app.py

import random
from typing import AnyStr
# import tensorflow_hub as hub

import itertools
import streamlit as st

import torch.nn.parameter
from bs4 import BeautifulSoup
import numpy as np
import base64

import validators
from spacy_streamlit.util import get_svg
from validators import ValidationFailure

from custom_renderer import render_sentence_custom
# from flair.data import Sentence
# from flair.models import SequenceTagger
from sentence_transformers import SentenceTransformer

import spacy
from spacy import displacy
from spacy_streamlit import visualize_parser

from transformers import AutoTokenizer, AutoModelForSequenceClassification, AutoModelForTokenClassification
from transformers import pipeline
import os
from transformers_interpret import SequenceClassificationExplainer

# USE_model = hub.load("https://tfhub.dev/google/universal-sentence-encoder/4")
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")


@st.experimental_singleton
def get_sentence_embedding_model():
    return SentenceTransformer('sentence-transformers/all-MiniLM-L6-v2')


@st.experimental_singleton
def get_spacy():
    nlp = spacy.load('en_core_web_lg')
    return nlp


# TODO: might look into which one is the best here
# TODO: might be useful to make an ml6 preloaded model for flair as this takes ridiculously long to load the first time
# @st.experimental_singleton
# @st.cache(suppress_st_warning=True, allow_output_mutation=True)
# def get_flair_tagger():
#     return SequenceTagger.load("flair/ner-english-ontonotes-fast")


@st.experimental_singleton
def get_transformer_pipeline():
    tokenizer = AutoTokenizer.from_pretrained("xlm-roberta-large-finetuned-conll03-english")
    model = AutoModelForTokenClassification.from_pretrained("xlm-roberta-large-finetuned-conll03-english")
    return pipeline("ner", model=model, tokenizer=tokenizer, grouped_entities=True)


# Page setup
st.set_page_config(
    page_title="Post-processing summarization fact checker",
    page_icon="",
    layout="centered",
    initial_sidebar_state="auto",
    menu_items={
        'Get help': None,
        'Report a bug': None,
        'About': None,
    }
)


def list_all_article_names() -> list:
    filenames = []
    for file in sorted(os.listdir('./sample-articles/')):
        if file.endswith('.txt'):
            filenames.append(file.replace('.txt', ''))
    return filenames


def fetch_article_contents(filename: str) -> AnyStr:
    with open(f'./sample-articles/{filename.lower()}.txt', 'r') as f:
        data = f.read()
    return data


def fetch_summary_contents(filename: str) -> AnyStr:
    with open(f'./sample-summaries/{filename.lower()}.txt', 'r') as f:
        data = f.read()
    return data


def fetch_entity_specific_contents(filename: str) -> AnyStr:
    with open(f'./entity-specific-text/{filename.lower()}.txt', 'r') as f:
        data = f.read()
    return data


def fetch_dependency_specific_contents(filename: str) -> AnyStr:
    with open(f'./dependency-specific-text/{filename.lower()}.txt', 'r') as f:
        data = f.read()
    return data


def fetch_dependency_svg(filename: str) -> AnyStr:
    with open(f'./dependency-images/{filename.lower()}.txt', 'r') as f:
        #data = f.read()
        lines=[line.rstrip() for line in f]
    return lines


def display_summary(article_name: str):
    summary_content = fetch_summary_contents(article_name)
    st.session_state.summary_output = summary_content
    soup = BeautifulSoup(summary_content, features="html.parser")
    HTML_WRAPPER = """<div style="overflow-x: auto; border: 1px solid #e6e9ef; border-radius: 0.25rem; padding: 1rem; margin-bottom: 2.5rem">{}</div>"""
    return HTML_WRAPPER.format(soup)


def get_all_entities_per_sentence(text):
    # load all NER models
    # nlp = get_spacy()
    # tagger = get_flair_tagger()
    doc = nlp(text)

    sentences = list(doc.sents)

    entities_all_sentences = []
    for sentence in sentences:
        entities_this_sentence = []

