A library to help users choose appropriate summarization tools based on their specific tasks or needs. Includes models, evaluation metrics, and datasets.
The library architecture is as follows:
NOTE: SummerTime is in active development, any helpful comments are highly encouraged, please open an issue or reach out to any of the team members.
# install extra dependencies first
pip install pyrouge@git+https://github.com/bheinzerling/pyrouge.git
pip install en_core_web_sm@https://github.com/explosion/spacy-models/releases/download/en_core_web_sm-3.0.0/en_core_web_sm-3.0.0-py3-none-any.whl
# install summertime from PyPI
pip install summertime
Alternatively, to enjoy the most recent features, you can install from the source:
git clone [email protected]:Yale-LILY/SummerTime
pip install -e .export ROUGE_HOME=/usr/local/lib/python3.7/dist-packages/summ_eval/ROUGE-1.5.5/Imports model, initializes default model, and summarizes sample documents.
from summertime import model
sample_model = model.summarizer()
documents = [
""" PG&E stated it scheduled the blackouts in response to forecasts for high winds amid dry conditions.
The aim is to reduce the risk of wildfires. Nearly 800 thousand customers were scheduled to be affected
by the shutoffs which were expected to last through at least midday tomorrow."""
]
sample_model.summarize(documents)
# ["California's largest electricity provider has turned off power to hundreds of thousands of customers."]Also, please run our colab notebook for a more hands-on demo and more examples.
SummerTime supports different models (e.g., TextRank, BART, Longformer) as well as model wrappers for more complex summarization tasks (e.g., JointModel for multi-doc summarzation, BM25 retrieval for query-based summarization). Several multilingual models are also supported (mT5 and mBART).
| Models | Single-doc | Multi-doc | Dialogue-based | Query-based | Multilingual |
|---|---|---|---|---|---|
| BartModel | ✔️ | ||||
| BM25SummModel | ✔️ | ||||
| HMNetModel | ✔️ | ||||
| LexRankModel | ✔️ | ||||
| LongformerModel | ✔️ | ||||
| MBartModel | ✔️ | 50 languages (full list here) | |||
| MT5Model | ✔️ | 101 languages (full list here) | |||
| TranslationPipelineModel | ✔️ | ~70 languages | |||
| MultiDocJointModel | ✔️ | ||||
| MultiDocSeparateModel | ✔️ | ||||
| PegasusModel | ✔️ | ||||
| TextRankModel | ✔️ | ||||
| TFIDFSummModel | ✔️ |
To see all supported models, run:
from summertime.model import SUPPORTED_SUMM_MODELS
print(SUPPORTED_SUMM_MODELS)from summertime import model
# To use a default model
default_model = model.summarizer()
# Or a specific model
bart_model = model.BartModel()
pegasus_model = model.PegasusModel()
lexrank_model = model.LexRankModel()
textrank_model = model.TextRankModel()Users can easily access documentation to assist with model selection
default_model.show_capability()
pegasus_model.show_capability()
textrank_model.show_capability()To use a model for summarization, simply run:
documents = [
""" PG&E stated it scheduled the blackouts in response to forecasts for high winds amid dry conditions.
The aim is to reduce the risk of wildfires. Nearly 800 thousand customers were scheduled to be affected
by the shutoffs which were expected to last through at least midday tomorrow."""
]
default_model.summarize(documents)
# or
pegasus_model.summarize(documents)All models can be initialized with the following optional options:
def __init__(self,
trained_domain: str=None,
max_input_length: int=None,
max_output_length: int=None,
):All models will implement the following methods:
def summarize(self,
corpus: Union[List[str], List[List[str]]],
queries: List[str]=None) -> List[str]:
def show_capability(cls) -> None:SummerTime supports different summarization datasets across different domains (e.g., CNNDM dataset - news article corpus, Samsum - dialogue corpus, QM-Sum - query-based dialogue corpus, MultiNews - multi-document corpus, ML-sum - multi-lingual corpus, PubMedQa - Medical domain, Arxiv - Science papers domain, among others.
