from typing import AnyStr, Dict import itertools import streamlit as st import en_core_web_lg import torch.nn.parameter from bs4 import BeautifulSoup import numpy as np import base64 from spacy_streamlit.util import get_svg from streamlit.proto.SessionState_pb2 import SessionState from custom_renderer import render_sentence_custom from sentence_transformers import SentenceTransformer from transformers import AutoTokenizer, AutoModelForTokenClassification from transformers import pipeline import os device = torch.device("cuda" if torch.cuda.is_available() else "cpu") HTML_WRAPPER = """
{}
""" @st.experimental_singleton def get_sentence_embedding_model(): return SentenceTransformer('sentence-transformers/all-MiniLM-L6-v2') @st.experimental_singleton def get_spacy(): nlp = en_core_web_lg.load() return nlp @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) @st.experimental_singleton def get_summarizer_model(): model_name = 'google/pegasus-cnn_dailymail' summarizer_model = pipeline("summarization", model=model_name, tokenizer=model_name, device=0 if torch.cuda.is_available() else -1) return summarizer_model # Page setup st.set_page_config( page_title="📜 Hallucination detection in summaries 📜", 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', '')) # Append free use possibility: filenames.append("Provide your own input") return filenames def fetch_article_contents(filename: str) -> AnyStr: if filename == "Provide your own input": return " " with open(f'./sample-articles/{filename}.txt', 'r') as f: data = f.read() return data def fetch_summary_contents(filename: str) -> AnyStr: with open(f'./sample-summaries/{filename}.txt', 'r') as f: data = f.read() return data def fetch_entity_specific_contents(filename: str) -> AnyStr: with open(f'./entity-specific-text/{filename}.txt', 'r') as f: data = f.read() return data def fetch_dependency_specific_contents(filename: str) -> AnyStr: with open(f'./dependency-specific-text/{filename}.txt', 'r') as f: data = f.read() return data def fetch_ranked_summaries(filename: str, ranknumber: int) -> AnyStr: with open(f'./ranked-summaries/{filename}/Rank{ranknumber}.txt', 'r') as f: data = f.read() return data def fetch_dependency_svg(filename: str) -> AnyStr: with open(f'./dependency-images/{filename}.txt', 'r') as f: lines = [line.rstrip() for line in f] return lines def display_summary(summary_content: str): st.session_state.summary_output = summary_content soup = BeautifulSoup(summary_content, features="html.parser") return HTML_WRAPPER.format(soup) def get_all_entities_per_sentence(text): 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 (CURRENTLY NOT USED) # 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)) def get_and_compare_entities(first_time: bool): if first_time: article_content = st.session_state.article_text all_entities_per_sentence = get_all_entities_per_sentence(article_content) entities_article = list(itertools.chain.from_iterable(all_entities_per_sentence)) st.session_state.entities_article = entities_article else: entities_article = st.session_state.entities_article summary_content = st.session_state.summary_output all_entities_per_sentence = get_all_entities_per_sentence(summary_content) entities_summary = list(itertools.chain.from_iterable(all_entities_per_sentence)) matched_entities = [] unmatched_entities = [] for entity in entities_summary: 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, show_progress_bar=False), sentence_embedding_model.encode(art_entity, show_progress_bar=False)) > 0.9 for art_entity in entities_article): matched_entities.append(entity) else: unmatched_entities.append(entity) matched_entities = list(dict.fromkeys(matched_entities)) unmatched_entities = list(dict.fromkeys(unmatched_entities)) matched_entities_to_remove = [] unmatched_entities_to_remove = [] for entity in matched_entities: for substring_entity in matched_entities: if entity != substring_entity and entity.lower() in substring_entity.lower(): matched_entities_to_remove.append(entity) for entity in unmatched_entities: for substring_entity in unmatched_entities: if entity != substring_entity and entity.lower() in substring_entity.lower(): unmatched_entities_to_remove.append(entity) matched_entities_to_remove = list(dict.fromkeys(matched_entities_to_remove)) unmatched_entities_to_remove = list(dict.fromkeys(unmatched_entities_to_remove)) for entity in matched_entities_to_remove: matched_entities.remove(entity) for entity in unmatched_entities_to_remove: unmatched_entities.remove(entity) return matched_entities, unmatched_entities def highlight_entities(): summary_content = st.session_state.summary_output markdown_start_red = "" markdown_start_green = "" markdown_end = "" matched_entities, unmatched_entities = get_and_compare_entities(True) 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") return