# Run with: streamlit run visualization.py import streamlit as st import os from io import StringIO import base64 import json import pandas as pd pd.options.mode.chained_assignment = None import numpy as np import matplotlib.pyplot as plt from filtering import LoadParameters, ModifyingDocuments, Filtering class Visualization: def __init__( self, path_instructions, path_data, lang, num_docs, num_docs_for_words, max_len_text_display, lang_dataset_id, path_fasttext_model, path_sentencepiece_model, path_kenlm_model, ): self.path_instructions = path_instructions self.path_data = path_data self.lang = lang self.num_docs = num_docs self.num_docs_for_words = num_docs_for_words self.max_len_text_display = max_len_text_display self.lang_dataset_id = lang_dataset_id self.param = LoadParameters.load_parameters(lang_dataset_id) self.stopwords = LoadParameters.load_stopwords(lang_dataset_id) self.flagged_words = LoadParameters.load_flagged_words(lang_dataset_id) self.model_lang_id = LoadParameters.load_model_lang_id( lang_dataset_id, path_fasttext_model ) self.sentencepiece_model = LoadParameters.load_sentencepiece_model( lang_dataset_id, path_sentencepiece_model ) self.sentencepiece_model_tok = ( self.sentencepiece_model if self.param["tokenization"] else None ) self.kenlm_model = LoadParameters.load_kenlm_model( lang_dataset_id, path_kenlm_model ) def warning_preamble(self): st.markdown( "This demo can be a little slow, and only allows you to process up to 5000 documents " "for a decent speed. If you want to display up to three times more documents and have " "a faster visualization, we invite you to run this " "[code](https://github.com/bigscience-workshop/data_tooling/tree/master/ac_dc/visualization) " "on your computer." ) def preamble(self): def get_binary_file_downloader_html(bin_file, file_label="File"): with open(bin_file, "rb") as f: data = f.read() bin_str = base64.b64encode(data).decode() href = f'{file_label}' return href st.markdown( "Before diving into this demo, you might want to take a look at how the filtering pipeline looks like in more detail in this " + get_binary_file_downloader_html( self.path_instructions, "pdf", ) + ".", unsafe_allow_html=True, ) def open_data(self): with open(self.path_data) as json_file: data = json.load(json_file) self.num_docs = min(self.num_docs, len(data)) self.num_docs_for_words = min(self.num_docs_for_words, len(data)) if "words" in data[0]: words = [doc["words"] for doc in data[: self.num_docs_for_words]] words = [word for doc in words for word in doc] self.words = pd.DataFrame(words) else: self.words = None docs = data[: self.num_docs] for doc in docs: if not (self.words is None): del doc["words"] if len(doc["text"]) > self.max_len_text_display: doc["text"] = ( doc["text"][: self.max_len_text_display] + " [...] [THIS LONG TEXT HAS BEEN TRUNCATED FOR DISPLAY REASONS]" ) self.docs_checkpoint = pd.DataFrame(docs) self.docs = self.docs_checkpoint def set_title(self): st.title(f"Filtering visualization") @staticmethod def plot_hist(dataframe, key, num_bins=50): checkbox = st.checkbox( "Diplay distribution", value=True, key=f"display_distribution_{key[0]}" ) if checkbox: fig, ax = plt.subplots() val = dataframe[key[0]].values if np.median(val) != 0: val = val[ abs(val - np.median(val)) < 9 * np.median(np.absolute(val - np.median(val))) ] ax.hist(val, bins=num_bins, density=True) ax.set_title(" ".join(key[0].split("_"))) ax.axvline(x=key[1], color="r", linestyle="dashed") st.pyplot(fig) def filtering_of_docs(self): st.sidebar.subheader("Parameters of the filtering on documents") def set_sliders(): columns = list(self.docs) keys = [] conds = {} def get_cond(key, cutoff, max_cutoff): if max_cutoff: return self.docs[key] <= cutoff return self.docs[key] >= cutoff def print_discared_by_cond(cond): st.caption( f"{(len(cond) - np.sum(1*cond)) / len(cond) * 100:.2f}% of the total is discarded with this filter." ) if "number_words" in columns: with st.sidebar.expander("Number of words"): cutoff_def = "If the number of words of a document is lower than this number, the document is removed." max_nb_words = int(np.max(self.docs["number_words"])) + 1 cutoff_min_number_words = st.slider( cutoff_def, 0, min(max_nb_words, 500), 0 ) new_key = ("number_words", cutoff_min_number_words, False) keys.append(new_key) Visualization.plot_hist(self.docs, new_key) cond_1 = get_cond(new_key[0], new_key[1], new_key[2]) print_discared_by_cond(cond_1) cutoff_def = "If the number of words of a document is higher than this number, the document is removed." cutoff_max_number_words = st.slider( cutoff_def, 0, max_nb_words, max_nb_words ) new_key = ("number_words", cutoff_max_number_words, True) keys.append(new_key) cond_2 = get_cond(new_key[0], new_key[1], new_key[2]) print_discared_by_cond(cond_2) conds["number_words"] = [cond_1, cond_2] if "repetitions_ratio" in columns: with st.sidebar.expander("Repetitions ratio"): val_repetitions_lengths = list( self.docs["repetitions_ratio"].iloc[0].keys() ) default_index = ( val_repetitions_lengths.index("10") if "10" in val_repetitions_lengths else 0 ) label_selectbox = "Length of the repetitions (that will determine the repetitions ratio)." repetitions_length = st.selectbox( label=label_selectbox, options=val_repetitions_lengths, index=default_index, ) st.caption( "Choosing a higher or lower number does not mean that the filtering " "is stronger or weaker. Be careful, choosing a low number (below 5 for languages like English) " "tends to associate a high repetitions ratio to very long documents (like book chapters), but with " "few or no repetitions, simply because their length gives them more diversity, and we do " "not want to discard such documents." ) self.docs["repetitions_ratio"] = self.docs_checkpoint[ "repetitions_ratio" ] for i in range(len(self.docs["repetitions_ratio"])): self.docs["repetitions_ratio"].iloc[i] = self.docs[ "repetitions_ratio" ].iloc[i][repetitions_length] cutoff_def = "If the repetitions ratio of a document is higher than this number, the document is removed." cutoff_repetitions_ratio = st.slider( cutoff_def, 0.0, 1.0, 1.0, step=0.01 ) new_key = ( "repetitions_ratio", cutoff_repetitions_ratio, True, repetitions_length, ) keys.append(new_key) Visualization.plot_hist(self.docs, new_key) cond = get_cond(new_key[0], new_key[1], new_key[2]) print_discared_by_cond(cond) conds["repetitions_ratio"] = [cond] if "special_characters_ratio" in columns: with st.sidebar.expander("Special characters ratio"): cutoff_def = "If the special characters ratio of a document is higher than this number, the document is removed." cutoff_special_characters_ratio = st.slider( cutoff_def, 0.0, 1.0, 1.0, step=0.01 ) new_key = ( "special_characters_ratio", cutoff_special_characters_ratio, True, ) keys.append(new_key) Visualization.plot_hist(self.docs, new_key) cond = get_cond(new_key[0], new_key[1], new_key[2]) print_discared_by_cond(cond) conds["special_characters_ratio"] = [cond] if "stopwords_ratio" in columns: with st.sidebar.expander("Stop words ratio"): stopwords_file = st.file_uploader( "Upload your own list of stop words (one per line). If there is none, the default one is used." ) if stopwords_file: new_stopwords = StringIO( stopwords_file.getvalue().decode("utf-8") ).read() new_stopwords = set(new_stopwords.split("\n")) self.docs["stopwords_ratio"] = self.docs_checkpoint[ "stopwords_ratio" ] for i in range(len(self.docs["stopwords_ratio"])): self.docs["stopwords_ratio"].iloc[ i ] = Filtering.compute_stopwords_ratio( self.docs["text"].iloc[i], self.sentencepiece_model_tok, self.param["strip_characters"], self.param["cond_words_augmentation"], self.param["words_augmentation_group_sizes"], self.param["words_augmentation_join_char"], new_stopwords, ) cutoff_def = "If the stop words ratio of a document is lower than this number, the document is removed." cutoff_stopwords_ratio = st.slider( cutoff_def, 0.0, 1.0, 0.0, step=0.01 ) new_key = ("stopwords_ratio", cutoff_stopwords_ratio, False) keys.append(new_key) Visualization.plot_hist(self.docs, new_key) cond = get_cond(new_key[0], new_key[1], new_key[2]) print_discared_by_cond(cond) conds["stopwords_ratio"] = [cond] if "flagged_words_ratio" in columns: with st.sidebar.expander("Flagged words ratio"): flagged_words_file = st.file_uploader( "Upload your own list of flagged words (one per line). If there is none, the default one is used." ) if flagged_words_file: new_flagged_words = StringIO( flagged_words_file.getvalue().decode("utf-8") ).read() new_flagged_words = set(new_flagged_words.split("\n")) self.docs["flagged_words_ratio"] = self.docs_checkpoint[ "flagged_words_ratio" ] for i in range(len(self.docs["flagged_words_ratio"])): self.docs["flagged_words_ratio"].iloc[ i ] = Filtering.compute_flagged_words_ratio( self.docs["text"].iloc[i], self.sentencepiece_model_tok, self.param["strip_characters"], self.param["cond_words_augmentation"], self.param["words_augmentation_group_sizes"], self.param["words_augmentation_join_char"], new_flagged_words, ) cutoff_def = "If the flagged words ratio of a document is higher than this number, the document is removed." cutoff_flagged_words_ratio = st.slider( cutoff_def, 0.0, 1.0, 1.0, step=0.01 ) new_key = ("flagged_words_ratio", cutoff_flagged_words_ratio, True) keys.append(new_key) Visualization.plot_hist(self.docs, new_key) cond = get_cond(new_key[0], new_key[1], new_key[2]) print_discared_by_cond(cond) conds["flagged_words_ratio"] = [cond] if "lang_id_score" in columns: with st.sidebar.expander("Language ID confidence score"): cutoff_def = "If the confidence score for the language identification prediction of a document is lower than this number, the document is removed." cutoff_lang_id_score = st.slider( cutoff_def, 0.0, 1.0, 0.0, step=0.01 ) new_key = ("lang_id_score", cutoff_lang_id_score, False) keys.append(new_key) Visualization.plot_hist(self.docs, new_key) cond = get_cond(new_key[0], new_key[1], new_key[2]) print_discared_by_cond(cond) conds["lang_id_score"] = [cond] if "perplexity_score" in columns: with st.sidebar.expander("Perplexity score"): cutoff_def = "If the perplexity score of a document is higher than this number, the document is removed." max_pp = int(np.max(self.docs["perplexity_score"])) + 1 cutoff_perplexity_score = st.slider(cutoff_def, 0, max_pp, max_pp) new_key = ("perplexity_score", cutoff_perplexity_score, True) keys.append(new_key) Visualization.plot_hist(self.docs, new_key) cond = get_cond(new_key[0], new_key[1], new_key[2]) print_discared_by_cond(cond) conds["perplexity_score"] = [cond] return keys, conds self.keys, conds = set_sliders() self.parameters = self.keys * 1 all_conds = [subcond for cond in list(conds.values()) for subcond in cond] all_conds = np.all(all_conds, axis=0) with st.expander( f"Filtering on documents, for {self.num_docs} {self.lang} documents" ): st.header( f"Filtering on documents, for {self.num_docs} {self.lang} documents" ) def display_dataset(cond, description): displayed_docs = self.docs.loc[cond] st.subheader( f"{description}: {len(displayed_docs)} docs ({len(displayed_docs) / self.num_docs * 100:.2f}%)" ) st.markdown( "Click on a column to sort by it, place the cursor on the text to display it." ) st.dataframe(displayed_docs) display_dataset(np.invert(all_conds), "Discarded documents") # st.subheader("Display discarded documents by filter") display_discarded_documents_by_filter = st.checkbox( "Display discarded documents by filter" ) if display_discarded_documents_by_filter: columns = list(self.docs) if "number_words" in columns: cond_filter = np.invert(np.all(conds["number_words"], axis=0)) display_dataset( cond_filter, "Discarded documents for the filter on the number of words", ) if "repetitions_ratio" in columns: cond_filter = np.invert(np.all(conds["repetitions_ratio"], axis=0)) display_dataset( cond_filter, "Discarded documents for the filter on the repetitions ratio", ) if "special_characters_ratio" in columns: cond_filter = np.invert( np.all(conds["special_characters_ratio"], axis=0) ) display_dataset( cond_filter, "Discarded documents for the filter on the special characters ratio", ) if "stopwords_ratio" in columns: cond_filter = np.invert(np.all(conds["stopwords_ratio"], axis=0)) display_dataset( cond_filter, "Discarded documents for the filter on the stop words ratio", ) if "flagged_words_ratio" in columns: cond_filter = np.invert( np.all(conds["flagged_words_ratio"], axis=0) ) display_dataset( cond_filter, "Discarded documents for the filter on the flagged words ratio", ) if "lang_id_score" in columns: cond_filter = np.invert(np.all(conds["lang_id_score"], axis=0)) display_dataset( cond_filter, "Discarded documents for the filter on the language identification confidence score", ) if "perplexity_score" in columns: cond_filter = np.invert(np.all(conds["perplexity_score"], axis=0)) display_dataset( cond_filter, "Discarded documents for the filter on the perplexity score", ) display_dataset(all_conds, "Retained documents") st.header("Download data") with open(self.path_data) as json_file: btn = st.download_button( label="Download data as json", data=json_file, file_name="data.json", ) def filtering_of_words(self): if not (self.words is None): st.sidebar.subheader("Parameter of the filtering