| Dataset | Domain | # Examples | Src. length | Tgt. length | Query | Multi-doc | Dialogue | Multi-lingual |
|---|---|---|---|---|---|---|---|---|
| ArXiv | Scientific articles | 215k | 4.9k | 220 | ||||
| CNN/DM(3.0.0) | News | 300k | 781 | 56 | ||||
| MlsumDataset | Multi-lingual News | 1.5M+ | 632 | 34 | ✔️ | German, Spanish, French, Russian, Turkish | ||
| Multi-News | News | 56k | 2.1k | 263.8 | ✔️ | |||
| SAMSum | Open-domain | 16k | 94 | 20 | ✔️ | |||
| Pubmedqa | Medical | 272k | 244 | 32 | ✔️ | |||
| QMSum | Meetings | 1k | 9.0k | 69.6 | ✔️ | ✔️ | ||
| ScisummNet | Scientific articles | 1k | 4.7k | 150 | ||||
| SummScreen | TV shows | 26.9k | 6.6k | 337.4 | ✔️ | |||
| XSum | News | 226k | 431 | 23.3 | ||||
| XLSum | News | 1.35m | ??? | ??? | 45 languages (see documentation) | |||
| MassiveSumm | News | 12m+ | ??? | ??? | 78 languages (see Multilingual Summarization section of README for details) |
To see all supported datasets, run:
from summertime import dataset
print(dataset.list_all_dataset())from summertime import dataset
cnn_dataset = dataset.CnndmDataset()
# or
xsum_dataset = dataset.XsumDataset()
# ..etcAll datasets are implementations of the SummDataset class. Their data splits can be accessed as follows:
dataset = dataset.CnndmDataset()
train_data = dataset.train_set
dev_data = dataset.dev_set
test_data = dataset.test_set To see the details of the datasets, run:
dataset = dataset.CnndmDataset()
dataset.show_description()The data in all datasets is contained in a SummInstance class object, which has the following properties:
data_instance.source = source # either `List[str]` or `str`, depending on the dataset itself, string joining may needed to fit into specific models.
data_instance.summary = summary # a string summary that serves as ground truth
data_instance.query = query # Optional, applies when a string query is present
print(data_instance) # to print the data instance in its entiretyData is loaded using a generator to save on space and time
data_instance = next(cnn_dataset.train_set)
print(data_instance)import itertools
# Get a slice from the train set generator - first 5 instances
train_set = itertools.islice(cnn_dataset.train_set, 5)
corpus = [instance.source for instance in train_set]
print(corpus)You can use custom data using the CustomDataset class that loads the data in the SummerTime dataset Class
from summertime.dataset import CustomDataset
''' The train_set, test_set and validation_set have the following format:
List[Dict], list of dictionaries that contain a data instance.
The dictionary is in the form:
{"source": "source_data", "summary": "summary_data", "query":"query_data"}
* source_data is either of type List[str] or str
* summary_data is of type str
* query_data is of type str
The list of dictionaries looks as follows:
[dictionary_instance_1, dictionary_instance_2, ...]
'''
# Create sample data
train_set = [
{
"source": "source1",
"summary": "summary1",
"query": "query1", # only included, if query is present
}
]
validation_set = [
{
"source": "source2",
"summary": "summary2",
"query": "query2",
}
]
test_set = [
{
"source": "source3",
"summary": "summary3",
"query": "query3",
}
]
# Depending on the dataset properties, you can specify the type of dataset
# i.e multi_doc, query_based, dialogue_based. If not specified, they default to false
custom_dataset = CustomDataset(
train_set=train_set,
validation_set=validation_set,
test_set=test_set,
query_based=True,
multi_doc=True
dialogue_based=False)import itertools
from summertime import dataset, model
cnn_dataset = dataset.CnndmDataset()
# Get a slice of the train set - first 5 instances
train_set = itertools.islice(cnn_dataset.train_set, 5)
corpus = [instance.source for instance in train_set]
# Example 1 - traditional non-neural model
# LexRank model
lexrank = model.LexRankModel(corpus)
print(lexrank.show_capability())
lexrank_summary = lexrank.summarize(corpus)
print(lexrank_summary)
# Example 2 - A spaCy pipeline for TextRank (another non-neueral extractive summarization model)
# TextRank model
textrank = model.TextRankModel()
print(textrank.show_capability())
textrank_summary = textrank.summarize(corpus)
print(textrank_summary)
# Example 3 - A neural model to handle large texts
# LongFormer Model
longformer = model.LongFormerModel()
longformer.show_capability()
longformer_summary = longformer.summarize(corpus)
print(longformer_summary)The summarize() method of multilingual models automatically checks for input document language.
