Skip to content

An open-source text summarization toolkit for non-experts. EMNLP'2021 Demo

License

Notifications You must be signed in to change notification settings

Yale-LILY/SummerTime

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

SummerTime - Text Summarization Toolkit for Non-experts

CI License Open In Colab

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.

Installation and setup

Install from PyPI (recommended)

# 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

Local pip installation

Alternatively, to enjoy the most recent features, you can install from the source:

git clone [email protected]:Yale-LILY/SummerTime
pip install -e .
Setup ROUGE (when using evaluation)
export ROUGE_HOME=/usr/local/lib/python3.7/dist-packages/summ_eval/ROUGE-1.5.5/

Quick Start

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.

Open In Colab

Models

Supported Models

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)

Import and initialization:

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:

Datasets

Datasets supported

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())

Dataset Initialization

from summertime import dataset

cnn_dataset = dataset.CnndmDataset()
# or 
xsum_dataset = dataset.XsumDataset()
# ..etc
Dataset Object

All 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()
Data instance

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 entirety

Loading and using data instances

Data is loaded using a generator to save on space and time

To get a single instance

data_instance = next(cnn_dataset.train_set)
print(data_instance)

To get a slice of the dataset

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)

Loading a custom dataset

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)

Using the datasets with the models - Examples

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)

Multilingual summarization

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.

Evaluation

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 and initialization:

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()

Evaluation Class

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):

Using evaluation metrics

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)

Using automatic pipeline assembly

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)')
# ]

=======

Visualizing performance of different models on your dataset

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]

Create a radar plot

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)

Create a scatter plot

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)

To contribute

Pull requests

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.

Code formatting

If you haven't already, install black and flake8:

pip install black
pip install flake8

Before 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.py

Ensure 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] --diff to preview changes as diffs instead of directly making changes
  • black [file.py] --check to preview changes with status codes instead of directly making changes
  • git diff -u | flake8 --diff to 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.

Tests

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.

Citation

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.