HTML_WRAPPER.format(soup) def render_dependency_parsing(text: Dict): html = render_sentence_custom(text, nlp) html = html.replace("\n\n", "\n") st.write(get_svg(html), unsafe_allow_html=True) def check_dependency(article: bool): if article: text = st.session_state.article_text all_entities = get_all_entities_per_sentence(text) else: text = st.session_state.summary_output all_entities = get_all_entities_per_sentence(text) doc = nlp(text) tok_l = doc.to_json()['tokens'] 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: 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]: 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]: 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 return test_list_dict_output 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'' % b64 return html def generate_abstractive_summary(text, type, min_len=120, max_len=512, **kwargs): text = text.strip().replace("\n", " ") if type == "top_p": text = summarization_model(text, min_length=min_len, max_length=max_len, top_k=50, top_p=0.95, clean_up_tokenization_spaces=True, truncation=True, **kwargs) elif type == "greedy": text = summarization_model(text, min_length=min_len, max_length=max_len, clean_up_tokenization_spaces=True, truncation=True, **kwargs) elif type == "top_k": text = summarization_model(text, min_length=min_len, max_length=max_len, top_k=50, clean_up_tokenization_spaces=True, truncation=True, **kwargs) elif type == "beam": text = summarization_model(text, min_length=min_len, max_length=max_len, clean_up_tokenization_spaces=True, truncation=True, **kwargs) summary = text[0]['summary_text'].replace("", " ") return summary # Load all different models (cached) at start time of the hugginface space sentence_embedding_model = get_sentence_embedding_model() ner_model = get_transformer_pipeline() nlp = get_spacy() summarization_model = get_summarizer_model() # Page st.title('📜 Hallucination detection 📜') st.subheader("🔎 Detecting errors in generated abstractive summaries") #st.title('📜 Error detection in summaries 📜') # INTRODUCTION st.header("🧑‍🏫 Introduction") #introduction_checkbox = st.checkbox("Show introduction text", value=True) #if introduction_checkbox: 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 summarization** merely copies informative fragments from the input. - **Abstractive summarization** may generate novel words. A good abstractive summary should cover principal information in the input and has to be linguistically fluent. This interactive blogpost will focus on this more difficult task of abstractive summary generation. Furthermore we will focus mainly on hallucination errors, and less on sentence fluency.""") st.markdown("###") st.markdown("🤔 **Why is this important?** 🤔 Let's say we want to summarize news articles for a popular " "newspaper. If an article tells the story of Elon Musk buying **Twitter**, we don't want our summarization " "model to say that he bought **Facebook** instead. Summarization could also be done for financial reports " "for example. In such environments, these errors can be very critical, so we want to find a way to " "detect them.") st.markdown("###") 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 detect errors from the generated summaries. 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.""") # 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 article, as otherwise the " "summary generated from it might not be optimal, leading to suboptimal performance of the post-processing " "steps. However, too long articles will be truncated and might miss information in the summary.") st.markdown("####") selected_article = st.selectbox('Select an article or provide your own:', list_all_article_names(), index=2) 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=250 ) summarize_button = st.button(label='🤯 Process article content', help="Start interactive blogpost") if summarize_button: st.session_state.article_text = article_text st.markdown("####") st.markdown( "*Below you can find the generated summary for the article. We will discuss two approaches that we found are " "able to detect some common errors. Based on these errors, one could then score different summaries, indicating 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.