app.py

import os
from dataclasses import dataclass
from operator import add, sub

import gradio as gr

import numpy as np
from datasets import load_dataset
from sklearn.metrics.pairwise import cosine_similarity

from pyparsing import Word, alphas, Char, ParseException

term = Word(alphas)
operator = Char("+ -")

expression = term + (operator + term)[...]
operations = {"+": add, "-": sub}


def parse_expression(input):
    try:
        return expression.parseString(input)
    except ParseException as e:
        raise gr.Error(f"Parsing error: {e.msg} at position [{e.loc}].")


def evaluate_expression(input):
    # Skip every other item
    words = input[::2]
    operators = input[1::2]

    result = word_to_vectors(words[0])

    for operator, word in zip(operators, words[1:]):
        result = operations[operator](result, word_to_vectors(word))

    return result


dataset = load_dataset("karmiq/glove", split="train")
df = dataset.to_pandas()

all_words = df["word"].to_numpy()
all_vectors = np.array(df["embeddings"].to_list())


def word_to_vectors(word):
    result = df.loc[df["word"] == word].embeddings.to_numpy()
    if len(result) < 1:
        raise gr.Error("Word not found in the dictionary.")
    else:
        return result[0]


def expression_to_vectors(input):
    return evaluate_expression(parse_expression(input))


def get_results(expression):
    if len(expression) < 1:
        raise gr.Error("Please provide an expression.")

    expression = expression.lower()
    vectors = expression_to_vectors(expression)
    similarity_scores = cosine_similarity([vectors], all_vectors)[0]
    top_indices = np.argsort(similarity_scores)[::-1]
    return dict(
        [
            (all_words[i], similarity_scores[i])
            for i in top_indices
            if not all_words[i] in expression.split()
        ][:10]
    )


examples = [
    "king - man + woman",
    "mother - woman + man",
    "berlin - germany + france",
    "saxophone - jazz + classical",
]

initial_output = get_results(examples[0])

css = """
button.gallery-item { color: var(--body-text-color) !important; }
.output-class { color: var(--color-red-700) !important; }
.confidence-set .label .text { font-weight: var(--weight-medium); }
.confidence-set:hover .label { color: var(--color-red-700) !important; }
"""

with gr.Blocks(
    css=css,
    theme=gr.themes.Monochrome(radius_size=gr.themes.sizes.radius_sm),
) as app:
    with gr.Row():
        with gr.Column():
            input = gr.Textbox(value=examples[0], label="Expression")
            with gr.Row():
                btn = gr.Button("Run", variant="primary")
            with gr.Row():
                gr.Markdown(
                    "Demonstration of computing cosine similarity of embeddings "
                    "from the [GloVe](https://nlp.stanford.edu/projects/glove/) dataset."
                )
            with gr.Row():
                gr.Examples(examples, inputs=input)

        with gr.Column():
            output = gr.Label(label="Closest words", value=initial_output)

    btn.click(fn=get_results, inputs=input, outputs=output)

app.launch()