PanPhon

Citing PanPhon

If you use PanPhon in research, please cite the following paper:

David R. Mortensen, Patrick Littell, Akash Bharadwaj, Kartik Goyal, Chris Dyer, Lori Levin (2016). "PanPhon: A Resource for Mapping IPA Segments to Articulatory Feature Vectors." Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 3475–3484, Osaka, Japan, December 11-17 2016.

Or in BibTeX:

@inproceedings{Mortensen-et-al:2016,
  author    = {David R. Mortensen and
               Patrick Littell and
               Akash Bharadwaj and
               Kartik Goyal and
               Chris Dyer and
               Lori S. Levin},
  title     = {PanPhon: {A} Resource for Mapping {IPA} Segments to Articulatory Feature Vectors},
  booktitle = {Proceedings of {COLING} 2016, the 26th International Conference on Computational Linguistics: Technical Papers},
  pages     = {3475--3484},
  publisher = {{ACL}},
  year      = {2016}
}

This package constitutes a database of segments in the International Phonetic Alphabet (IPA) and their equivalents in terms of (articulatory) phonological features. They include both data files and the tool generate_ipa_all.py, which allows the application of rules for diacritics and modifiers to collections of IPA characters, data files, and configuration/rule files and well as the tool validate_ipa.py, which checks Unicode IPA text from STDIN for well-formedness.

Python API for Accessing Phonological Features of IPA Segments

The FeatureTable class in the panphon module provides a straightforward API that allows users and developers to access the segment-feature relationships encoded in the IPA database consisting of the files panphon/data/ipa_bases.csv and diacritic_definitions.yml.

Note that a new API using faster, more rational, data structures (see the Segment class in panphon.segment) has been introduced. The old API is still available in the module _panphon.

>>> import panphon
>>> ft = panphon.FeatureTable()
>>> ft.word_fts(u'swit')
[<Segment [-syl, -son, +cons, +cont, -delrel, -lat, -nas, 0strid, -voi, -sg, -cg, +ant, +cor, -distr, -lab, -hi, -lo, -back, -round, -velaric, 0tense, -long]>, <Segment [-syl, +son, -cons, +cont, -delrel, -lat, -nas, 0strid, +voi, -sg, -cg, -ant, -cor, 0distr, +lab, +hi, -lo, +back, +round, -velaric, 0tense, -long]>, <Segment [+syl, +son, -cons, +cont, -delrel, -lat, -nas, 0strid, +voi, -sg, -cg, 0ant, -cor, 0distr, -lab, +hi, -lo, -back, -round, -velaric, +tense, -long]>, <Segment [-syl, -son, +cons, -cont, -delrel, -lat, -nas, 0strid, -voi, -sg, -cg, +ant, +cor, -distr, -lab, -hi, -lo, -back, -round, -velaric, 0tense, -long]>]
>>> ft.word_fts(u'swit')[0].match({'cor': 1})
True
>>> ft.word_fts(u'swit')[0] >= {'cor': 1}
True
>>> ft.word_fts(u'swit')[1] >= {'cor': 1}
False
>>> ft.word_to_vector_list(u'sauɹ', numeric=False)
[[u'-', u'-', u'+', u'+', u'-', u'-', u'-', u'0', u'-', u'-', u'-', u'+', u'+', u'-', u'-', u'-', u'-', u'-', u'-', u'-', u'0', u'-'], [u'+', u'+', u'-', u'+', u'-', u'-', u'-', u'0', u'+', u'-', u'-', u'0', u'-', u'0', u'-', u'-', u'+', u'+', u'-', u'-', u'+', u'-'], [u'+', u'+', u'-', u'+', u'-', u'-', u'-', u'0', u'+', u'-', u'-', u'0', u'-', u'0', u'+', u'+', u'-', u'+', u'+', u'-', u'+', u'-'], [u'-', u'+', u'-', u'+', u'-', u'-', u'-', u'0', u'+', u'-', u'-', u'+', u'+', u'-', u'-', u'+', u'-', u'+', u'+', u'-', u'0', u'-']]

Summary of Functionality

Operations on feature sets and segments

The FeatureTable class includes a broad range of operations on features and segments (consonants and vowels).

