused infections among patients with traumatic injuries , an outbreak in watsonville , california , was the first to spread in a community .",
                 'score': -1.0,
                 'log_pmi': 9.19114915613,
                 'log_p_f_e': -5.34472373936,
                 'trans': "microbe"}, ...],
}

Attributes are as follows:

• pos : the part of speech of the paraphrase pair
• pp1 : the source or target, whichever is first alphabetically
• pp2 : the source or target, whichever is second alphabetically
• ppdb2score : the maximum of the ppdbscore of (pp1,pp2) or (pp2,pp1)
• pp1_sents : PSTS sentences containing pp1 (target) in its sense as a paraphrase of pp2 (pp)
• pp2_sents : PSTS sentences containing pp2 (target) in its sense as a paraphrase of pp1 (pp)

Within the sentence lists, attributes of each dict are:

• lang : the language of the translation used to extract the sentence
• trans : the translation used to extract the sentence
• tokens : the (tokenized) sentence
• s_idx : the starting index of the English target word in the tokenized sentence
• e_idx : the ending index of the English target word in the tokenized sentence
• log_pmi : the PMI of the foreign translation with the pp in the paraphrase pair
• log_p_f : the translation probability of the translation
• log_p_f_e : the alignment probability between the pp in the paraphrase pair, and the foreign translation
• score : the regression model predicted quality score for the sentence (should correlate with how 'characteristic' the sentence is of the shared meaning of the paraphrase pair)