The Dissimilarity WordGame

Fun project that checks if you can come up with words that are as dissimilar as possible. (Implementation details at the bottom.)

Play it at https://romanoneg.github.io/WordGame/

Implementation:

Code:

Uses tensorflow.js to load tensorflow's Universal-Sentence-Encoder in order to create vector representations of each word (or multiple words) entered in the text boxes.

The library is imported with the following in index.html:

<script src="https://cdn.jsdelivr.net/npm/@tensorflow/tfjs"></script>
<script src="https://cdn.jsdelivr.net/npm/@tensorflow-models/universal-sentence-encoder"></script>

The model is then loaded in script.js:

use.load().then(loadedModel => {
  model = loadedModel;
};

And inference is done with:

// sentences is a vector of strings to be embedded
model.embed(sentences).then(embeddings => {
  // calculating scores with embeddings
};

---

Score Calculation

These $n$ vectors are then used calculate the final score: $$\text{GameScore} = \frac{1 - \mu_{\text{sim}}}{\text{normalization factor}}$$ where $\mu$ is the average of all the pairwise cosine-similarities between vectors calculated as: $$\mu_{\text{sim}} = {n \choose 2}^{-1} \sum_{i=0}^{n} \sum_{j=i+1}^{n} \text{cosine-similarity} (v_{i}, v_{j})$$ Where $v_{i}$ denotes the i-th word embedding vector and the $\text{normalization factor} = 0.8$ and is hand-chosen to make the scores roughly go from $0$ to $100$.