        # SPACY ENTITIES
        for entity in sentence.ents:
            entities_this_sentence.append(str(entity))

        # FLAIR ENTITIES
        # sentence_entities = Sentence(str(sentence))
        # tagger.predict(sentence_entities)
        # for entity in sentence_entities.get_spans('ner'):
        #     entities_this_sentence.append(entity.text)

        # XLM ENTITIES
        entities_xlm = [entity["word"] for entity in ner_model(str(sentence))]
        for entity in entities_xlm:
            entities_this_sentence.append(str(entity))

        entities_all_sentences.append(entities_this_sentence)

    return entities_all_sentences


def get_all_entities(text):
    all_entities_per_sentence = get_all_entities_per_sentence(text)
    return list(itertools.chain.from_iterable(all_entities_per_sentence))


# TODO: this functionality can be cached (e.g. by storing html file output) if wanted (or just store list of entities idk)
def get_and_compare_entities(article_name: str):
    article_content = fetch_article_contents(article_name)
    all_entities_per_sentence = get_all_entities_per_sentence(article_content)
    # st.session_state.entities_per_sentence_article = all_entities_per_sentence
    entities_article = list(itertools.chain.from_iterable(all_entities_per_sentence))

    summary_content = fetch_summary_contents(article_name)
    all_entities_per_sentence = get_all_entities_per_sentence(summary_content)
    # st.session_state.entities_per_sentence_summary = all_entities_per_sentence
    entities_summary = list(itertools.chain.from_iterable(all_entities_per_sentence))

    matched_entities = []
    unmatched_entities = []
    for entity in entities_summary:
        # TODO: currently substring matching but probably should do embedding method or idk?
        if any(entity.lower() in substring_entity.lower() for substring_entity in entities_article):
            matched_entities.append(entity)
        elif any(
                np.inner(sentence_embedding_model.encode(entity), sentence_embedding_model.encode(art_entity)) > 0.9 for
                art_entity in entities_article):
            matched_entities.append(entity)
        else:
            unmatched_entities.append(entity)
    return matched_entities, unmatched_entities


def highlight_entities(article_name: str):
    summary_content = fetch_summary_contents(article_name)

    markdown_start_red = "<mark class=\"entity\" style=\"background: rgb(238, 135, 135);\">"
    markdown_start_green = "<mark class=\"entity\" style=\"background: rgb(121, 236, 121);\">"
    markdown_end = "</mark>"

    matched_entities, unmatched_entities = get_and_compare_entities(article_name)

    for entity in matched_entities:
        summary_content = summary_content.replace(entity, markdown_start_green + entity + markdown_end)

    for entity in unmatched_entities:
        summary_content = summary_content.replace(entity, markdown_start_red + entity + markdown_end)
    soup = BeautifulSoup(summary_content, features="html.parser")

    HTML_WRAPPER = """<div style="overflow-x: auto; border: 1px solid #e6e9ef; border-radius: 0.25rem; padding: 1rem; 
    margin-bottom: 2.5rem">{}</div> """

    return HTML_WRAPPER.format(soup)


def render_dependency_parsing(text: str):
    html = render_sentence_custom(text)
    html = html.replace("\n\n", "\n")
    print(get_svg(html))
    st.write(get_svg(html), unsafe_allow_html=True)