Single-doc multilingual models can be initialized and used in the same way as monolingual models. They return an error if a language not supported by the model is input.
mbart_model = st_model.MBartModel()
mt5_model = st_model.MT5Model()
# load Spanish portion of MLSum dataset
mlsum = datasets.MlsumDataset(["es"])
corpus = itertools.islice(mlsum.train_set, 5)
corpus = [instance.source for instance in train_set]
# mt5 model will automatically detect Spanish as the language and indicate that this is supported!
mt5_model.summarize(corpus)The following languages are currently supported in our implementation of the MassiveSumm dataset: Afrikaans, Amharic, Arabic, Assamese, Aymara, Azerbaijani, Bambara, Bengali, Tibetan, Bosnian, Bulgarian, Catalan, Czech, Welsh, Danish, German, Greek, English, Esperanto, Persian, Filipino, French, Fulah, Irish, Gujarati, Haitian, Hausa, Hebrew, Hindi, Croatian, Hungarian, Armenian,Igbo, Indonesian, Icelandic, Italian, Japanese, Kannada, Georgian, Khmer, Kinyarwanda, Kyrgyz, Korean, Kurdish, Lao, Latvian, Lingala, Lithuanian, Malayalam, Marathi, Macedonian, Malagasy, Mongolian, Burmese, South Ndebele, Nepali, Dutch, Oriya, Oromo, Punjabi, Polish, Portuguese, Dari, Pashto, Romanian, Rundi, Russian, Sinhala, Slovak, Slovenian, Shona, Somali, Spanish, Albanian, Serbian, Swahili, Swedish, Tamil, Telugu, Tetum, Tajik, Thai, Tigrinya, Turkish, Ukrainian, Urdu, Uzbek, Vietnamese, Xhosa, Yoruba, Yue Chinese, Chinese, Bislama, and Gaelic.
SummerTime supports different evaluation metrics including: BertScore, Bleu, Meteor, Rouge, RougeWe
To print all supported metrics:
from summertime.evaluation import SUPPORTED_EVALUATION_METRICS
print(SUPPORTED_EVALUATION_METRICS)import summertime.evaluation as st_eval
bert_eval = st_eval.bertscore()
bleu_eval = st_eval.bleu_eval()
meteor_eval = st_eval.bleu_eval()
rouge_eval = st_eval.rouge()
rougewe_eval = st_eval.rougewe()All evaluation metrics can be initialized with the following optional arguments:
def __init__(self, metric_name):All evaluation metric objects implement the following methods:
def evaluate(self, model, data):
def get_dict(self, keys):Get sample summary data
from summertime.evaluation.base_metric import SummMetric
from summertime.evaluation import Rouge, RougeWe, BertScore
import itertools
# Evaluates model on subset of cnn_dailymail
# Get a slice of the train set - first 5 instances
train_set = itertools.islice(cnn_dataset.train_set, 5)
corpus = [instance for instance in train_set]
print(corpus)
articles = [instance.source for instance in corpus]
summaries = sample_model.summarize(articles)
targets = [instance.summary for instance in corpus]Evaluate the data on different metrics
from summertime.evaluation import BertScore, Rouge, RougeWe,
# Calculate BertScore
bert_metric = BertScore()
bert_score = bert_metric.evaluate(summaries, targets)
print(bert_score)
# Calculate Rouge
rouge_metric = Rouge()
rouge_score = rouge_metric.evaluate(summaries, targets)
print(rouge_score)
# Calculate RougeWe
rougewe_metric = RougeWe()
rougwe_score = rougewe_metric.evaluate(summaries, targets)
print(rougewe_score)Given a SummerTime dataset, you may use the pipelines.assemble_model_pipeline function to retrieve a list of initialized SummerTime models that are compatible with the dataset provided.
