opusfilter 2.1.0

Toolbox for filtering parallel corpora

OpusFilter

OpusFilter is a tool for filtering and combining parallel corpora. It uses the OpusTools library (Aulamo et al., 2020) to download data from the OPUS corpus collection (Tiedemann, 2012), but can be used with any corpora in raw text format.

Features:

• Preprocessing pipelines configured with YAML
• Simple downloading of parallel corpora from OPUS with OpusTools
• Implementations for many common text file operations on parallel files
• Memory-efficient processing of large files
• Implemented filters based e.g. on language identification, word aligment, and n-gram language models
• Extendable with your own filters written in Python

OpusFilter has been presented in ACL 2020 system demonstrations.

A changelog is available in docs/CHANGELOG.md.

Table of contents

• Installing
  • Required libraries
  • Optional libraries and tools
• Citing and references
• Overview
  • Examples
  • Variables and constants
  • Running a single command
• Available functions
  • Downloading and selecting data
    • opus_read
    • concatenate
    • download
    • head
    • tail
    • slice
    • split
    • subset
    • product
    • unzip
    • write
  • Preprocessing text
    • preprocess
  • Filtering and scoring
    • remove_duplicates
    • filter
    • score
  • Using score files
    • join
    • sort
  • Training language and alignment models
    • train_ngram
    • train_aligment
  • Training and using classifiers
    • train_classifier
    • classify
• Available filters
  • Length filters
    • LengthFilter
    • LengthRatioFilter
    • AverageWordLengthFilter
    • LongWordFilter
  • Script and language identification filters
    • CharacterScoreFilter
    • LanguageIDFilter
  • Special character and similarity filters
    • HtmlTagFilter
    • TerminalPunctuationFilter
    • NonZeroNumeralsFilter
    • LongestCommonSubstringFilter
  • Language model filters
    • CrossEntropyFilter
    • CrossEntropyDifferenceFilter
  • Alignment model filters
    • WordAlignFilter
• Custom filters
• Available preprocessors
  • Tokenizer
  • Detokenizer
  • WhitespaceNormalizer
  • RegExpSub
• Custom preprocessors
• Other tools
  • opusfilter-diagram
  • opusfilter-duplicates
  • opusfilter-scores
  • opusfilter-test
• How to contribute

Installing

Install the latest release from PyPI:

• pip install opusfilter

Install from source:

• pip install . or
• python setup.py install

Required libraries

• beautifulsoup4
• graphviz
• langid
• fast-mosestokenizer
• fasttext
• OpusTools
• pandas
• pycld2
• pyhash
• ruamel.yaml
• regex
• scikit-learn
• tqdm

Optional libraries and tools

For using n-gram language model filters, you need to install VariKN (https://github.com/vsiivola/variKN) and its Python wrapper. Include the library files compiled to build/lib/python to your Python library path (e.g. by setting the PYTHONPATH environment variable).

For using word alignment filters, you need to install elfomal (https://github.com/robertostling/eflomal) and set environment variable EFLOMAL_PATH to eflomal's root directory, which contains the Python scripts align.py and makepriors.py.

Citing and references

If you use OpusFilter in your research, please cite our ACL 2020 paper:

@inproceedings{aulamo-etal-2020-opusfilter,
    title = "{O}pus{F}ilter: A Configurable Parallel Corpus Filtering Toolbox",
    author = {Aulamo, Mikko and Virpioja, Sami and Tiedemann, J{\"o}rg},
    booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations",
    month = jul,
    year = "2020",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.acl-demos.20",
    doi = "10.18653/v1/2020.acl-demos.20",
    pages = "150--156"
}

A bibliography of other papers cited in the documentation and code can be found from docs/references.bib.

Overview

The main script provided by the package is opusfilter, which takes a configuration file as an input. The configuration files are in YAML format. At the top level, they have to sections:

• common, which may include output_directory for setting where to write the output files. If it is not set, the current working directory is used.

• steps, which is a list of the corpus processing steps.

The syntax for the opusfilter is

opusfilter [--overwrite] [--last LAST] [--single SINGLE] CONFIG

where CONFIG is path to the configuration file. The script will run the steps one by one and stops when the final step has been processed (if no exceptions were raised). The script has options for setting the last step to run (--last) and running only a single step (--single). It the latter, the user has to make sure that all input files for the step already exist. The first step has number 1, and -1 points to the last step, -2 to the second to last, and so on.

By default, existing output files will be re-used, and the steps producing them skipped. The --overwrite option will force overwrite for all the steps.

Each step in steps is a dictionary (mapping) with two keys: type and parameters. Type is a string that defines the function that should be run, and parameters is a dictionary with keys that depend on the function; at minimum the output files are defined there.

Examples

A very simple configuration file that downloads a parallel corpus (here Finnish-English ParaCrawl v4) from OPUS and stores its segments to the files paracrawl.fi.gz and paracrawl.en.gz looks like this:

steps:
  - type: opus_read
    parameters:
      corpus_name: ParaCrawl
      source_language: fi
      target_language: en
      release: v4
      preprocessing: raw
      src_output: paracrawl.fi.gz
      tgt_output: paracrawl.en.gz

The corpus files processed by OpusFilter are UTF-8 text files that contain one segment per line. Compressed files are read and written if the file ends with .gz (gzip) or .bz2 (bzip2).

