Updated: Nov 10, 2020
The full pack of data and code for the official evaluation is at: .data/SemEval-PIT2015-github.zip
- Wei Xu, University of Pennsylvania
- Chris Callison-Burch, University of Pennsylvania
- Bill Dolan, Microsoft Research
paper about the dataset, baselines, and the MultiP model (multiple-instance learning paraphrase):
@article{Xu-EtAl-2014:TACL,
author = {Xu, Wei and Ritter, Alan and Callison-Burch, Chris and Dolan, William B. and Ji, Yangfeng"}
title = {Extracting Lexically Divergent Paraphrases from {Twitter}},
journal = {Transactions of the Association for Computational Linguistics},
volume = {2},
year = {2014},
pages = {435--448},
publisher = {Association for Computational Linguistics},
url = {https://www.aclweb.org/anthology/Q14-1034},
doi = {10.1162/tacl_a_00194}
}
overview paper of the shared task:
@inproceedings{xu2015semeval,
author = {Xu, Wei and Callison-Burch, Chris and Dolan, Bill},
title = {{SemEval-2015 Task} 1: Paraphrase and Semantic Similarity in {Twitter} ({PIT})},
booktitle = {Proceedings of the 9th International Workshop on Semantic Evaluation ({S}em{E}val 2015)},
year = {2015},
address. = {Denver, Colorado},
publisher = {Association for Computational Linguistics},
url = {https://www.aclweb.org/anthology/S15-2001},
doi = {10.18653/v1/S15-2001},
pages = {1--11},
}
The dataset contains the following files:
./data/train.data (13063 sentence pairs)
./data/dev.data (4727 sentence pairs)
./data/test.data (972 sentences pairs)
./data/test.label (a separate file of labels only, used by evaluation scripts)
Both data files come in the tab-separated format. Each line contains 7 columns:
Topic_Id | Topic_Name | Sent_1 | Sent_2 | Label | Sent_1_tag | Sent_2_tag |
The "Topic_Name" are the names of trends provided by Twitter, which are not hashtags.
The "Sent_1" and "Sent_2" are the two sentences, which are not necessarily full tweets. Tweets were tokenized by Brendan O'Connor et al.'s toolkit (ICWSM 2010) and split into sentences.
The "Sent_1_tag" and "Sent_2_tag" are the two sentences with part-of-speech and named entity tags by Alan Ritter et al.'s toolkit (RANLP 2013, EMNLP 2011).
The "Label" column for *dev/train data * is in a format like "(1, 4)", which means among 5 votes from Amazon Mechanical turkers only 1 is positive and 4 are negative. We would suggest map them to binary labels as follows:
paraphrases: (3, 2) (4, 1) (5, 0)
non-paraphrases: (1, 4) (0, 5)
debatable: (2, 3) which you may discard if training binary classifier
The "Label" column for test data is in a format of a single digit between
between 0 (no relation) and 5 (semantic equivalence), annotated by expert.
We would suggest map them to binary labels as follows:
paraphrases: 4 or 5
non-paraphrases: 0 or 1 or 2
debatable: 3 which we discarded in Paraphrase Identification evaluation
We discarded the debatable cases in the evaluation of Paraphrase Identification task, but kept them in the evaluation of Semantic Similarity task.
There are two scripts for the official evaluation:
./scripts/pit2015_checkformat.py (checks the format or the system output file)
./scripts/pit2015_eval_single.py (evaluation metrics)
The participants are required to produce a binary label (paraphrase) for each sentence
pair, and optionally a real number between 0 (no relation) and 1 (semantic equivalence)
for measuring semantic similarity.
The system output file should match the lines of the test data. Each line has 2 columns and separated by a tab in between, like this: | Binary Label (true/false) | Degreed Score (between 0 and 1, in the 4 decimal format) | if your system only gives binary labels, put "0.0000" in all second columns.
The output file names look like this: PIT2015_TEAMNAME_01_nameofthisrun.output PIT2015_TEAMNAME_02_nameofthisrun.output
There are scripts for two baselines:
./scripts/baseline_random.py
./scripts/baseline_logisticregression.py
and their outputs on the test data, plus outputs from two state-of-the-art systems:
./systemoutputs/PIT2015_BASELINE_01_random.output
./systemoutputs/PIT2015_BASELINE_02_LG.output
./systemoutputs/PIT2015_BASELINE_03_WTMF.output
./systemoutputs/PIT2015_BASELINE_04_MultiP.output
(1) The logistic regression (LG) model using simple lexical overlap features:
It is our reimplementation in Python. This is a baseline originally used by Dipanjan Das and Noah A. Smith (ACL 2009): "Paraphrase Identification as Probabilistic Quasi-Synchronous Recognition".
To run the script, you will need to install NLTK and Megam packages: http://www.nltk.org/_modules/nltk/classify/megam.html http://www.umiacs.umd.edu/~hal/megam/index.html
If you have troubles with Megam, you may need to rebuild it from source code: http://stackoverflow.com/questions/11071901/stuck-in-using-megam-in-python-nltk-classify-maxentclassifier
Example output, if training on train.data and test on dev.data will look like:
Read in 11513 training data ... (after discarding the data with debatable cases)
Read in 4139 test data ... (see details in TRAIN/DEV DATA section)
PRECISION: 0.704069050555
RECALL: 0.389229720518
F1: 0.501316944688
ACCURACY: 0.725537569461
The script will provide the numbers for plotting precision/recall curves, or a single precision/recall/F1 score with 0.5 cutoff of predicated probability.
(2) The Weighted Matrix Factorization (WTMF) model is a unsupervised approach developed by Weiwei Guo and Mona Diab (ACL 2012): "Modeling Sentences in the Latent Space" Its code is available at: http://www.cs.columbia.edu/~weiwei/code.html
(3) The Multiple-instance Learning Paraphrase model (MultiP) is a supervised approach developed by Wei Xu et al. (TACL 2014): "Extracting Lexically Divergent Paraphrases from Twitter" Its code is available at: http://www.cis.upenn.edu/~xwe/multip/