app.py

import os
import re
from datetime import datetime

import gradio as gr
import numpy as np
import pandas as pd
from gensim.models.fasttext import load_facebook_model
from huggingface_hub import hf_hub_download


ACCESS_KEY = os.environ.get('ACCESS_KEY')
url = hf_hub_download(repo_id="simonschoe/call2vec", filename="model.bin") #hf_hub_url(repo_id="simonschoe/call2vec", filename="model.bin")
model = load_facebook_model(url)

def semantic_search(_input, neighbours):
    """ Perform semantic search """

    _input = re.split('[,;\n]', _input)
    _input = [s.strip().lower().replace(' ', '_') for s in _input if s]

    if _input[0] != ACCESS_KEY:
        with open('log.txt', 'a', encoding='utf-8') as f:
            f.write(str(datetime.now()) + '+++' + '___'.join(_input) + '\n')

    if len(_input) > 1:
        avg_input = np.stack([model.wv[w] for w in _input], axis=0).mean(axis=0)
        nearest_neighbours = model.wv.most_similar(positive=avg_input, topn=neighbours)
        frequencies = [model.wv.get_vecattr(nn[0], 'count') for nn in nearest_neighbours]
    else:
        nearest_neighbours = model.wv.most_similar(positive=_input[0], topn=neighbours)
        frequencies = [model.wv.get_vecattr(nn[0], 'count') for nn in nearest_neighbours]

    if _input[0] == ACCESS_KEY:
        with open('log.txt', 'r', encoding='utf-8') as f:
            prompts = f.readlines()
            prompts = [p.strip().split('+++') for p in prompts]
        result = pd.DataFrame(prompts,
                              columns=['Time', 'Prompt'])
    else:
        result = pd.DataFrame([(a[0],a[1],b) for a,b in zip(nearest_neighbours, frequencies)],
                              columns=['Token', 'Cosine Similarity', 'Corpus Frequency'])

    result.to_csv('result.csv')
    return result, 'result.csv', '\n'.join(_input)

app = gr.Blocks(theme=gr.themes.Default(), css='#component-0 {max-width: 730px; margin: auto; padding-top: 1.5rem}')

with app:
    gr.Markdown(
        """
        # Call2Vec
        ## Semantic Search in Quarterly Earnings Conference Calls
        """
    )

    with gr.Tabs() as tabs:
        with gr.TabItem("🔍 Model", id=0):
            text_in = gr.Textbox(lines=1, placeholder="Insert text", label="Search Query")
            with gr.Row():
                n = gr.Slider(value=50, minimum=5, maximum=500, step=5, label="Number of Neighbours")
                btn = gr.Button("Search")
            df_out = gr.Dataframe(interactive=False)
            f_out = gr.File(interactive=False, label="Download")
            gr.Examples(
                examples = [
                    ["transformation", 20],
                    ["climate_change", 50],
                    ["risk, political_risk, uncertainty", 250],
                ],
                inputs = [text_in, n],
                outputs = [df_out, f_out, text_in],
                fn = semantic_search,
                cache_examples=True
            )
        with gr.TabItem("📝 Usage", id=1):
            gr.Markdown(
                """
                #### App usage
                The model is intended to be used for **semantic search**: It encodes the search query (entered in the textbox on the right) in a dense vector space and finds semantic neighbours, i.e., token which frequently occur within similar contexts in the underlying training data.
                The model allows for two use cases:
                1. *Single Search:* The input query consists of a single word. When provided a bi-, tri-, or even fourgram, the quality of the model output depends on the presence of the query token in the model's vocabulary. N-grams should be concated by an underscore (e.g., "machine_learning" or "artifical_intelligence").
                2. *Multi Search:* The input query may consist of several words or n-grams, seperated by comma, semi-colon or newline. It then computes the average vector over all inputs and performs semantic search based on the average input token.
                """
            )
        with gr.TabItem("📖 About", id=2):
            gr.Markdown(
                """
                #### Project Description
                Call2Vec is a [fastText](https://fasttext.cc/) word embedding model trained via [Gensim](https://radimrehurek.com/gensim/). It maps each token in the vocabulary into a dense, 300-dimensional vector space, designed for performing semantic search.
                The model is trained on a large sample of quarterly earnings conference calls, held by U.S. firms during the 2006-2022 period. In particular, the training data is restriced to the (rather sponentous) executives' remarks of the Q&A section of the call. The data has been preprocessed prior to model training via stop word removal, lemmatization, named entity masking, and coocurrence modeling.
                """
            )

    with gr.Accordion("📙 Citation", open=False):
        citation_button = gr.Textbox(
            value='Placeholder',
            label='Copy to cite these results.',
            show_copy_button=True
        )

    btn.click(semantic_search, inputs=[text_in, n], outputs=[df_out, f_out, text_in])

app.launch()