A bit more complex example that downloads both ParaCrawl and WMT-News sets from OPUS, concatenates the output files, and filters them so that only the segment pairs for which both languages have segment length of 1-100 words and the ratio of the lengths is at most 3 remain:

steps:
  - type: opus_read
    parameters:
      corpus_name: ParaCrawl
      source_language: fi
      target_language: en
      release: v4
      preprocessing: raw
      src_output: paracrawl.fi.gz
      tgt_output: paracrawl.en.gz

  - type: opus_read
    parameters:
      corpus_name: WMT-News
      source_language: fi
      target_language: en
      release: v2019
      preprocessing: raw
      src_output: wmt.fi.gz
      tgt_output: wmt.en.gz

  - type: concatenate
    parameters:
      inputs:
      - paracrawl.fi.gz
      - wmt.fi.gz
      output: all.fi.gz

  - type: concatenate
    parameters:
      inputs:
      - paracrawl.en.gz
      - wmt.en.gz
      output: all.en.gz

  - type: filter
    parameters:
      inputs:
      - all.fi.gz
      - all.en.gz
      outputs:
      - filtered.fi.gz
      - filtered.en.gz
      filters:
        - LengthFilter:
            unit: word
            min_length: 1
            max_length: 100

        - LengthRatioFilter:
            unit: word
            threshold: 3

YAML node anchors (&name) and references (*name) can be used. They are especially useful when defining a set of filters you want to use for different data sets. For example, if in the previous example you wanted to use the same filters separately for the ParaCrawl and WMT-News data, you can have:

  - type: filter
    parameters:
      inputs:
      - paracrawl.fi.gz
      - paracrawl.en.gz
      outputs:
      - paracrawl_filtered.fi.gz
      - paracrawl_filtered.en.gz
      filters: &myfilters
        - LengthFilter:
            unit: word
            min_length: 1
            max_length: 100

        - LengthRatioFilter:
            unit: word
            threshold: 3

  - type: filter
    parameters:
      inputs:
      - wmt.fi.gz
      - wmt.en.gz
      outputs:
      - wmt_filtered.fi.gz
      - wmt_filtered.en.gz
      filters: *myfilters

Variables and constants

If you have for example multiple bitexts, for which you want to use the same preprocessing steps, writing separate steps for each of them requires a lot of manually written configuration. While generating the YAML configuration programmatically is a recommended option especially for very complex setups, OpusFilter provides a couple of custom YAML tags for defining objects that can be replaced with constant or variable values.

The first tag, !var, is used for mapping the variable name (string) to its value. For example, if the variable or constant x is one in the current scope, {key: !var x} will be replaced by {key: 1}. The value can be any kind of object.

The second tag, !varstr, can be used for one or more variable substitutions within a single string, as in Python's format() method. For example, if l1 and l2 have the values en and fi, respectively, in the current scope, {output: !varstr "cleaned.{l1}-{l2}.txt"} will be replaced by {output: cleaned.en-fi.txt}. Note that the string template has to be quoted in order that the YAML loader will not consider the expressions inside the braces as objects.

Constant values for the variables can be defined in the common section of the configuration, having a scope in all steps, or in individual steps, having a local scope within the step. In either case, the definitions are placed in a dictionary under the constants key, with variable names as keys. For example,

common:
  constants:
    source: en

steps:
  - type: concatenate
    parameters:
      inputs:
      - !varstr "file1.{source}-{target}.gz"
      - !varstr "file2.{source}-{target}.gz"
      output: !varstr "all.{source}-{target}.gz"
    constants:
      target: fi

will produce the output file "all.en-fi.gz". The local scope of the step overrides definitions in common.

Another possibility is to define multiple choices for the variable(s) within a single step. The definitions are placed in a dictionary under the variables key, with variable names as keys and a list of intended variable values as values. For example,

common:
  constants:
    source: en

steps:
  - type: concatenate
    parameters:
      inputs:
      - !varstr "file1.{source}-{target}.gz"
      - !varstr "file2.{source}-{target}.gz"
      output: !varstr "all.{source}-{target}.gz"
    variables:
      target: [fi, sv]

will produce output files "all.en-fi.gz" and "all.en-sv.gz". If values are defined for multiple variables, the list are required to have the same length. The step is expanded into as many substeps as there are values in the lists. Note that if you need to use the same lists of variable values in multiple steps, you can exploit the standard YAML node anchors and references.

Running a single command

If you need to run a single OpusFilter function wihtout the need of writing a complete configuration, the script opusfilter-cmd is provided for convenience. With the command, you can define a single-step configuration using just command line arguments.

The syntax for the opusfilter-cmd is:

opusfilter-cmd [--overwrite] [--outputdir OUTDIR] FUNCTION [--parameters PARAMETERS] [PARAMETER-OPTION] ...

The --outputdir option defines the work directory; if not given the current directory is used. Existing output files will not be overwritten by default; the --overwrite option will force overwrite.

The FUNCTION argument defines the function to use. Parameters for the function can be defined in two ways: Providing a single JSON object (as a string) that corresponds to the parameter definition in YAML configuration, or setting the parameters with custom options.