# If deps for article: True, otherwise deps for summary calc
def check_dependency(article: bool):
    # nlp = spacy.load('en_core_web_lg')
    if article:
        text = st.session_state.article_text
        all_entities = get_all_entities_per_sentence(text)
        # all_entities = st.session_state.entities_per_sentence_article
    else:
        text = st.session_state.summary_output
        all_entities = get_all_entities_per_sentence(text)
        # all_entities = st.session_state.entities_per_sentence_summary
    doc = nlp(text)
    tok_l = doc.to_json()['tokens']
    # all_deps = ""
    test_list_dict_output = []

    sentences = list(doc.sents)
    for i, sentence in enumerate(sentences):
        start_id = sentence.start
        end_id = sentence.end
        for t in tok_l:
            # print(t)
            if t["id"] < start_id or t["id"] > end_id:
                continue
            head = tok_l[t['head']]
            if t['dep'] == 'amod' or t['dep'] == "pobj":
                object_here = text[t['start']:t['end']]
                object_target = text[head['start']:head['end']]
                if t['dep'] == "pobj" and str.lower(object_target) != "in":
                    continue
                # ONE NEEDS TO BE ENTITY
                if object_here in all_entities[i]:
                    # all_deps = all_deps.join(str(sentence))
                    identifier = object_here + t['dep'] + object_target
                    test_list_dict_output.append({"dep": t['dep'], "cur_word_index": (t['id'] - sentence.start),
                                                  "target_word_index": (t['head'] - sentence.start),
                                                  "identifier": identifier, "sentence": str(sentence)})
                elif object_target in all_entities[i]:
                    # all_deps = all_deps.join(str(sentence))
                    identifier = object_here + t['dep'] + object_target
                    test_list_dict_output.append({"dep": t['dep'], "cur_word_index": (t['id'] - sentence.start),
                                                  "target_word_index": (t['head'] - sentence.start),
                                                  "identifier": identifier, "sentence": str(sentence)})
                else:
                    continue
    # print(f'NOW TEST LIST DICT: {test_list_dict_output}')
    return test_list_dict_output
    # return all_deps


def is_valid_url(url: str) -> bool:
    result = validators.url(url)
    if isinstance(result, ValidationFailure):
        return False
    return True


# Start session
if 'results' not in st.session_state:
    st.session_state.results = []

# Page
st.title('Summarization fact checker')

# INTRODUCTION
st.header("Introduction")
st.markdown("""Recent work using transformers on large text corpora has shown great success when fine-tuned on 
several different downstream NLP tasks. One such task is that of text summarization. The goal of text summarization 
is to generate concise and accurate summaries from input document(s). There are 2 types of summarization: extractive 
and abstractive. **Extractive summarization** merely copies informative fragments from the input, 
whereas **abstractive summarization** may generate novel words. A good abstractive summary should cover principal 
information in the input and has to be linguistically fluent. This blogpost will focus on this more difficult task of 
abstractive summary generation.""")

st.markdown("""To generate summaries we will use the [PEGASUS] (https://huggingface.co/google/pegasus-cnn_dailymail) 
model, producing abstractive summaries from large articles. These summaries often contain sentences with different 
kinds of errors. Rather than improving the core model, we will look into possible post-processing steps to improve 
the generated summaries. By comparing contents of the summary with the source text, we come up with a factualness 
metric, indicating the trustworthiness of the generated summary. Throughout this blog, we will also explain the 
results for some methods on specific examples. These text blocks will be indicated and they change according to the 
currently selected article.""")


sentence_embedding_model = get_sentence_embedding_model()
# tagger = get_flair_tagger()
ner_model = get_transformer_pipeline()
nlp = get_spacy()

# GENERATING SUMMARIES PART
st.header("Generating summaries")
st.markdown("Let’s start by selecting an article text for which we want to generate a summary, or you can provide "
            "text yourself. Note that it’s suggested to provide a sufficiently large text, as otherwise the summary "
            "generated from it might not be optimal, leading to suboptimal performance of the post-processing steps.")

