app.py

from transformers import AutoModel, AutoTokenizer
import faiss
import numpy as np
import pandas as pd
import streamlit as st
import torch

import math
import os
import re

os.environ['KMP_DUPLICATE_LIB_OK']='True'


@st.cache(allow_output_mutation=True)
def load_model_and_tokenizer():
    tokenizer = AutoTokenizer.from_pretrained("kaisugi/scitoricsbert")
    model = AutoModel.from_pretrained("kaisugi/scitoricsbert", output_attentions=True)
    model.eval()
    
    return model, tokenizer


@st.cache(allow_output_mutation=True)
def load_sentence_data():
    sentence_df = pd.read_csv("sentence_data_858k.csv.gz")
    
    return sentence_df


@st.cache(allow_output_mutation=True)
def load_sentence_embeddings_and_index():
    npz_comp = np.load("sentence_embeddings_858k.npz")
    sentence_embeddings = npz_comp["arr_0"]

    faiss.normalize_L2(sentence_embeddings)
    D = 768
    N = 857610
    Xt = sentence_embeddings[:100000]
    X = sentence_embeddings

    # Param of PQ
    M = 16  # The number of sub-vector. Typically this is 8, 16, 32, etc.
    nbits = 8 # bits per sub-vector. This is typically 8, so that each sub-vec is encoded by 1 byte
    # Param of IVF
    nlist = int(math.sqrt(N))  # The number of cells (space partition). Typical value is sqrt(N)
    # Param of HNSW
    hnsw_m = 32  # The number of neighbors for HNSW. This is typically 32

    # Setup
    quantizer = faiss.IndexHNSWFlat(D, hnsw_m)
    index = faiss.IndexIVFPQ(quantizer, D, nlist, M, nbits)

    # Train
    index.train(Xt)

    # Add
    index.add(X)

    # Search
    index.nprobe = 8  # Runtime param. The number of cells that are visited for search.

    return sentence_embeddings, index


@st.cache(allow_output_mutation=True)
def formulaic_phrase_extraction(sentences, model, tokenizer):
    THRESHOLD = 0.01
    LAYER = 10

    output_sentences = []

    with torch.no_grad():
        inputs = tokenizer.batch_encode_plus(
            sentences, 
            padding=True,
            truncation=True,
            max_length=512,
            return_tensors='pt'
        )
        outputs = model(**inputs)
        attention = outputs[-1]

        cls_attentions = torch.mean(attention[LAYER][0], dim=0)

        for sentence, cls_attention in zip(sentences, cls_attentions):
            check_bool_arr = list((cls_attention > THRESHOLD).numpy())[1:-1]
            tokens = tokenizer.tokenize(sentence)

            cur_tokens = tokens.copy()

            while True:
                flg = False

                for idx, token in enumerate(cur_tokens):
                    if token.startswith("##"):
                        flg = True
                        back_token = token.replace("##", "")
                        front_token = cur_tokens.pop(idx-1)
                        cur_tokens[idx-1] = front_token + back_token

                        back_bool_val = check_bool_arr[idx]
                        front_bool_val = check_bool_arr.pop(idx-1)
                        check_bool_arr[idx-1] = front_bool_val and back_bool_val

                if not flg:
                    break

            result = " ".join([f'<font color="coral">{original_word}</font>' if b else original_word for (b, original_word) in zip(check_bool_arr, sentence.split())])
            output_sentences.append(result)

    return output_sentences


@st.cache(allow_output_mutation=True)
def get_retrieval_results(index, input_text, top_k, model, tokenizer, sentence_df, exclude_word_list, phrase_annotated=True):
    with torch.no_grad():
        inputs = tokenizer.encode_plus(
            input_text,
            padding=True,
            truncation=True,
            max_length=512,
            return_tensors='pt'
        )
        outputs = model(**inputs)
        query_embeddings = outputs.last_hidden_state[:, 0, :][0]
        query_embeddings = query_embeddings.detach().cpu().numpy()
        query_embeddings = query_embeddings / np.linalg.norm(query_embeddings, ord=2)

    _, ids = index.search(x=np.array([query_embeddings]), k=top_k)
    retrieved_sentences = []
    retrieved_paper_ids = []

    for id in ids[0]:
        cur_sentence = sentence_df.loc[id, "sentence"]
        cur_link = f"https://aclanthology.org/{sentence_df.loc[id, 'file_id']}"

        if len(exclude_word_list) == 0:
            retrieved_sentences.append(cur_sentence)
            retrieved_paper_ids.append(cur_link)

        else:
            exclude_word_list_regex = '|'.join(exclude_word_list)
            pat = re.compile(f'{exclude_word_list_regex}')
            
            if not bool(pat.search(cur_sentence)):
                retrieved_sentences.append(cur_sentence)
                retrieved_paper_ids.append(cur_link)

    if phrase_annotated:
        retrieved_sentences = formulaic_phrase_extraction(retrieved_sentences, model, tokenizer)

    return retrieved_sentences, retrieved_paper_ids


if __name__ == "__main__":
    model, tokenizer = load_model_and_tokenizer()
    sentence_df = load_sentence_data()
    sentence_embeddings, index = load_sentence_embeddings_and_index()


    st.markdown("## AI-based Paraphrasing for Academic Writing")

    input_text = st.text_area("text input", "Our model shows good results.", placeholder="Write something here...")
    top_k = st.number_input('top_k (upperbound)', min_value=1, value=30, step=1)
    input_words = st.text_input("exclude words (comma separated)", "good, result")

    agree = st.checkbox('Include phrase annotation')

    if st.button('search'):
        exclude_word_list = [s.strip() for s in input_words.split(",") if s.strip() != ""]
        retrieved_sentences, retrieved_paper_ids = get_retrieval_results(index, input_text, top_k, model, tokenizer, sentence_df, exclude_word_list, phrase_annotated=agree)

        result_table_markdown = "|  sentence  |  source link  |\n|:---|:---|\n"

        for (retrieved_sentence, retrieved_paper_id) in zip(retrieved_sentences, retrieved_paper_ids):
            result_table_markdown += f"| {retrieved_sentence} | {retrieved_paper_id} |\n"
        
        st.markdown(result_table_markdown, unsafe_allow_html=True)

    st.markdown("---\n#### How this works")

    st.markdown("This app uses ScitoricsBERT [(Sugimoto and Aizawa, 2022)](https://aclanthology.org/2022.sdp-1.7/), a functional sentence representation model, to retrieve sentences that are functionally similar to the input. It also extracts phrasal patterns that accord to the function, by leveraging self-attention patterns within ScitoricsBERT.")