For the former, use the --parameters option. For example, the following performs filtering with a single word length filter:

opusfilter-cmd filter --parameters {"inputs": ["wmt.fi.gz", "wmt.en.gz"], "outputs": ["wmt_filtered.fi.gz", "wmt_filtered.en.gz"], "filters": [{"LengthFilter": {"unit": "word", "min_length": 1, "max_length": 100}}]}

Writing a valid complex JSON object may be difficult, and the custom parameter options help with that. You can replace each top-level key (e.g. inputs) in the parameters with a corresponding option (--inputs) followed by its value, again as a JSON object. A value that cannot be parsed as JSON will be assumed to be a string. If multiple values are given for the same option, or the same option is given multiple times, the value is assumed to be a list containing all the values. For providing a list of a single value, use the JSON notation (e.g. '["value"]'). Any dashes in the option name will be replaced by underscores (e.g. --corpus-name can be used to produce the corpus_name parameter).

In this manner, the above example can be written also as:

opusfilter-cmd filter --inputs wmt.fi.gz wmt.en.gz --outputs wmt_filtered.fi.gz wmt_filtered.en.gz --filters '[{"LengthFilter": {"unit": "word", "min_length": 1, "max_length": 100}}]'

Note that the filters still need to be defined as a complex JSON objects. The created configuration will be shown as YAML for easier checking and storing.

Available functions

Downloading and selecting data

opus_read

Read a corpus from the OPUS corpus collection (Tiedemann, 2012) using the OpusTools (Aulamo et al., 2020) interface.

Parameters:

• corpus_name: name of the corpus in OPUS
• source_language: language code for the source language
• target_language: language code for the target language
• release: version of the corpus in OPUS
• preprocessing: raw for untokenized and xml for tokenized segments
• src_output: output file for source language
• tgt_output: output file for target language
• suppress_prompts: false (default) prompts user to confirm before download, true to download without prompting

concatenate

Concatenate two or more text files.

Parameters:

• inputs: a list of input files
• output: output file

download

Download a file from URL.

Parameters:

• url: URL for the file to download
• output: output file

head

Take the first n lines from files.

Parameters:

• inputs: a list of input files
• outputs: a list of output files
• n: number of output lines

tail

Take the last n lines from files.

Parameters:

• inputs: a list of input files
• outputs: a list of output files
• n: number of output lines

Note: The memory requirement of tail is proportional to n. Use slice if you need all except the first n lines.

slice

Take slice of lines from files.

Parameters:

• inputs: a list of input files
• outputs: a list of output files
• start: start index (optional; default 0)
• stop: stop index (optional; default null)
• step: step size (optional; default 1)

Either start, stop, or both of them should be given. If stop is not given, reads until the end of the file.

split

Split files to two parts giving the approximative proportions as fractions.

Parameters:

• inputs: input file(s)
• outputs: output file(s) for selected lines
• outputs_2: output file(s) for the rest of the lines (optional)
• divisor: divisor for the modulo operation (e.g. 2 for splitting to equal sized parts)
• threshold: threshold for the output of the modulo operation (optional; default 1)
• compare: select files to use for hash operation (optional; default all or a list of indices)
• hash: select hash algorithm from pyhash (optional; default xx_64)
• seed: integer seed for the hash algorithm (optional; default 0)

Input files are processed line by line in parallel. If the condition hash(content) % divisor < threshold, where the content is a concatenation of the input lines and the hash function returns an integer, holds, the lines are written to the outputs. If the condition does not hold, and outputs_2 are defined, the lines are written there.

Compared to random splitting (see subset) or using the modulo operation on the line number, the benefit of the hash-based approach is that the decision is fully deterministic and based only on the content of the lines. Consequently, identical content always goes to the the same output file(s). For example, if you split a parallel corpus into test and training sets, and you can be sure that your test data does not contain exactly same samples as the training data even if the original data has duplicates.

The downside is that you need to be careful if you use several splits for the same data. The divisors used in consecutive splits should not themselves have common divisors, or the proportion of the data in the output files may be unexpected. Distinct prime numbers are good choices. Also setting a different seed value for the hash functions prevents the issue.

The compare parameter can be used to select which input files are used to generate the content for the hash function. For example, if you have source and target language files, and you want that the split depends only on the source or target sentence, set compare to [0] or [1], respectively.

subset

Take a random subset from parallel corpus files.

Parameters:

• inputs: input files
• outputs: output files for storing the subset
• size: number of lines to select for the subset
• seed: seed for the random generator; set to ensure that two runs select the same lines (optional; default null)
• shuffle_subset: shuffle the order of the selected lines for each language except for the first; can be used to produce noisy examples for training a corpus filtering model (optional; default false)

product

Create a Cartesian product of parallel segments and optionally sample from them.

Parameters:

• inputs: a list of input files lists
• outputs: a list of output files
• skip_empty: skip empty lines (optional, default true)
• skip_duplicates: skip duplicate lines per language (optional, default true)
• k: sample at most k random items per product (optional, default null)
• seed: seed for the random generator; set to ensure that two runs produce the same lines (optional; default null)

Can be used to combine parallel files of the same language that contain alternative translations or other meaningful variation (e.g. alternative subword segmenatations). For example, if you have the same text translated to language A by N translators and to language B by M translators, you can combine the N + M files into two files having N x M lines for each original line.

unzip

Unzip parallel segments joined in a single file into multiple files.

