Computational Construction Grammar, or c2xg, is a Python package for the unsupervised learning of CxG representations along with tools for vectorizing these representations for computational tasks. Why CxG? Constructions are grammatical entities that support a straight-forward quantification of linguistic structure.
The first task is to initialize an instance of c2xg:
from c2xg import C2xG
CxG = C2xG(data_dir, language, s3, s3_bucket, nickname, model, zho_split)
This references the following variables:
data_dir (str) Either the path to the main data folder or, if using s3, the prefix name
language (str) Currently supports ara, deu, eng, fra, por, rus, spa, zho
s3 (boolean) If True, operates on data stored on an s3 bucket
s3_bucket (str) If s3 == True, this contains the name of the s3 bucket to connect to (assumes credentials are set in AWS-CLI)
nickname (str) If learning a new model, this creates a new namespace for saving temp files
model (str) If provided, loads a specific model; otherwise loads default grammar for the language
zho_split(boolean) Chinese text needs to be segmented into words; if False, the input text is already split
The Parse method takes a text or string and returns a sparse matrix with construction frequencies.
vectors = CxG.parse_return(input, mode, workers)
This references the following settings:
input (str / list of [strs]) The input, either filenames or texts, specified using **mode**
mode (str) "files" assumes input as filenames; "lines" takes a list of texts
workers (int) Number of processes to use
A generator function is also available.
for vector in CxG.parse_yield(input, mode, workers):
print(vector)
This references the following settings:
input (str / list of [strs]) The input, either filenames or texts, specified using **mode**
mode (str) "files" assumes input as filenames; "lines" takes a list of texts
The second task is to learn a new CxG. Most users will not need to train a new model.
CxG.learn(nickname, cycles, cycle_size, freq_threshold, beam_freq_threshold, turn_limit, workers, mdl_workers)
This references the following variables:
nickname (str): Creates a unique namespace for saving temp files
cycles (int): Number of unique folds to use; final grammars are merged across fold-specific grammars
cycle_size (tuple of ints): The number of files to use for optimization data, for candidate extraction, and for background data
freq_threshold (int): The number of occurrences required before a candidate construction is considered
beam_freq_threshold (int): The frequency threshold used when searching for the best beam search parameters
turn_limit (int): For the tabu search, how many turns to evaluate for making each move (x3 for the direct tabu search)
workers (int): Number of processes to use; not every stage distributes well.
mdl_workers (int): Number of processes to use for evaluating MDL during construction search; uses more memory
Each learning fold consists of three tasks: (i) estimating association values from background data; this requires a large amount of data (e.g., 20 files); (ii) extracting candidate constructions; this requires a moderate amount of data (e.g., 5 files); (iii) evaluating potential grammars against a test set; this requires a small amount of data (e.g., 1 file or 10 mil words).
The freq_threshold is used to control the number of potential constructions to consider. It can be set at 20. The turn limit controls how far the search process can go. It can be set at 10.
Installation files will be provided in the c2xg/whl directory.
pip install <whl file>
This package works best with properly-compiled (e.g., Intel Python versions of) the following dependencies:
Python 3.6.3
cytoolz 0.9.0.1
gensim 3.4.0
numpy 1.14.0
scipy 1.0.0
scikit-learn 0.19.1
numba 0.35.0
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent