philipperemy/urban-dictionary-transformers

Urban Dictionary Transformers

Transformers applied to Urban Dictionary for fun. 🍑 🍆

Model name	Dataset	Num samples (train)	Num samples (test)	Cross Entropy (test)	Num Epochs	Download model	Results (test)
Facebook BART Base	Urban Dictionary	1,190,865	62,677	1.5180	10	Click (zip)	Click (json)

The checkpoints are compatible with simpletransformers.

Goal

From the description we try to guess the word.

Input: One who has a mania for music.
Target: melomaniac

Some Results

Isn't that good?

{
    "target": "Oh, my back!",
    "input": "Used when something unpleasant or distasteful is about to happen.",
    "predicted": "Shit is goin down"
},
{
    "target": "ho",
    "input": "A bitch.",
    "predicted": "Kanye West"
},
{
  "target": "tequila suicide",
  "input": "An alternative to the [tequila slammer]. Snort the salt, down the shot, squeeze lime in the eye.\r\n\r\n*WARNING* May cause nasal bleeding, bloodshot eyes and being totally [muntered].",
  "predicted": "tequila sniper"
},
{
  "target": "A.F.I.D.",
  "input": "asian female immigrant driver",
  "predicted": "a.f.i.d."
},
{
  "target": "ecuador",
  "input": "The best country in the world.  It is home to a superior race of men who is beginning world domination in Rochester, Minnesota and in Rogers, (the whole) Arkansas.",
  "predicted": "America"
},
{
  "target": "rich",
  "input": "Someone who has a lot of money",
  "predicted": "rich"
},
{
  "target": "high-altitude fuck",
  "input": "1. An instance of sexual intercourse performed in a mountainous region that is usually accompanied by an intense over-bearing shame inflicted on the participants by the conservative views of the society in which they live.  It is also usually directly preceded by stream fishing.  ",
  "predicted": "mountain fuck"
},
{
  "target": "old girl",
  "input": "Your mom. Mother",
  "predicted": "old girl"
},

How to?

Download the model checkpoints

# Download the ZIP https://drive.google.com/drive/folders/1dI3o4yTBWHv5s15LxowCY3FDtFyGWkgO.
unzip urban-checkpoints-20210421.zip

Install the dependencies

pip install -r requirements.txt

Download the data (only for training or full evaluation)

wget https://s3-eu-west-1.amazonaws.com/pfigshare-u-files/8010727/UT_raw_plus_lowercase.7z
7z x UT_raw_plus_lowercase.7z # sudo apt-get install p7zip-full
md5sum words.json # 5aca6e9bb2c8b9eb7fc5ebc9f947ec33  words.json

Full evaluation

export CUDA_VISIBLE_DEVICES=1; python urban.py --resume_from urban-checkpoints-20210421 --eval_only

Query the model with your own text

After the call to Seq2SeqModel you can query the model like this (just edit the file):

model.predict(['Singer of Nirvana', 'A jerk', 'The thing I do first in the morning.', 'Duck two girls at the same time', 'Highest mountain in the world'])
# ['Kurt Cobain', 'dickwad', 'piss', 'double dip', 'mount dew']

Judge the results by yourself but I am sure you will find it funny.

philipperemy / urban-dictionary-transformers