Parameters:

• input: input file
• outputs: a list of output files
• separator: a string separator in the input file

Can be used to split e.g. Moses-style (|||) or tab-separated parallel text files into parts.

write

Write a specified string into a file.

Parameters:

• output: output file
• data: input data to write to the output (converted to a string if not already)

Useful mostly for testing.

Preprocessing text

preprocess

Filter parallel data with a combination of filters.

Parameters:

• inputs: input files for segments to preprocess
• outputs: output files for preprocessed segments
• preprocessors: a list of preprocessors to apply; see below

The preprocessors parameter is a list of dictionaries, each representing one preprocessor. The top level should typically include a single key that defines the class name for the preprocessor (e.g. WhitespaceNormalizer). Additionally it can include a special key module for defining module name for custom preprocessors.

Under the class name there is a dictionary the defines the parameters of the preprocessors. The are mostly specific to the preprocessor class; see section Available preprocessors for ready-made preprocessors.

Filtering and scoring

remove_duplicates

Filter out duplicate lines from parallel corpus files.

Parameters:

• inputs: input file(s)
• outputs: output file(s)
• compare: select files for duplicate comparison (optional; default all or a list of indices)
• hash: select hash algorithm from pyhash (optional; default xx_64)
• overlap: remove overlap with a second set of files (optional; default null)

Duplicate filtering is recommended as a first step especially if you combine different corpus collections (e.g. data crawled from web) and cannot be sure that the same data sources have not been used in many of them.

The remove_duplicates function works for any number of files. The compare parameter can be used to select which input files are used to generate the key for duplicate comparison. For example, if you have source and target language files, and you want that each source or target sentence occurs only once, set compare to [0] or [1], respectively.

Instead of removing duplicates from a single set, the optional overlap argument can be used to remove all segments from inputs that match the segments in overlap. For example, you can remove exact duplicates of the test data from your training data.

Non-cryptographic hashing is used to reduce memory consumption for the case that the files are very large. The lines defined by the compare option are concatenated together, and the hash algorithm is applied on the result to produce the final key for storing the counts. You can use any of the hash algorithms implemented in the pyhash library. The default 64-bit XXHash algorithm should be fine for any practical data sizes, but if do not care about memory use and want to be extra sure there are no collisions, you can disable hashing by setting the hash parameter as empty string or null.

filter

Filter parallel data with a combination of filters.

Parameters:

• inputs: input files for segments to filter
• outputs: output files for filtered sentences
• filters: a list of filters to apply; see below
• filterfalse: Yield segment pairs that do not pass at least one of the filters (optional; default false)

The filters parameter is a list of dictionaries, each representing one filter. The top level should typically include a single key that defines the class name for the filter (e.g. LenghtFilter). Additionally it can include a special key module for defining module name for custom filters.

Under the class name there is a dictionary the defines the parameters of the filters. The are mostly specific to the filter class; see section Available filters for ready-made filters. An exception is a parameter name that is available for all filters. It has no effect for the filter function, but is useful for the score function below.

The output of the step is only those segment pairs that are accepted by all the filters, unless filterfalse is set true, in which case the output is the opposite (i.e., those segment pairs that are rejected by at least one filter).

score

Calculate filtering scores for the lines of parallel data.

Parameters:

• inputs: input files for segments to score
• output: output file for the scores
• filters: a list of filters to apply; see below

The filters are defined in the same manner as in the filter function. The possible accept threshold parameters of the filters do not have an effect in scoring; each filter simply outputs one or more numerical values.

The scores are written in the JSON Lines format: each line contains a single JSON object. The top level of the object contains class names for the filters. If there is only of filter of the specific class, and its score is a single number, the value of the score is simply below the class name. The the filter outputs more scores, they are represented in a list or dictionary. Typically there is one score for each segment (language) if there are multiple scores.

The filters may contain the same filter class multiple times so that the same filter can be used with different parameters (e.g. both words and characters as units for length-based filters). In this case, under the top-level filter class key there is another dictionary that separates the filter instances. The keys for the instances can be defined by using the name parameter that is available for all filters. If the name is not defined, the first filter of the class is given key "1", the second "2", and so on. (Note: make sure to give a name to either all or none of the filters, or at least do not manually give integers as names.)

The output can be used e.g. for analyzing the distribution of the scores or training a classifier for filtering. The JSON Lines data is easy to load as a pandas DataFrame using the json_normalize method.

Using score files

join

Join two or more score files.

Parameters:

• inputs: input files containing scores in JSON Lines format
• output: output file for joined scores
• keys: a list containing dictionary keys for each input file (optional; default null)

If the list of keys is provided, the input objects are inserted under the corresponding key. The objects can also be inserted deeper in a hierarchical score dictionary by using a key that has dot-separated parts. For example, x.y means setting key y under the key x. If the keys are not provided, or the key corresponding to the input file is null, output object will be updated with the input object and existing keys will be overwritten.

