Autologistic model

About

This is an implementation of Autologistic model in Python for the experiments as shown in the Section 5 of the paper "Contrasting Vertical and Horizontal Transmission of Typological Features" in COLING 2016.

This software is released under the MIT License, see LICENSE.

Requirements

• Python 3.5
• NetworkX
• Numpy
• WALS data (CSV) should be downloaded to data directory

Usage overview

Print WALS features to process

• You can see the indices of WALS features to process by:

$ python ./scripts/experiment.py print-features
Index	Feature name
----------------------------------------------
0	101A Expression of Pronominal Subjects
1	112A Negative Morphemes
2	116A Polar Questions
...

Run experiments

• To run the experiments, use scripts/experiment.py with following arguments.

$ python ./scripts/experiment.py exp [first index of WALS feature to process] [last index of WALS feature to process] [experiment type]

• You can specify the path of input files and the directory to output. See more information by help:

$ python ./scripts/experiment.py exp --help

Missing value imputation

Run

• To run the missing value imputation experiment like:

$ python ./scripts/experiment.py exp 0 10 mvi

• It outputs the results to the directory (output directory)/mvi/(number of the split)/(feature name).json.

Output

Output JSON file is like:

{
    "feature": "101A Expression of Pronominal Subjects",
    "lambda": 0.002049842660738542,
    "theta": 0.009065555215474861,
    "beta": [
        -0.09558277923882781,
        0.9942113364129587,
        -0.5985722330582819,
        -0.15532445575421136,
        -0.24775781040167438,
        -0.5600623232969136
    ],
    "accuracy": {
        "baseline": 0.4788732394366197,
        "proposed": 0.5915492957746479
    },
    "estimate_results": [ ... ]
}

estimate_results field contains each estimated or original feature value of each language like:

{
   "value": "2 Subject affixes on verb",
   "is_hidden_estimated": true,
   "language": "Aja",
   "original": "2 Subject affixes on verb"
},

if is_hidden_estimated is true, it denotes the original feature value (original) of the language ("Aja" in example) is hidden to the missing value imputation test and estimated as value by the model.

if is_original is true, it denotes the feature value (value) of the language is the original one, otherwise, is estimated by the model.

Parameter estimation

Run

• To run the parameter estimation experiment like:

$ python ./scripts/experiment.py exp 0 10 param

• It outputs the results to the directory (output directory)/param/(feature name).json.

Output

Output JSON file is like:

{
    "feature": "116A Polar Questions",
    "lambda": 0.010144166778914547,
    "theta": 0.013618914159751434,
    "beta": [
        0.5496708416159658,
        0.2262238700953768,
        -0.07882853251432281,
        -0.07955149753492666,
        -0.09035882308183195,
        0.2525423664878199,
        -0.09236543248194536
    ],
    "estimate_results": [ ... ]
}

estimate_results field contains each estimated or original feature value of each language like:

{
    "is_hidden_estimated": false,
    "language": "Ajagbe",
    "is_original": false,
    "value": "1 Question particle"
},

if is_original is true, it denotes the feature value (value) of the language ("Ajagbe" in example) is the original, otherwise, is estimated by the model.