from summertime.pipeline import assemble_model_pipeline
from summertime.dataset import CnndmDataset, QMsumDataset
single_doc_models = assemble_model_pipeline(CnndmDataset)
# [
# (<model.single_doc.bart_model.BartModel object at 0x7fcd43aa12e0>, 'BART'),
# (<model.single_doc.lexrank_model.LexRankModel object at 0x7fcd43aa1460>, 'LexRank'),
# (<model.single_doc.longformer_model.LongformerModel object at 0x7fcd43b17670>, 'Longformer'),
# (<model.single_doc.pegasus_model.PegasusModel object at 0x7fccb84f2910>, 'Pegasus'),
# (<model.single_doc.textrank_model.TextRankModel object at 0x7fccb84f2880>, 'TextRank')
# ]
query_based_multi_doc_models = assemble_model_pipeline(QMsumDataset)
# [
# (<model.query_based.tf_idf_model.TFIDFSummModel object at 0x7fc9c9c81e20>, 'TF-IDF (HMNET)'),
# (<model.query_based.bm25_model.BM25SummModel object at 0x7fc8b4fa8c10>, 'BM25 (HMNET)')
# ]=======
Given a SummerTime dataset, you may use the pipelines.assemble_model_pipeline function to retrieve a list of initialized SummerTime models that are compatible with the dataset provided.
# Get test data
import itertools
from summertime.dataset import XsumDataset
# Get a slice of the train set - first 5 instances
sample_dataset = XsumDataset()
sample_data = itertools.islice(sample_dataset.train_set, 100)
generator1 = iter(sample_data)
generator2 = iter(sample_data)
bart_model = BartModel()
pegasus_model = PegasusModel()
models = [bart_model, pegasus_model]
metrics = [metric() for metric in SUPPORTED_EVALUATION_METRICS]from summertime.evaluation.model_selector import ModelSelector
selector = ModelSelector(models, generator1, metrics)
table = selector.run()
print(table)
visualization = selector.visualize(table)
from summertime.evaluation.model_selector import ModelSelector
new_selector = ModelSelector(models, generator2, metrics)
smart_table = new_selector.run_halving(min_instances=2, factor=2)
print(smart_table)
visualization_smart = selector.visualize(smart_table)from summertime.evaluation.model_selector import ModelSelector
from summertime.evaluation.error_viz import scatter
keys = ("bert_score_f1", "bleu", "rouge_1_f_score", "rouge_2_f_score", "rouge_l_f_score", "rouge_we_3_f", "meteor")
scatter(models, sample_data, metrics[1:3], keys=keys[1:3], max_instances=5)
Create a pull request and name it [your_gh_username]/[your_branch_name]. If needed, resolve your own branch's merge conflicts with main. Do not push directly to main.
If you haven't already, install black and flake8:
pip install black
pip install flake8Before pushing commits or merging branches, run the following commands from the project root. Note that black will write to files, and that you should add and commit changes made by black before pushing:
black .
flake8 .Or if you would like to lint specific files:
black path/to/specific/file.py
flake8 path/to/specific/file.pyEnsure that black does not reformat any files and that flake8 does not print any errors. If you would like to override or ignore any of the preferences or practices enforced by black or flake8, please leave a comment in your PR for any lines of code that generate warning or error logs. Do not directly edit config files such as setup.cfg.
See the black docs and flake8 docs for documentation on installation, ignoring files/lines, and advanced usage. In addition, the following may be useful:
black [file.py] --diffto preview changes as diffs instead of directly making changesblack [file.py] --checkto preview changes with status codes instead of directly making changesgit diff -u | flake8 --diffto only run flake8 on working branch changes
Note that our CI test suite will include invoking black --check . and flake8 --count . on all non-unittest and non-setup Python files, and zero error-level output is required for all tests to pass.
Our continuous integration system is provided through Github actions. When any pull request is created or updated or whenever main is updated, the repository's unit tests will be run as build jobs on tangra for that pull request. Build jobs will either pass or fail within a few minutes, and build statuses and logs are visible under Actions. Please ensure that the most recent commit in pull requests passes all checks (i.e. all steps in all jobs run to completion) before merging, or request a review. To skip a build on any particular commit, append [skip ci] to the commit message. Note that PRs with the substring /no-ci/ anywhere in the branch name will not be included in CI.
This repository is built by the LILY Lab at Yale University, led by Prof. Dragomir Radev. The main contributors are Ansong Ni, Zhangir Azerbayev, Troy Feng, Murori Mutuma, Hailey Schoelkopf, and Yusen Zhang (Penn State).
If you use SummerTime in your work, consider citing:
@article{ni2021summertime,
title={SummerTime: Text Summarization Toolkit for Non-experts},
author={Ansong Ni and Zhangir Azerbayev and Mutethia Mutuma and Troy Feng and Yusen Zhang and Tao Yu and Ahmed Hassan Awadallah and Dragomir Radev},
journal={arXiv preprint arXiv:2108.12738},
year={2021}
}
For comments and question, please open an issue.