For example, if you have scores for the source and target sentences created by external tools (myscores.src and myscores.tgt containing one number per line), and you want to join them with an existing score file created by OpusFilter (scores.jsonl.gz), you can do it like this:

  - type: join
    parameters:
      inputs:
      - scores.jsonl.gz
      - myscores.src
      - myscores.tgt
      keys:
      - null
      - MyScore.src
      - MyScore.tgt
      output: scores-joined.jsonl.gz

Apart from the old scores from scores.jsonl.gz, each line should now contain {"MyScore": {"src": ..., "tgt": ...}}.

sort

Sort files based on score values.

Parameters:

• inputs: input files to sort
• outputs: sorted output files
• values: input file for values used in sorting
• reverse: true for descending sort (optional; default false)
• key: if values file contain JSON objects, use the key to select field (optional; default null)
• type: force type conversion for the value (optional; float, int, str, or default null)

The values file should contain one JSON object per line. If a line cannot be interpreted as a JSON object, it is read as a plain unicode string. Dots (.) in the key are interpreted as multiple get operations (e.g. x.y expects that there is key x under the key y). List items can be accessed with integer keys. The type conversion can be used e.g. for forcing numerical values to be compared as strings.

Training language and alignment models

train_ngram

Train a character-based varigram language model with VariKN (Siivola et al. 2007). Can be used for CrossEntropyFilter and CrossEntropyDifferenceFilter.

Parameters:

• data: input file name for training data
• model: output file name for the model
• parameters: training options for VariKN and tokenization
  • optdata: filename for optimization data (optional; default empty string "" = use leave-one-out estimation instead)
  • norder: limit model order (optional; default 0 = no limit)
  • dscale: model size scale factor (optional; smaller value gives a larger model; default 0.001)
  • dscale2: model size scaling during pruning step (optional; default 0 = no pruning)
  • arpa: output ARPA instead of binary LM (optional; default true)
  • use_3nzer: use 3 discounts per order instead of one (optional; default false)
  • absolute: use absolute discounting instead of Kneser-Ney smoothing (optional; default false)
  • cutoffs: use the specified cutoffs (optional; default "0 0 1"). The last value is used for all higher order n-grams.
  • mb: word-internal boundary marking (optional; default "")
  • wb: word boundary tag (optional; default "<w>")

Apart from the scale, cutoff, and order parameters the size of the model depends on the size of the training data. Typically you want to at least change the dscale value to get a model of a reasonable size. If unsure, start with high values, look at the number of the n-grams in the output file, and divide by 10 if it looks too small. The dscale2 option is useful mostly if you want to optimize the balance between the model size and accuracy at the cost of longer training time; a suitable rule of thumb is double the value of dscale.

The default boundary settings are suitable for character-based models and are not recommended to edit.

See VariKN documentation for details.

train_aligment

Train word alignment priors for eflomal (Östling and Tiedemann, 2016). Can be used in WordAlignFilter.

Parameters:

• src_data: input file for the source language
• tgt_data: input file for the target language
• parameters: training options for the aligment and tokenization
  • src_tokenizer: tokenizer for source language (optional; default null)
  • tgt_tokenizer: tokenizer for target language (optional; default null)
  • model: eflomal model type (optional; default 3)
• output: output file name for the priors

See WordAlignFilter for details of the training parameters.

Training and using classifiers

train_classifier

Train an sklearn classifier to produce a cleanness score for sentence pairs.

Parameters:

• training_scores: a file containing filter scores for training in JSON lines format produced with score function.
• criterion: criterion to be used in classifier optimization (valid options are CE, ROC_AUC, SSE, AIC and BIC)
• dev_scores: a file containing filter scores for training in JSON lines format produced with score function with and added item label added to each entry. label has value 1 for clean pairs and 0 for noisy pairs (optional; dev_scores is only used when the criterion is ROC_AUC)
• model_type: classifier model type selected from sklearn classifiers (default LogisticRegression)
• model_parameters: parameters for the sklearn classifier
• model: output model file
• features: the features given to the classifier to be trained on, defined as a list of filter names
  • ExampleFilter:
    • clean-direction: the direction that indicates higher cleanness (valid options are high and low)
    • quantiles: a dictionary the items of which (min, max and initial) specify the minimum, maximum and inital quantile value that are used in classifier optimization to select negative and positive training examples (default {'min': 0, 'max': 1, 'initial': 0.1})

The classifier is optimized by training multiple classifier model with the training data divided differently into positive and negative examples based on the quantile boundaries specified in each feature. The model that achieves the highest criterion score is then saved in the output file.

classify

Use a classifier model trained with train_classifier to assign a cleanness score or label to sentence pairs that have been scored with score.

Parameters:

• model: classifier model trained with train_classifier
• scores: scores of the sentence pairs to be classifed in JSON lines format produced with the score function
• output_probabilities: file to write the cleanness scores to, 1 is cleanest and 0 is noisiest (optional)
• output_labels: file to write the cleanness labels to, 1 is a clean and 0 is a noisy pair (optional)

The probabilities and labels are written to the output files line by line, corresponding to the scores on each line in scores.

Available filters

Length filters

LengthFilter

Filtering based on absolute segment lengths.

