xtsv – A generic TSV-style format based intermodular communication framework and REST API implemented in Python

• inter-module communication via a TSV-style format
  • processing can be started or stopped at any module
  • module dependency checks before processing
  • easy to add new modules
  • multiple alternative modules for some tasks
• easy to use command-line interface
• convenient REST API with simple web frontend
• Python library API
• Can be turned into a docker image and runnable docker form

If a bug is found please leave feedback with the exact details.

Citing and License

xtsv is licensed under the LGPL 3.0 license. The submodules have their own license.

If you use this library, please cite the following paper:

Indig, Balázs, Bálint Sass, and Iván Mittelholcz. "The xtsv Framework and the Twelve Virtues of Pipelines." Proceedings of The 12th Language Resources and Evaluation Conference. 2020.

@inproceedings{indig-etal-2020-xtsv,
    title = "The xtsv Framework and the Twelve Virtues of Pipelines",
    author = "Indig, Bal{\'a}zs  and
      Sass, B{\'a}lint  and
      Mittelholcz, Iv{\'a}n",
    booktitle = "Proceedings of The 12th Language Resources and Evaluation Conference",
    month = may,
    year = "2020",
    address = "Marseille, France",
    publisher = "European Language Resources Association",
    url = "https://www.aclweb.org/anthology/2020.lrec-1.871",
    pages = "7044--7052",
    abstract = "We present xtsv, an abstract framework for building NLP pipelines. It covers several kinds of functionalities which can be implemented at an abstract level. We survey these features and argue that all are desired in a modern pipeline. The framework has a simple yet powerful internal communication format which is essentially tsv (tab separated values) with header plus some additional features. We put emphasis on the capabilities of the presented framework, for example its ability to allow new modules to be easily integrated or replaced, or the variety of its usage options. When a module is put into xtsv, all functionalities of the system are immediately available for that module, and the module can be be a part of an xtsv pipeline. The design also allows convenient investigation and manual correction of the data flow from one module to another. We demonstrate the power of our framework with a successful application: a concrete NLP pipeline for Hungarian called e-magyar text processing system (emtsv) which integrates Hungarian NLP tools in xtsv. All the advantages of the pipeline come from the inherent properties of the xtsv framework.",
    language = "English",
    ISBN = "979-10-95546-34-4",
}

Requirements

• Python 3.7 <=
• [Optional, if required by any module] PyJNIus and OpenJDK 11 JDK

API documentation

• ModuleError: The exception thrown when something bad happened to the modules (e.g. the module could not be found or the ordering of the modules is not feasible because of the required and supplied fields)
• HeaderError: The exception thrown when the input could not satisfy the required fields in its header
• jnius_config: Set JAVA VM options and CLASSPATH for the PyJNIus library
• build_pipeline(inp_data, used_tools, available_tools, presets, conll_comments=False) -> iterator_on_output_lines: Build the current pipeline from the input data (stream, iterable or string), the list of the elements of the desired pipeline chosen from the available tools and presets returning an output iterator
• pipeline_rest_api(name, available_tools, presets, conll_comments, singleton_store=None, form_title, doc_link) -> app: Create a Flask application with the REST API and web frontend on the available initialised tools and presets with the desired name. Run with a wsgi server or Flask's built-in server with with app.run() (see REST API section)
• singleton_store_factory() -> singleton: Singletons can be used for initialisation of modules (eg. when the application is restarted frequently and not all modules are used between restarts)
• process(stream, initialised_app, conll_comments=False) -> iterator_on_output_lines: A low-level API to run a specific member of the pipeline on a specific input stream, returning an output iterator
• parser_skeleton(...) -> argparse.ArgumentParser(...): A CLI argument parser skeleton can be further customized when needed
• add_bool_arg(parser, name, help_text, default=False, has_negative_variant=True): A helper function to easily add BOOL arguments to the ArgumentParser class

To be defined by the actual pipeline:

• tools: The list of tools (see configuration for details)
• presets: The dictionary of shorthands for tasks which are defined as list of tools to be run in a pipeline (see configuration for details)

Data format

The input and output can be one of the following:

• Free form text file
• TSV file with fixed column order and without header (like CoNLL-U)
• TSV file with arbitrary column order where the columns are identified by the TSV header (main format of xtsv)

The TSV files are formatted as follows (closely resembling the CoNLL-U, vertical format):