# TODO: NEED TO CHECK ARTICLE TEXT INSTEAD OF ARTICLE NAME ALSO FREE INPUT OPTION
selected_article = st.selectbox('Select an article or provide your own:',
                                list_all_article_names())  # index=0, format_func=special_internal_function, key=None, help=None, on_change=None, args=None, kwargs=None, *, disabled=False)
st.session_state.article_text = fetch_article_contents(selected_article)
article_text = st.text_area(
    label='Full article text',
    value=st.session_state.article_text,
    height=150
)

st.markdown("Below you can find the generated summary for the article. Based on empirical research, we will discuss "
            "two main methods that detect some common errors. We can then score different summaries, to indicate how "
            "factual a summary is for a given article. The idea is that in production, you could generate a set of "
            "summaries for the same article, with different parameters (or even different models). By using "
            "post-processing error detection, we can then select the best possible summary.")
if st.session_state.article_text:
    with st.spinner('Generating summary...'):
        # classify_comment(article_text, selected_model)

        summary_displayed = display_summary(selected_article)

        st.write("**Generated summary:**", summary_displayed, unsafe_allow_html=True)
else:
    st.error('**Error**: No comment to classify. Please provide a comment.',
             help="Generate summary for the given article text")


def render_svg(svg_file):
    with open(svg_file, "r") as f:
        lines = f.readlines()
        svg = "".join(lines)

        # """Renders the given svg string."""
        b64 = base64.b64encode(svg.encode("utf-8")).decode("utf-8")
        html = r'<img src="data:image/svg+xml;base64,%s"/>' % b64
        return html


# ENTITY MATCHING PART
st.header("Entity matching")
st.markdown("The first method we will discuss is called **Named Entity Recognition** (NER). NER is the task of "
            "identifying and categorising key information (entities) in text. An entity can be a singular word or a "
            "series of words that consistently refers to the same thing. Common entity classes are person names, "
            "organisations, locations and so on. By applying NER to both the article and its summary, we can spot "
            "possible **hallucinations**. Hallucinations are words generated by the model that are not supported by "
            "the source input. In theory all entities in the summary (such as dates, locations and so on), "
            "should also be present in the article. Thus we can extract all entities from the summary and compare "
            "them to the entities of the original article, spotting potential hallucinations. The more unmatched "
            "entities we find, the lower the factualness score of the summary. ")
with st.spinner("Calculating and matching entities..."):
    entity_match_html = highlight_entities(selected_article)
    st.write(entity_match_html, unsafe_allow_html=True)
    red_text = """<font color="black"><span style="background-color: rgb(238, 135, 135); opacity: 
    1;">red</span></font> """
    green_text = """<font color="black">
        <span style="background-color: rgb(121, 236, 121); opacity: 1;">green</span>
    </font>"""

    markdown_start_red = "<mark class=\"entity\" style=\"background: rgb(238, 135, 135);\">"
    markdown_start_green = "<mark class=\"entity\" style=\"background: rgb(121, 236, 121);\">"
    st.markdown("We call this technique “entity matching” and here you can see what this looks like when we apply "
                "this method on the summary. Entities in the summary are marked  " + green_text + " when the entity "
                                                                                                  "also exists in the "
                                                                                                  "article, "
                                                                                                  "while unmatched "
                                                                                                  "entities are "
                                                                                                  "marked " +
                red_text + ". Several of the example articles and their summaries indicate different errors we find "
                           "by using this technique. Based on which article you choose, we provide a short "
                           "explanation of the results below.",
                unsafe_allow_html=True)
    entity_specific_text = fetch_entity_specific_contents(selected_article)
    soup = BeautifulSoup(entity_specific_text, features="html.parser")
    HTML_WRAPPER = """<div style="overflow-x: auto; border: 1px solid #e6e9ef; border-radius: 0.25rem; padding: 1rem;
    margin-bottom: 2.5rem">{}</div> """
    st.write("💡👇 **Specific example explanation** 👇💡", HTML_WRAPPER.format(soup), unsafe_allow_html=True)

