By Mike Pham and Jackson Lee

Academic publication:

Pham, Mike, & Jackson L. Lee. 2018. Mincing words: Balancing recovery and deletion in word truncation. Glossa: A Journal of General Linguistics, 3(1), 36. DOI: http://doi.org/10.5334/gjgl.269

This repo contains code, datasets, and results associated with this publication.

• main.py: Python script for running the different truncation models over the data files in the data/ folder

• data/: directory for datasets

  • pt_br_full.txt: Brazilian Portuguese lexicon (8.7 MB)
  • gold_standard.txt: gold standard nouns with attested truncation, all annotated with predictions from models such as the binary foot models

• plots_for_words/: directory for output plots for individual words for log(left-complete counts) and log(right-complete counts)

• results/: directory for output files

  • error details in CSV
  • error distribution boxplot
  • evaluation
  • L- and R-complete counts for individual in LaTeX and PDF
  • R code for making individual plots for test words

• readme.md: this readme file

NumPy, Pandas, and Seaborn are required to run main.py. For reproducibility, the exact versions we use are pinned down in requirements.txt, and you can install these dependencies by pip install -r requirements.txt. We are using Python 3.6.3.

Optional -- If the following commands are recognized in your path environment:

• Rscript: plot_word.R (generated by main.py) is run and the plots for individual words are saved in plots_for_words/.

• xelatex: individual_word_details.tex (generated by main.py) is compiled and the resultant PDF is saved in results/.

Rscript comes from R, whereas xelatex comes from a LaTeX distribution (such as texlive-full).

Download this repository to your local drive by one of these two methods:

• Download and unzip https://github.com/jacksonllee/BP-truncation/archive/master.zip

• Clone this repository:

  $ git clone https://github.com/jacksonllee/BP-truncation.git
  $ cd BP-truncation
  

main.py can take optional arguments. The argument -h brings up the help page with details of these arguments:

$ python main.py -h
usage: main.py [-h] [-f] [-l] [-r] [-d] [-x LEXICON] [-g GOLDSTANDARD]

Modeling truncation in Brazilian Portuguese, by Mike Pham and Jackson Lee

optional arguments:
  -h, --help            show this help message and exit
  -f, --freqtoken       Use token frequencies in lexicon (default: False)
  -l, --latex           Compile the output LaTeX file (default: False)
  -r, --run_r_script    Run R script (default: False)
  -d, --digraphsfixed   Change orthographic digraphs into monographs (default:
                        False)
  -x LEXICON, --lexicon LEXICON
                        Lexicon file (default: data/pt_br_full.txt)
  -g GOLDSTANDARD, --goldstandard GOLDSTANDARD
                        Gold standard file (default: data/gold_standard.txt)

The sample output files included in this repository are generated by this command:

$ python main.py -lr

This command has most of the default settings as described in the help page shown above, except that R code is run and LaTeX compilation is triggered. If you don't want to run R and LaTeX, simply run python main.py with no other arguments.

If, for instance, you'd like to make use of word token frequency information in the models that involve right-completes and left-completes, you should run python main.py -flr (still running R and LaTeX). All output files for this command bear the suffix "-tokenfreq".

To activate orthographic digraph replacements, run python main.py -dlr. All output files are suffixed by "-nodigraphs".

The arguments --lexicon and -goldstandard may be used to override the default lexicon and gold standard files, respectively. See the sections below on their file format.

The lexicon file data/pt_br_full.txt is a plain text file where each line begins with a word, and then a space, and finally the frequency count of that word. Here are the first ten lines of the the lexicon file:

que 12021478
não 9712854
o 9578625
de 8089861
a 7188507
é 6843557
você 6211533
e 5863939
eu 5741437
um 4589127

This lexicon file is from here (released with an MIT license), which in turn derived the lexicon and frequency counts from movie subtitles. The data is therefore highly representative of the spoken language.

The gold standard file is a plain text file where each line has one original (untruncated) word, and then a space/tab, and finally the truncated form (TF) of that original word; the TF is just for reference and is not used by main.py in any way. The original word is annotated by the following symbols:

• |: where the true truncation point is for forming the truncated stem (TS). For instance, the TS for adrenalina (see below) is adren.
• $: where the truncation point is as predicted by the binLR model.
• #: where the truncation point is as predicted by the binRL model.

Here are the first ten lines of the default gold standard file in this repository (data/gold_standard.txt):

adr$en|#alina	adrena
an$alf|#abeto	analfa
bat$#er|ista	batera
bel$e|z#a	belê
berm|$ud#a	bermas
bij$#u|teria	biju
bis|$av#ó	bisa
bob|$eir#a	bobis
bot$equ|#im	boteco
burg|$#ês	burga

MIT License