• The first line is the header (when the column order is not fixed, therefore the next module identifies columns by their names)
• Columns are separated by TAB characters
• One token per line (one column), the other columns contain the information (stem, POS-tag, etc.) of that individual token
• Sentences are separated by emtpy lines
• If allowed by settings, zero or more comment lines (e.g. lines starting with hashtag and space) immediately precede the sentences

The fields (represented by TSV columns) are identified by the header in the first line of the input. Each module can (but does not necessarily have to) define:

• A set of source fields which is required to present in the input
• A list of target fields which are to be generated to the output in order
  • Newly generated fields are started from the right of the rightmost column, the existing columns should not be modified at all

The following types of modules can be defined by their input and output format requirements:

• Tokeniser: No source fields, no header, has target fields, free-format text as input, TSV+header output
• Internal module: Has source fields, has header, has target fields, TSV+header input, TSV+header output
• Finalizer: Has source fields, no header, no target fields, TSV+header input, free-format text as output
• Fixed-order TSV importer: No source fields, no header, has target fields, Fixed-order TSV w/o header as input, TSV+header output
• Fixed-order TSV processor: No source fields, no header, no target fields, Fixed-order TSV w/o header as input, Fixed-order TSV w/o header as output

Creating a module that can be used with xtsv

We strive to be a welcoming open source community. In agreement with the license, everybody is free to create a new compatible module without asking for permission.

The following requirements apply for a new module:

1. It must provide (at least) the mandatory API (see emDummy for a well-documented example)
2. It must conform to the (to be defined) field-name conventions and the format conventions
3. It must have an LGPL 3.0 compatible license (as all modules communicate through the thin xtsv API, there is no restriction or obligation to commit for the module license. This is not legal advice!)

The following technical steps are needed to insert the new module into the pipeline:

1. Add the new module package as a requirement to the requirements.txt of the pipeline's main repository (e.g. emtsv)

2. Insert the configuration in config.py:

   # Setup the tuple:
   #   module name,
   #   class,
   #   friendly name,
   #   args (tuple),
   #   kwargs (dict)
   em_dummy = (
       'emdummy',
       'EmDummy',
       'EXAMPLE (The friendly name of DummyTagger used in REST API form)',
       ('Params', 'goes', 'here'),
       {
           'source_fields': {'Source field names'},
           'target_fields': ['Target field names'],
           'other': 'kwargs as needed',
       }
   )

3. Add the new module to tools list in config.py, optionally also to presets dictionary

   tools = [
       ...,
       (em_dummy, ('dummy-tagger', 'emDummy')),
   ]

4. Update README.md with the short description of the newly added module and add neccessary documentaion (e.g. extra installation instructions)

5. Test, commit and push (create a pull request if you want to include your module in other's pipeline)

Installation

• Can be installed as pip package: pip3 install xtsv
• Or by using the git repository as submodule for another git repository

Usage

Here we present the usage scenarios.

To extend the toolchain with new modules, just add new modules to config.py.

Some examples of the realised applications:

• emtsv
• emmorphpy
• HunTag3

Command-line interface

• Multiple modules at once (not necessarily starting with raw text):

  echo "Input text." | python3 ./main.py modules,separated,by,comas

• Modules glued together one by one with the standard *nix pipelines where users can interact with the data between the modules:

  echo "Input text." | \
      python3 main.py module | \
      python3 main.py separated | \
      python3 main.py by | \
      python3 main.py comas

• Independently from the other options, xtsv can also be used with input or output streams redirected or with string input (this applies to the runnable docker form as well):

  python3 ./main.py modules,separated,by,comas -i input.txt -o output.txt
  python3 ./main.py modules,separated,by,comas --text "Input text."

Docker image

With the appropriate Dockerfile xtsv can be used as follows

• Runnable docker form (CLI usage of docker image):

  cat input.txt | docker run -i xtsv-docker task,separated,by,comas > output.txt

As service through Rest API (docker container)

docker run --rm -p5000:5000 -it xtsv-docker  # REST API listening on http://0.0.0.0:5000

REST API

Server

• RECOMMENDED WAY: Docker image (see above)

• Any wsgi server (uwsgi, gunicorn, waitress, etc.) can be configured to run with a prepared wsgi file .

• Debug server (Flask) only for development (single threaded, one request at a time):

  When the server outputs a message like * Running on then it is ready to accept requests on http://127.0.0.1:5000. (We do not recommend using this method in production as it is built atop of Flask debug server! Please consider using the Docker image for REST API in production!)