# DEPENDENCY PARSING PART
st.header("Dependency comparison")
st.markdown("The second method we use for post-processing is called **Dependency parsing**: the process in which the "
            "grammatical structure in a sentence is analysed, to find out related words as well as the type of the "
            "relationship between them. For the sentence “Jan’s wife is called Sarah” you would get the following "
            "dependency graph:")

# TODO: I wonder why the first doesn't work but the second does (it doesn't show deps otherwise)
# st.image("ExampleParsing.svg")
st.write(render_svg('ExampleParsing.svg'), unsafe_allow_html=True)
st.markdown("Here, “Jan” is the “poss” (possession modifier) of “wife”. If suddenly the summary would read “Jan’s "
            "husband…”, there would be a dependency in the summary that is non-existent in the article itself (namely "
            "“Jan” is the “poss” of “husband”). However, often new dependencies are introduced in the summary that "
            "are still correct. “The borders of Ukraine” have a different dependency between “borders” and “Ukraine” "
            "than “Ukraine’s borders”, while both descriptions have the same meaning. So just matching all "
            "dependencies between article and summary (as we did with entity matching) would not be a robust method.")
st.markdown("However, by empirical testing, we have found that there are certain dependencies which can be used for "
            "such matching techniques. When unmatched, these specific dependencies are often an indication of a "
            "wrongly constructed sentence. **Should I explain this more/better or is it enough that I explain by "
            "example specific run throughs?**. We found 2(/3 TODO) common dependencies which, when present in the "
            "summary but not in the article, are highly indicative of factualness errors. Furthermore, we only check "
            "dependencies between an existing **entity** and its direct connections. Below we highlight all unmatched "
            "dependencies that satisfy the discussed constraints. We also discuss the specific results for the "
            "currently selected article.")
with st.spinner("Doing dependency parsing..."):
    # TODO RIGHT IF FUNCTION (IF EXAMPLE AND IF INPUT UNCHANGED)
    #if selected_article == 'article11':
    if True:
        for cur_svg_image in fetch_dependency_svg(selected_article):
            st.write(cur_svg_image, unsafe_allow_html=True)
    else:
        summary_deps = check_dependency(False)
        article_deps = check_dependency(True)
        total_unmatched_deps = []
        for summ_dep in summary_deps:
            if not any(summ_dep['identifier'] in art_dep['identifier'] for art_dep in article_deps):
                total_unmatched_deps.append(summ_dep)
        # print(f'ALL UNMATCHED DEPS ARE: {total_unmatched_deps}')
        # render_dependency_parsing(check_dependency(False))
        if total_unmatched_deps:
            for current_drawing_list in total_unmatched_deps:
                render_dependency_parsing(current_drawing_list)
    dep_specific_text = fetch_dependency_specific_contents(selected_article)
    soup = BeautifulSoup(dep_specific_text, features="html.parser")
    HTML_WRAPPER = """<div style="overflow-x: auto; border: 1px solid #e6e9ef; border-radius: 0.25rem; padding: 1rem;
    margin-bottom: 2.5rem">{}</div> """
    st.write("💡👇 **Specific example explanation** 👇💡", HTML_WRAPPER.format(soup), unsafe_allow_html=True)

# OUTRO/CONCLUSION
st.header("Wrapping up")
st.markdown("We have presented 2 methods that try to improve summaries via post-processing steps. Entity matching can "
            "be used to solve hallucinations, while dependency comparison can be used to filter out some bad "
            "sentences (and thus worse summaries). These methods highlight the possibilities of post-processing "
            "AI-made summaries, but are only a basic introduction. As the methods were empirically tested they are "
            "definitely not sufficiently robust for general use-cases. (something about that we tested also RE and "
            "maybe other things).")
st.markdown("####")
st.markdown("Below we generated 5 different kind of summaries from the article in which their ranks are estimated, "
            "and hopefully the best summary (read: the one that a human would prefer or indicate as the best one) "
            "will be at the top. TODO: implement this (at the end I think) and also put something in the text with "
            "the actual parameters or something? ")