Converting words to feature arrays

The panphon class includes the function word2array which takes a list of feature names (as a list of strings) and a panphon word (from FeatureTable().word_fts()) and returns a NumPy array where each row corresponds to a segment in the word and each column corresponds to one of the specified features. Basic usage is illustrated in the following example:

>>> import panphon
>>> ft=panphon.FeatureTable()
>>> ft.word_array(['syl', 'son', 'cont'], u'sɑlti')
array([[-1, -1,  1],
       [ 1,  1,  1],
       [-1,  1,  1],
       [-1, -1, -1],
       [ 1,  1,  1]])

Segment manipulations

The Segment class, defined in the panphon.segment module, is used to represent analyzed segments in the new panphon.FeatureTable class (code found in panphon.featuretable). It provides performance advantages over the old list-of-tuples representation, is more Pythonic, and provides additional functionality.

Construction

There are two main ways to construct a Segment object:

>>> from panphon.segment import Segment
>>> Segment(['syl', 'son', 'cont'], {'syl': -1, 'son': -1, 'cont': 1})
<Segment [-syl, -son, +cont]>
>>> Segment(['syl', 'son', 'cont'], ftstr='[-syl, -son, +cont]')
<Segment [-syl, -son, +cont]>

In both cases, the first argument passed to the constructor is a list of feature names. This specifies what features a segment has as well as their canonical ordering (used, for example, when a feature vector for a segment is returned as a list). The second argument is a dictionary of feature name-feature value pairs. The feature values are integers from the set {-1, 0 1} (equivalent to {-, 0, +}). This dictionary can be omitted if the keyword argument ftstr is included. This string is scanned for sequences of (-|0|+)(\w+), which are interpreted as name-value (really value-name) pairs.

Basic querying and updating

Segment objects implement a dictionary-like interface for manipulating key-value pairs:

>>> a = Segment(['syl', 'son', 'cont'], {'syl': -1, 'son': -1, 'cont': 1})
>>> a
<Segment [-syl, -son, +cont]>
>>> a['syl']
-1
>>> a['son'] = 1
>>> a
<Segment [-syl, +son, +cont]>
>>> a.update({'son': -1, 'cont': -1})
>>> a
<Segment [-syl, -son, -cont]>

Set operations

The match method asks whether the Segment object on which it is called has a superset of the features contained in the dictionary passed to it as an argument. The >= operator is an alias for the match method:

>>> a = Segment(['syl', 'son', 'cont'], {'syl': -1, 'son': -1, 'cont': 1})
>>> a.match({'son': -1, 'cont': 1})
True
>>> a.match({'son': -1, 'cont': -1})
False
>>> a >= {'son': -1, 'cont': 1}
True
>>> a >= {'son': 1, 'cont': 1}
False

The intersection method asks which features the Segment object on which it is called and the dictionary or Segment object that is passed to it as an argument share. The & operator is an alias for the intersection method:

>>> a = Segment(['syl', 'son', 'cont'], {'syl': -1, 'son': -1, 'cont': 1})
>>> a.intersection({'syl': -1, 'son': 1, 'cont': -1})
<Segment [-syl]>
>>> a & {'syl': -1, 'son': 1, 'cont': -1}
<Segment [-syl]>

Vector representations

Segment objects can return their vector representations, either as a list of integers or as a list of strings, using the numeric and string methods:

>>> a = Segment(['syl', 'son', 'cont'], {'syl': -1, 'son': -1, 'cont': 1})
>>> a.numeric()
[-1, -1, 1]
>>> a.strings()
[u'-', u'-', u'+']

Fixed-width pattern matching

The FeatureTable classes also allows matching of fixed-width, feature-based patterns.

Sonority calculations

The Sonority class has methods for computing sonority scores for segments.

Feature edit distance

The Distance class includes methods for calculating edit distance, both in which the cost of substitutions is based upon Hamming distance between the feature vectors and in which the cost of substitutions are based upon edit weights for individual features.

The `panphon.distance` Module

This module includes the Distance class, which includes various methods for computing the distance between Unicode IPA strings, including convenience methods (really "inconvenience methods") for computing Levenshtein distance, but--more importantly--methods for computing similarity metrics related to articulatory features. The methods include the following:

panphon.distance.Distance .levenshtein_distance

A Python implementation of Levenshtein's string edit distance.

panphon.distance.Distance .fast_levenshtein_distance

A C implementation of Levenshtein's string edit distance. Unsurprisingly, must faster than the former.

panphon.distance.Distance .dolgo_prime_distance

Fast Levenshtein distance after collapsing segments into an enhanced version of Dolgopolsky's equivalence classes.

panphon.distance.Distance .feature_edit_distance

Edit distance where each feature-edit has cost 1/22. Edits from unspecified to specified cost 1/44.

panphon.distance.Distance .hamming_feature_edit_distance

Edit distance where each feature-edit has cost 1/22. Edits from unspecified to specified also cost 1/22. Insertions and substitutions each cost 1.

panphon.distance.Distance .weighted_feature_edit_distance

Edit distance where costs of feature edits are differently weighted depending on their class and subjective variability. All of these methods have the same interface and patterns of usage, demonstrated below:

>>> import panphon.distance
>>> dst = panphon.distance.Distance()
>>> dst.dolgo_prime_distance(u'pops', u'bobz')
0
>>> dst.dolgo_prime_distance(u'pops', u'bobo')
1

Scripts

The `generate_ipa_all.py` Script

Summary

This small Python program allows the user to apply sets of rules, defined in YAML, for adding diacritics and modifiers to IPA segments based upon their phonological features.

Usage

To generate a segment features file (ipa_all.csv), use the following in the panphon data directory:

$ generate_ipa_all.py ipa_bases.csv -d diacritic_definitions.yml -s sort_order.yml ipa_all.csv

Note that this will overwrite your existing ipa_all.csv file, which is often what you want.

Data Files

This package also includes multiple data files. The most important of these is ipa_bases.csv, a CSV table of IPA characters with definitions in terms of phonological features. From it, and the diacritics_definitions.yml file, the comprehensive ipa_all.csv is generated.

IPA Character Databases: `ipa_bases.csv` and `ipa_all.csv`

The IPA Character Table is a CSV file in which the first column contains an IPA segment and each subsequent column contains a phonological feature, coded as +, -, or 0. The features are as follows:

syl: syllabic
son: sonorant
cons: consonantal
cont: continuant
delrel: delayed release
lat: lateral
nas: nasal
strid: strident
voi: voice
sg: spread glottis
cg: constricted glottis
ant: anterior
cor: coronal
distr: distributed
lab: labial
hi: high (vowel/consonant, not tone)
lo: low (vowel/consonant, not tone)
back: back
round: round
velaric: velaric airstream mechanism (click)
tense: tense
long: long

Inspiration for the data in these tables is drawn primarily from two sources: the data files for HsSPE and Bruce Hayes's feature spreadsheet. It has since be re-rationalizeds based on evidence from a wide range of sources. As such, any special relationship to these prior inspirations has been eliminated.

The IPA Character Table ipa_bases.csv is intended to contain all of the unmodified segmental symbols in IPA, as well as all common affricates and dually-articulated segments. It is meant to be augmented by the rule-driven application of diacritics and modifiers.

Configuration and Rule Files

This package includes two files that control the behavior of generate_ipa_all.py. These are intended to be edited by the end user. Both are written in YAML, a standardized, human-readable and human-editable data serialization language.

Sort Order Specification: sort_order.yml

The file sort_order.yml controls the ordering of segments in the output of the Diacritic Application Tool. It is a sequence of maps, each with two fields:

name The name of a feature.
reverse A boolean value (True or False) specifying whether sorting on the named feature will be reversed or not.

The order of the features determines the priority of sorting.

The file sort_order_schema_.yml is a Kwalify schema that defines a syntactically valid sort order file.

Diacritic and Modifier Rules: diacritic_definitions.yml

The most important file for controlling the Diacritic Application Tool is diacritic_definitions.yml, a list of rules for applying diacritics and modifiers to IPA segments based on their phonological features. It has two sections, diacritics and combinations. Each of these is the key to an item in the top-level map.

Diacritics

The key diacritics points to a list of rules for applying diacritics/modifiers to bases. Each rule is a map with the following fields:

marker. The Unicode diacritic or modifier.
name. The name of the series derived from applying the diacritic or modifier.
postion. The position of the diacritic relative to the base (pre or post).
conditions. A list of conditions, each of them consisting of an associative array of feature specifications, under which the diacritic or modifier will be applied to a base.
exclude. A sequence of segments to be excluded from the application of the diacritic/modifier even if they match the conditions.
content. The feature specifications that will be set if the diacritic or modifier is applied, given as a map of feature specifications.

Combinations

The key combinations likewise points to a list of rules for combining the rules in diacritics. These rules are very simple, and include only the following fields:

name. The name of the combined category.
combines. A sequence of the names of the rules from diacritics that are to be combined.

The file diacritic_definitions_schema.yml is a Kwalify schema that defines a syntactically valid diacritics definition file.

dmort27/panphon