Parameters:

• min_length: minimum segment length (optional; default 1)
• max_length: maximum segment length (optional; default 100)
• unit: type of unit for calculating the lengths (optional; word for words or any whitespace-separated units, and character or char for characters; the default is word)
• pass_empty: if true, always accept if all segment lengths are zero (default false)

Returned scores are lengths for the source and target segment. In filtering, all segments have to be between the minimum and maximum length thresholds.

LengthRatioFilter

Filtering based on ratio of the segment lengths.

Parameters:

• threshold: threshold for the length ratio
• unit: type of unit for calculating the lengths (optional; word for words or any whitespace-separated units, and character or char for characters; the default is word)

Returned score is the higher length divided by the lower length, or infinity of either of the lengths are zero. In filtering, segment pairs is accepted of the ratio is below the given threshold.

AverageWordLengthFilter

Filtering based on average word lengths.

Parameters:

• min_length: minimum length (optional; default 2)
• max_length: maximum length (optional; default 20)
• pass_empty: if true, always accept if all segment lengths are zero (default false)

Returned scores are average words lengths for the segments. In filtering, all segments have to be between the minimum and maximum length thresholds.

LongWordFilter

Filtering based on maximum word length.

Parameters:

• threshold: maximum length (optional; default 40)

Returned score is the length of the longests words across the segments. The length has to below the threshold.

Script and language identification filters

CharacterScoreFilter

Filter segments based on what proportion of their alphabetic characters are in a given script. For a list of valid scripts, see e.g. https://www.regular-expressions.info/unicode.html

Parameters:

• scripts: scripts for input segments
• thresholds: minimum proportion of characters in a script (default 1)

Returned scores are proportions of valid characters in the segments. In filtering, all values have to be equal to or greater than the minimum thresholds.

LanguageIDFilter

Filter segments based on their language identification confidence scores.

Parameters:

• languages: expected languages for the segments
• id_method: language indentification method (langid for using the langid library, cld2 for using the cld2 library, or fasttext for using a fasttext model; the default is langid)
• thresholds: minimum identification confidence score for the segments (a single float or a list of floats per language)
• fasttext_model_path: path for a fasttext model (required only for the fasttext method; default null)

Returned scores are the language identification confidence scores from a given identification method for the segments. The scores range from 0 to 1. In filtering, all values have to be greater than the minimum thresholds. Negative threshold can be used to skip filtering for a language.

A pretrained fasttext model can be downloaded from https://fasttext.cc/docs/en/language-identification.html

Special character and similarity filters

HtmlTagFilter

Filter segments based on whether they contain HTML tags or not.

The returned scores are two boolean values indicating whether the segments contain HTML tags. In filtering, a segment pair is accepted if none of the segments contains HTML tags.

TerminalPunctuationFilter

Filter segments based on a penalty score with respect to the co-occurrence of therminal punctuation marks ('.', '...', '?', '!') in source and target segments (Vázquez et al., 2019). The score is formulated as follows: the initial score is the absolute difference in source and target terminal punctuation counts, the score is then incremented by the number of terminal punctuation beyond the first occurence in both segments, and finally, the score is updated with score=-log(score+1). The score of the greatest co-occurrence is 0 and smaller values indicate greater penalty.

This filter works only for bilingual input.

Parameters:

• threshold: minimum score threshold (default -2)

The returned score is a single terminal punctuation score. In filtering, the score has to equal to of be greater than the minimum threshold.

NonZeroNumeralsFilter

Filter segments based on a similarity measure of numerals between the segments with zeros removed (Vázquez et al., 2019). Non-zero numerals are extracted from all segments preserving the relative order of the numerals. The similarity score between the numeral sequences is produced with SequenceMatcher.ratio() from Python's difflib library.

Parameters:

• threshold: minimum score threshold (default 0.5)
• require_all: if True, all scores (for pairs of n segments) have to be reach threshold; otherwise at least one the ratios has to reach the threshold

The returned value is a list of similarity scores for all language pairs. For n-lingual input, the scores will include C(n, 2) values. In filtering, all pairwise scores has to equal to or be greater than the minimum threshold.

LongestCommonSubstringFilter

Filter segments based on similarity of the strings.

Parameters:

• threshold: filter segments if the similarity is equal or above the threshold (optional; default 0.9)
• require_all: if True, all ratios (for pairs of n segments) have to be below the threshold; otherwise at least one the ratios have to be below the threshold

Returned scores are ratios between the length of the longest common substring and the length of the shorter of the compared strings for all language pairs. For n-lingual input, the scores will include C(n, 2) values.

Language model filters

CrossEntropyFilter

Filter segments by n-gram language model probabilities.

Parameters:

• lm_params: a list of dictionaries for the parameters of the language models; see below
• score_type: select whether to calculate cross-entropy (entropy; default), perplixty (perplexity) or negative log-probability (logprob) scores
• thresholds: upper thresholds for scores when filtering (optional; default is 50.0 for all languages)
• low_thresholds: lower thresholds for scores when filtering (optional; default is no threshold)
• diff_threshold: upper threshold for absolute difference of source and target language scores when filtering (optional; default 10.0)
• score_for_empty: set score values manually for empty input pairs (default null)

Language model parameters for lm_params:

• filename: filename for the language model to use
• arpa: LM is in ARPA format instead of binary LM (optional; default true)
• unk: unknown token symbol (optional; default <UNK>, case sensitive)
• include_unks: include unknown tokens in perplexity calculations (optional; default false)
• ccs: list of context cues ignored in perplexity calculations (optional; default null)
• mb: morph boundary marking (optional; default "")
• wb: word boundary tag (optional; default "<w>")
• init_hist: ignore n first tokens after the end-of-sentence tag </s> in perplexity calculations (optional; default 2)
• interpolate: list of language models (in ARPA format) and interpolation weights (optional; default null)

See train_ngram for training the models. Note that the format, perplexity calculation, and boundary marking options should match the parameters used in model training; do not change them unless you know what you are doing.