Client

• Web fronted provided by xtsv

• From Python (the URL contains the tools to be run separated by /):

  >>> import requests
  >>> # With input file
  >>> r = requests.post('http://127.0.0.1:5000/tools/separated/by/slashes', files={'file': open('input.file', encoding='UTF-8')})
  >>> print(r.text)
  ...
  >>> # With input text
  >>> r = requests.post('http://127.0.0.1:5000/tools/separated/by/slashes', data={'text': 'Input text.'})
  >>> print(r.text)
  ...
  >>> # CoNLL style comments can be enabled per request (disabled by default):
  >>> r = requests.post('http://127.0.0.1:5000/tools/separated/by/slashes', files={'file':open('input.file', encoding='UTF-8')}, data={'conll_comments': True})
  >>> print(r.text)
  ...

  The server checks whether the module order is feasible, and returns an error message if there are any problems.

As Python Library

1. Install xtsv package or make sure the main pipeline's installation is in the PYTHONPATH environment variable.

2. import xtsv

3. Example:

   import sys
   from xtsv import build_pipeline, parser_skeleton, jnius_config, process, pipeline_rest_api, singleton_store_factory
   # Imports end here. Must do only once per Python session
   
   argparser = parser_skeleton(description='An example pipeline for xtsv')
   opts = argparser.parse_args()
   
   jnius_config.classpath_show_warning = opts.verbose  #  False to suppress warning
   
   # Set input from any stream, iterator or raw string in any acceptable format
   if opts.input_text is not None:
       # Raw, or processed TSV input list and output file...
       # input_data = ['A kutya', 'elment sétálni.']  # Raw text line by line
       # Processed data: header and the token POS-tag pairs line by line
       # input_data = [['form', 'xpostag'], ['A', '[/Det|Art.Def]'], ['kutya', '[/N][Nom]'], ['elment', '[/V][Pst.NDef.3Sg]'], ['sétálni', '[/V][Inf]'], ['.', '.']]
       input_data = opts.input_text
   else:
       # Set input from any stream or iterable and output stream...
       input_data = opts.input_stream
   
   # Set output iterator: e.g. output_iterator = open('output.txt', 'w', encoding='UTF-8')  # File
   output_iterator = opts.output_stream
   
   # Select a predefined task to do or provide your own list of pipeline elements
   # i.e. set the tagger name as in the _tools dictionary in the config.py_ e.g. used_tools = ['dummy']
   used_tools = ['tools', 'in', 'a', 'list']
   presets = []
   
   # The relevant part of config.py
   # from emdummy import EmDummy
   em_dummy = ('emdummy', 'EmDummy', 'EXAMPLE (The friendly name of EmDummy used in REST API form)',
               ('Params', 'goes', 'here'), {'source_fields': {'form'},  # Source field names
                                            'target_fields': {'star'}})  # Target field names
   tools = [(em_dummy, ('dummy', 'dummy-tagger', 'emDummy'))]
   
   
   # Run the pipeline on input and write result to the output...
   # You can enable or disable CoNLL-U style comments here (default: disabled)
   output_iterator.writelines(build_pipeline(input_data, used_tools, tools, presets, opts.conllu_comments,
                                             opts.output_header))
   
   # Alternative: Run specific tool for input streams (still in emtsv format).
   # Useful for training a module (see Huntag3 for details):
   # e.g. output_iterator.writelines(process(input_data, EmDummy(*em_dummy[3], **em_dummy[4])))
   output_iterator.writelines(process(sys.stdin, Module('with', 'params')))
   
   # Or process individual tokens further... WARNING: The header will be the
   # first item in the iterator!
   for tok in build_pipeline(input_data, used_tools, tools, presets, opts.conllu_comments, opts.output_header):
       if len(tok) > 1:  # Empty line (='\n') means end of sentence
           form, xpostag, *rest = tok.strip().split('\t')  # Split to the expected columns
   
   # Alternative2: Run REST API debug server
   app = pipeline_rest_api(name='TEST', available_tools=tools, presets=presets,
                           conll_comments=opts.conllu_comments, singleton_store=singleton_store_factory(),
                           form_title='TEST TITLE', doc_link='https://github.com/nytud/xtsv',
                           output_header=opts.output_header)
   # And run the Flask debug server separately
   app.run()