app.py

import streamlit as st
import matplotlib.pyplot as plt
import spacy
import transformers
import os


from spacy.lang.en import English
from transformers import AutoModel, AutoTokenizer
from utils.utils import *

transformers.utils.logging.disable_progress_bar()
os.system("python3 -m spacy download en")

st.markdown("""### TL;DR: give me the keywords!
Здесь вы можете получить отранжированный список ключевых слов по названию и аннотации статьи.

Единственным поддерживаемым языком является английский.""")

st.markdown("<p style=\"text-align:center\"><img width=100% src='https://c.tenor.com/IKt-6tAk9CUAAAAd/thats-a-lot-of-words-lots-of-words.gif'></p>", unsafe_allow_html=True)

#from transformers import pipeline

#pipe = pipeline("ner", "Davlan/distilbert-base-multilingual-cased-ner-hrl")

#st.markdown("#### Title:")
title = st.text_area("Заголовок:", value="How to cook a neural network", height=16, help="Заголовок статьи")
abstract = st.text_area("Аннотация:",
    value="""My dad fits hellish models in general.
Well, this is about an average recipe, because there are a lot of variations.
The model is taken, it is not finetuned, finetuning is not about my dad.
He takes this model, dumps it into the tensorboard and starts frying it.
Adds a huge amount of noize, convolutions, batch and spectral normalization DROPOUT! for regularization, maxpooling on top.
All this is fitted to smoke.
Then the computer is removed from the fire and cools on the balcony.
Then dad brings it in and generously sprinkles it with crossvalidation and starts predicting.
At the same time, he gets data from the web, scraping it with a fork.
Predicts and sentences in a half-whisper oh god.
At the same time, he has sweat on his forehead.
Kindly offers me sometimes, but I refuse.
Do I need to talk about what the wildest overfitting then?
The overfitting is such that the val loss peels off the walls.
    """,
    height=512, help="Аннотация статьи")

# Spacy

@st.cache(hash_funcs={English: lambda _: None})
def get_nlp(nlp_name):
    return spacy.load(nlp_name)

# Вообще, стоит найти pipeline, заточенный под научный текст.
# Но этим займёмся потом, если будет время.
nlp_name = 'en_core_web_sm'
main_nlp = get_nlp(nlp_name)


# Получение модели.

#@st.cache(hash_funcs={transformers.tokenizers.Tokenizer: lambda _: None})
def get_model_and_tokenizer(model_name):
    model = AutoModel.from_pretrained(model_name)
    tokenizer = AutoTokenizer.from_pretrained(model_name)

    return model, tokenizer

model_name = "distilroberta-base"
main_model, main_tokenizer = get_model_and_tokenizer(model_name)


# Обработка текста.

text = preprocess([title + ". " + abstract])[0]

if not text is None and len(text) > 0:
    #keywords = get_candidates(text, main_nlp)
    keywords = get_keywords(text, main_nlp, main_model, main_tokenizer)
    labels = [kw[0].replace(' ', '\n') for kw in keywords]
    scores = [kw[1] for kw in keywords]
    #st.markdown(f"{keywords}")

    # Топ 5 слов.
    top = 5
    top = min(len(labels, top))
    st.markdown("Топ %d ключевых слов: **%s**" % (top, ', '.join(labels[0:5])))

    # График важности слов.
    fig, ax = plt.subplots(figsize=(8, 16))
    ax.set_title("95% самых важных ключевых слов")
    ax.grid(color='#000000', alpha=0.15, linestyle='-', linewidth=1, which='major')
    ax.grid(color='#000000', alpha=0.1, linestyle='-', linewidth=0.5, which='minor')

    bar_width = 0.75
    indexes = -np.arange(len(labels))
    ax.barh(indexes, scores, bar_width)
    plt.yticks(indexes, labels=labels)

    st.pyplot(fig)
else:
    st.markdown("Please, try to enter something.")