Separate scores (entropy, perplexity, or negative log-probability) are returned for the source and target segment. In filtering, the segment pair is accepted if all values are below the respective thresholds, and their absolute differences are below the difference threshold.

CrossEntropyDifferenceFilter

Filter segments by using cross-entropy difference method by Moore and Lewis (2010).

Parameters:

• id_lm_params: a list of dictionaries for the parameters of the in-domain language models
• nd_lm_params: a list of dictionaries for the parameters of the non-domain language models
• thresholds: upper thresholds for scores when filtering (optional; default is 0.0 for all languages)
• score_for_empty: set score values manually for empty input pairs (default null)

For contents of the id_lm_params and nd_lm_params, see CrossEntropyFilter.

See train_ngram for training the models. Note that the format, perplexity calculation, and boundary marking options should match the parameters used in model training; do not change them unless you know what you are doing.

The filter returns difference between the in-domain LM and non-domain LM cross-entropy.

Alignment model filters

WordAlignFilter

Filter segments by word aligment scores using eflomal (Östling and Tiedemann, 2016).

Parameters:

• src_threshold: score threshold for source language (default 0)
• tgt_threshold: score threshold for target language (default 0)
• src_tokenizer: tokenizer for source language (optional; default null)
• tgt_tokenizer: tokenizer for target language (optional; default null)
• model: eflomal model type (optional; default 3)
• priors: priors for the aligment (optional; default null)
• score_for_empty: score values for empty input pairs (default -100)

A segment pair is accepted if scores for both directions are lower than the corresponding thresholds.

The only tokenizer supported at the moment is the fast-mosestokenizer that re-implements the tokenizer script from the Moses toolkit. To enable it, provide a tuple containing moses and an appropriate two-letter language code, e.g. [moses, en] for English.

The eflomal model types are 1 for IBM1, 2 for IBM1 + HMM, and 3 for IBM1 + HMM + fertility. See https://github.com/robertostling/eflomal for details.

See train_aligment for training priors. Compatible tokenizer and model parameters should be used.

Custom filters

You can also import your own filters by defining the module key in the filter configuration entries.

The custom filters should inherit the abstract base class FilterABC from the opusfilter package. They should implement two abstract methods: score and accept.

The score method is a generator that takes an iterator over tuples of parallel sentences, and yields a score object for each pair. The score may either be a single number, or if multiple score values need to be yielded, a dictionary that has the numbers as values.

The accept method takes a single output yielded by the score method, and returns whether the sentence pair should be accepted based on the score.

If the filter requires any parameters (e.g. score thresholds for the accept method), the class should implement also the __init__ method. Arbitrary keyword arguments should be accepted (with **kwargs), and the __init__ method of the base class (FilterABC) should be called in the end with the remaining keyword arguments. The keyword argument name is reserved for giving names to the filters.

Based on the score and accept methods, the abstract class FilterABC implements the following three generators that take iterator over segment pairs as input:

• decisions yields results of the accept method
• filter yields only accepted segments
• filterfalse yields only rejected segments

These should not be redefined except for a good reason.

The example below shows code for simple filter that calculates the proportion of uppercase letters in the sentences, and accepts the pair only if all sentences have less than 50% (or given threshold) of uppercase characters:

import opusfilter

class UppercaseFilter(opusfilter.FilterABC):

    def __init__(self, threshold=0.5, **kwargs):
        self.threshold = threshold
        super().__init__(**kwargs)

    def uppercase_ratio(self, sentence):
        length = len(sentence)
        if length > 0:
            return sum(1 for char in sent if char.isupper()) / length
        return 0

    def score(self, pairs):
        for pair in pairs:
            yield [self.uppercase_ratio(sentence) for sentence in pair]

    def accept(self, score):
        return all(ratio < self.threshold for ratio in score)

Assuming that the above code is in a module named customfilter in the Python evironment (e.g. save the code as customfilter.py and add the directory that contains it to PYTHONPATH environment variable), it can be selected in the filter configurations as follows:

steps:

  ...

  - type: filter
    parameters:

      ...

      filters:

        - UppercaseFilter:
            threshold: 0.5
          module: customfilter

Available preprocessors

Tokenizer

Tokenize parallel texts.

Parameters:

• tokenizer: tokenizer type
• languages: a list of language codes for the tokenizer
• options: a dictionary of tokenizer options (optional)

Currently there is only one type of tokenizer available: moses (uses the fast-mosestokenizer package). Options are passed to the mosestokenizer.MosesTokenizer class; see its documentation for the available options.

The list of language codes should match to the languages of the input files given in the preprocess step.

