A Blog About D4T4 & M47H

Using Word2Vec for "Code Names"

12 May ’18

"Code Names" Rules: People are divided into two teams. The board is comprised of 25 words divided into 4 categories: blue team, red team, neutral, and the death word. People are divided evenly into two teams (red and blue). In each round, two people from either team take turns giving 1 word clues. The goal is to get the other members of your team to guess your teams' words and NOT the other words, especially not the death word; if your team guesses the death word, you immediately lose.

It is a really fun game. I also thought it might be an interesting application for Word2Vec. Word2Vec is a two-layer neural network which models the linguistic contexts of words. There are two approaches to training Word2Vec: CBOW (continuous bag of words) and skip-gram. CBOW predicts a word from a window of surrounding words. Skip-gram uses a single word to predict words in the surrounding window. This is a nice summary. Also cool, you don't need to train your own Word2Vec model! Lots of people/organizations provide pre-trained word vectors that you can easily implement, e.g. Google News and Facebook.

I built a small app that uses Word2Vec to generate word hints for "Code Names". I used Python's gensim package to measure word similarities / generate hints using pre-trained word vectors from Stanford NLP's GloVe. The app itself is built using Plotly's Dash, which is analogous to Shiny for R. I packaged the entire thing in a Docker container.

Dash app (

import os

import dash
import dash_core_components as dcc
import dash_html_components as html
from dash.dependencies import Input, Output

import pandas as pd
import numpy as np

from gensim.models import KeyedVectors

import plotly.figure_factory as ff

# initialize app
app = dash.Dash()
server = app.server

# load model
model = 'glove/w2v.{}.txt.gz'.format(os.getenv('GLOVE_MODEL', 'glove.6B.50d'))
word_vectors = KeyedVectors.load_word2vec_format(model, binary=False)

# precompute L2-normalized vectors (saves lots of memory)

# pandas df to html
def generate_table(df, max_rows=10):
    return html.Table(
        # header
        [html.Tr([html.Th(col) for col in df.columns])] +

        # body
            html.Td(df.iloc[i][col]) for col in df.columns
        ]) for i in range(min(len(df), max_rows))]

# generate some clues
def generate_hints(words):
        hints = word_vectors.most_similar(positive=words)
        hints = pd.DataFrame.from_records(hints, columns=['word','similarity'])
        return generate_table(hints)
    except KeyError as e:
        return html.Div(str(e))

# generate dendrogram for word similarity
def generate_dendro(words):
        similarities = np.array([word_vectors.distances(w, words) for w in words])
        figure = ff.create_dendrogram(similarities, labels=words)
        figure['layout'].update({'width': 800, 'height': 500})
        return figure
    except KeyError as e:

# set up app layout
app.layout = html.Div(children=[
    html.H1(children='Code Names Hints'),
        html.Tr([html.Td("All Words:"), html.Td("Words for Hints:")]),
        html.Tr([html.Td(dcc.Textarea(id='words-all', value='god zeus bat ball mountain cold snow', style={'width': 500})),
                 html.Td(dcc.Input(id='words', value='bat ball', type='text'))]),
        html.Tr([html.Td(dcc.Graph(id='dendro')), html.Td(html.Div(id='hints'))])

# set up app callbacks
    Output(component_id='dendro', component_property='figure'),
    [Input(component_id='words-all', component_property='value')]
def update_dendro(input_value):
    words = [w.lower() for w in input_value.strip().split(' ')]
    return generate_dendro(words)

    Output(component_id='hints', component_property='children'),
    [Input(component_id='words', component_property='value')]
def update_hints(input_value):
    words = [w.lower() for w in input_value.strip().split(' ')]
    return generate_hints(words)

# run
if __name__ == '__main__':


FROM python:3.5-slim

ENV GLOVE_MODEL glove.6B.200d


RUN apt-get update \
  && apt-get install -y unzip gzip wget \
  && rm -rf /var/lib/apt/lists/*

COPY requirements.txt ./
RUN chmod +x

RUN pip install -r requirements.txt

# requirements.txt:
# dash==0.21.1                    gensim==3.4.0
# dash-core-components==0.22.1    pandas==0.22.0
# dash-html-components==0.10.1    gunicorn==19.8.1
# dash-renderer==0.12.1           gevent==1.2.2

RUN wget -q \
  && unzip -d glove \
  && rm \
  && python -m gensim.scripts.glove2word2vec --input glove/${GLOVE_MODEL}.txt --output glove/w2v.${GLOVE_MODEL}.txt \
  && gzip glove/w2v.${GLOVE_MODEL}.txt \
  && rm glove/*.txt



echo Starting Gunicorn...
gunicorn app:server \
    --name code-names \
    --bind$PORT \
    --workers $GUNICORN_WORKERS \
    --preload \
    --worker-class gevent \
    --timeout 600 \
    --log-level info \

Example output:

Overall, it does...okay haha. In some cases, it does surprisingly well. For instance, the app provides "published" as a top hint for "book" and "penguin". However, the algorithm struggles to identify commonalities that may not be explicitly collocated in text. For instance, for "dog" and "whale", "mammal" might be a good hint. However, our app simply lists other animals, e.g. "cat" and "shark".

I'm hosting a version of the app here on And a link to my repo on GH. Cheers!