*") st.markdown("####") if st.session_state.article_text: with st.spinner('Generating summary, this might take a while...'): if selected_article != "Provide your own input" and article_text == fetch_article_contents( selected_article): st.session_state.unchanged_text = True summary_content = fetch_summary_contents(selected_article) else: summary_content = generate_abstractive_summary(article_text, type="beam", do_sample=True, num_beams=15, no_repeat_ngram_size=4) st.session_state.unchanged_text = False summary_displayed = display_summary(summary_content) st.write("✍ **Generated summary:** ✍", summary_displayed, unsafe_allow_html=True) else: st.error('**Error**: No comment to classify. Please provide a comment.') # ENTITY MATCHING PART st.header("1️⃣ 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**. ") st.markdown("Hallucinations are words generated by the model that are not supported by " "the source input. Deep learning based generation is [prone to hallucinate](" "https://arxiv.org/pdf/2202.03629.pdf) unintended text. These hallucinations degrade " "system performance and fail to meet user expectations in many real-world scenarios. By applying entity matching, we can improve this problem" " for the downstream task of summary generation.") st.markdown(" 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, this takes about 10-20 seconds..."): entity_match_html = highlight_entities() st.markdown("####") st.write(entity_match_html, unsafe_allow_html=True) red_text = """red """ green_text = """ green """ markdown_start_red = "" markdown_start_green = "" 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 the current article, we provide a short explanation of the results below **(only for " "example articles)**. ", unsafe_allow_html=True) if st.session_state.unchanged_text: entity_specific_text = fetch_entity_specific_contents(selected_article) soup = BeautifulSoup(entity_specific_text, features="html.parser") st.markdown("####") st.write("💡👇 **Specific example explanation** 👇💡", HTML_WRAPPER.format(soup), unsafe_allow_html=True) # DEPENDENCY PARSING PART st.header("2️⃣ 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, as can be seen in the example below. ") st.write(render_svg('SecondExampleParsing.svg'), unsafe_allow_html=True) st.markdown("*“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." " More on the different sorts of dependencies and their description can be found [here](https://universaldependencies.org/docs/en/dep/).") st.markdown("However, we have found that **there are specific dependencies that are often an " "indication of a wrongly constructed sentence** -when there is no article match. We (currently) use 2 " "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 example article.") with st.spinner("Doing dependency parsing..."): if st.session_state.unchanged_text: for cur_svg_image in fetch_dependency_svg(selected_article): st.write(cur_svg_image, unsafe_allow_html=True) dep_specific_text = fetch_dependency_specific_contents(selected_article) soup = BeautifulSoup(dep_specific_text, features="html.parser") st.write("💡👇 **Specific example explanation** 👇💡", HTML_WRAPPER.format(soup), 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) if total_unmatched_deps: for current_drawing_list in total_unmatched_deps: render_dependency_parsing(current_drawing_list) # CURRENTLY DISABLED # OUTRO/CONCLUSION st.header("🤝 Bringing it together") st.markdown("We have presented 2 methods that try to detect errors in 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 first introduction. As the methods were " "empirically tested they are definitely not sufficiently robust for general use-cases.") st.markdown("####") st.markdown( "Below we generate 3 different kind of summaries, and based on the two discussed methods, their errors are " "detected to estimate a factualness score. Based on this basic approach, " "the best summary (read: the one that a human would prefer or indicate as the best one) " "will hopefully be at the top. Summaries with the same scores will get the same rank displayed. We currently " "only do this for the example articles (for which the different summmaries are already generated). The reason " "for this is that HuggingFace spaces are limited in their CPU memory.") st.markdown("####") if selected_article != "Provide your own input" and article_text == fetch_article_contents(selected_article): with st.spinner("Calculating more summaries and scoring them, this might take a minute or two..."): summaries_list = [] deduction_points = [] # FOR NEW GENERATED SUMMARY for i in range(1 , 4): st.session_state.summary_output = fetch_ranked_summaries(selected_article, i) _, amount_unmatched = get_and_compare_entities(False) 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) summaries_list.append(st.session_state.summary_output) deduction_points.append(len(amount_unmatched) + len(total_unmatched_deps)) # RANKING AND SHOWING THE SUMMARIES deduction_points, summaries_list = (list(t) for t in zip(*sorted(zip(deduction_points, summaries_list)))) cur_rank = 1 rank_downgrade = 0 for i in range(len(deduction_points)): st.write(f'🏆 Rank {cur_rank} summary: 🏆', display_summary(summaries_list[i]), unsafe_allow_html=True) if i < len(deduction_points) - 1: rank_downgrade += 1 if not deduction_points[i + 1] == deduction_points[i]: cur_rank += rank_downgrade rank_downgrade = 0