Detokenizer

Detokenize parallel texts.

Parameters:

• tokenizer: tokenizer type
• languages: a list of language codes for the detokenizer
• options: a dictionary of tokenizer options (optional)

Currently there is only one type of detokenizer available: moses (uses the fast-mosestokenizer package). Options are passed to the mosestokenizer.MosesTokenizer class; see its documentation for the available options.

The list of language codes should match to the languages of the input files given in the preprocess step.

WhitespaceNormalizer

Normalizes whitespace characters.

The whitespace normalization consists of two steps: First, any sequence of one or more unicode whitespace characters (as matched by \s in the re standard library) are replaced by a single space character. Second, any leading and trailing whitespace is removed.

RegExpSub

Run a sequence of arbitrary regular expression substitutions.

Parameters:

• patterns: a list of patterns to use by default
• lang_patterns: a dictionary or list of patterns to use for specific languages

Multiple substitutions are applied in the given order. The default patterns are replaced with language-specific patterns when the corresponding index (starting from 0) is found in the lang_patterns dictionary. The lang_patterns argument may also be a list, if you e.g. want to use separate patterns for all languages.

The substitution patterns are 4-tuples containing the regular expression, replacement, count (0 = substitute all) and flags (list of flag constants in the re library, e.g. ["I", "A"]). The regular expressions are first compiled with re.compile, and then the substitutions are applied with re.sub.

Custom preprocessors

Similarly to filters, You can import your own preprocessors by defining the module key in the filter configuration entries.

The custom preprocessors should inherit the abstract base class PreprocessorABC from the opusfilter package, and implement the abstract methods process. The process method is a generator that takes an iterator over segments and yields preprocessed (modified) segments of text. It also has an additional argument, f_idx, which is the index of the current file being processed by the preprocess step. This argument enables preprocessing parallel files in the same step even if the preprocessing options (such as language code for a tokenizer) varies.

Other tools

Apart from the main opusfilter and opusfilter-cmd scripts, the package also provides some analysis tools.

opusfilter-diagram

Draws a diagram (a directed acyclic graph) from OpusFilter configuration file using the graphviz library.

opusfilter-diagram [--rankdir {TB,LR}] FILE FILE

The --rankdir option changes the direction of the graph from left-to-right (default) to top-to-bottom. If the output file ends with .dot, the raw dot format is used; otherwise the graph is rendered to the format indicated by the extension (e.g. PDF or PNG).

opusfilter-duplicates

This is a simple script based on the remove_duplicates function, that instead of filtering the data, prints out statistics of the duplicate entries. You can either provide a single corpus (as one monolingual file or multiple parallel files) for calculating the number of duplicates in it, or two corpora for calculating the overlap between them. The syntax for the opusfilter-duplicates is:

opusfilter-duplicates [--overlap FILE [FILE ...]] [--hash HASH] [--letters-only] [--lowercase] FILE [FILE ...]

The options are essentially the same as for remove_duplicates.

opusfilter-scores

This is a tool that can be used to calculate and plot statistics from scores produced by the score function. The tool has several subcommands that all take the JSON Lines score file as the input, and either print or plot the output:

• list: Print score column names
• describe: Print basic score statistics
• corr: Plot score correlation matrix
• hist: Plot score histograms
• scatter-matrix: Plot scatter matrix for scores
• values: Plot score values by line number

opusfilter-test

This is a simple script based on the filter function that can be used to calculate the amount of segments that the given filter(s) would remove from the parallel data, and optionally output the to-be-removed segments.

The syntax for the opusfilter-test is:

opusfilter-test [--yaml FILE] [--add CLASS JSON] [--removed FILE] FILE [FILE ...]

The filters to test can be defined either from a YAML file (--yaml) using a similar definition as the filters parameter for the filter function, or adding the one by one with the --add option, which takes the filter class as the first argument and filter parameters in as a JSON object as the second argument. For default filter parameters, an empty dictionary ('{}') should be provided.

The scripts first calculates the total number of segments in the input files, and then runs the filters on them one by one. The number and proportion of removed segments is printed. In addition, it is possible to write the removed segments to a file in JSON Lines format (--removed).

How to contribute

Questions, bug reports, and feature wishes are welcome in the GitHub issues page. We are also happy to consider pull requests. There are a few rules for pull requests:

• Make a pull request to the develop branch instead of master.
• The code should support at least Python versions from 3.6 to 3.8.
• Please follow PEP 8. Exception: The maximum line length is 127 characters instead of 79.
• Especially for new features, please include test cases for unit testing.

PEP 8 compatibility can be checked with flake8. Install it e.g. via pip and run flake8 opusfilter/ in the project root.

The unit tests are located in the tests directory. To run them, install pytest and run python -m pytest tests/ in the project root. Also nosetests should work, if you have VariKN and eflomal set up as instructed (pytest skips the respective tests if not).

GitHub workflows defined in the project run automatically flake8 checks and unit testing with pytest using Python 3.6, 3.7, and 3.8.

Especially for larger contributions, consider using a code analysis tool like Pylint. Install it e.g. via pip, run pylint opusfilter/ in the project root and fix at least everything that is simple to fix in the new code (note that the current code yields a few warnings from pylint).