on words") with st.sidebar.expander("Length of words"): cutoff_def = "If the length of a word is higher than this number, the word is removed." max_len_word = min(int(np.max(self.words["len_word"])) + 1, 200) cutoff_word = st.slider(cutoff_def, 0, max_len_word, max_len_word) new_key = ("len_word", cutoff_word, True) self.parameters.append(new_key) Visualization.plot_hist(self.words, new_key) with st.sidebar.expander("Words with incorrect substrings"): incorrect_substrings = st.checkbox( "Remove words with incorrect substrings." ) self.parameters.append(("incorrect_substrings", incorrect_substrings)) cond_words = self.words["len_word"] <= cutoff_word if incorrect_substrings: cond_words = cond_words & np.invert( self.words["incorrect_substring"] ) with st.expander( f"Filtering on words, for {self.num_docs} {self.lang} documents" ): st.header( f"Filtering on words, for {self.num_docs} {self.lang} documents" ) st.markdown( f"Since the number of words is way larger than the number of documents, " f"we consider in this section words for the first {self.num_docs_for_words} documents only." ) discarded_words = self.words.loc[np.invert(cond_words)] st.subheader( f"Discarded words: {len(discarded_words)} words ({len(discarded_words) / len(self.words) * 100:.2f}%)" ) st.markdown( "Click on a column to sort by it, place the cursor on the text to display it." ) st.dataframe(discarded_words) retained_words = self.words.loc[cond_words] st.subheader( f"Retained words: {len(retained_words)} words ({len(retained_words) / len(self.words) * 100:.2f}%)" ) st.markdown( "Click on a column to sort by it, place the cursor on the text to display it." ) st.dataframe(retained_words) def download_parameters(self): st.sidebar.subheader("Download parameters") btn = st.sidebar.download_button( label="Download current parameters as json", data=json.dumps(self.parameters), file_name=f"parameters_{self.lang_dataset_id}.json", ) """ def plot_zipf_law(self): if not (self.words is None): st.header("Zipf's Law") display_zipf_law = st.checkbox("Display Zipf's Law") if display_zipf_law: freq_words = {} for _, row in self.words.iterrows(): freq_words[row["word"]] = freq_words.get(row["word"], 0) + 1 freq_words = np.array(list(freq_words.values())) freq_words = -np.sort(-freq_words) fig, ax = plt.subplots() ax.loglog(freq_words) ax.set_title("Zipf's Law") ax.set_xlabel("$i$-th most frequent word") ax.set_ylabel("frequency in the documents") st.pyplot(fig) """ def analyse_personal_doc(self): with st.expander("Analyse your own document"): st.header("Analyse your own document") personal_doc = st.text_area( label="Paste here the document you want to analyse", value="", max_chars=10000, ) is_discarded = False def is_doc_discarded(key, score): if key[2]: # max cutoff return score > key[1] else: return score < key[1] if personal_doc: st.markdown("Statistics of the document:") for key in self.keys: if key[0] == "number_words": words = ModifyingDocuments.get_words_from_document( personal_doc, self.sentencepiece_model_tok, lower_case=False, strip_characters=self.param["strip_characters"], ) if key[2]: st.markdown(f"Number of words: {len(words)}") if is_doc_discarded(key, len(words)): is_discarded = True elif key[0] == "repetitions_ratio": repetitions_ratio = Filtering.compute_repetitions_ratio( personal_doc, int(key[3]) ) repetitions_ratio = round(repetitions_ratio, 3) st.markdown(f"Repetitions ratio: {repetitions_ratio}") if is_doc_discarded(key, repetitions_ratio): is_discarded = True elif key[0] == "special_characters_ratio": special_characters_ratio = ( Filtering.compute_special_characters_ratio( personal_doc, self.param["special_characters"] ) ) special_characters_ratio = round(special_characters_ratio, 3) st.markdown( f"Special characters ratio: {special_characters_ratio}" ) if is_doc_discarded(key, special_characters_ratio): is_discarded = True elif key[0] == "stopwords_ratio": stopwords_ratio = Filtering.compute_stopwords_ratio( personal_doc, self.sentencepiece_model_tok, self.param["strip_characters"], self.param["cond_words_augmentation"], self.param["words_augmentation_group_sizes"], self.param["words_augmentation_join_char"], self.stopwords, ) stopwords_ratio = round(stopwords_ratio, 3) st.markdown(f"Stop words ratio: {stopwords_ratio}") if is_doc_discarded(key, stopwords_ratio): is_discarded = True elif key[0] == "flagged_words_ratio": flagged_words_ratio = Filtering.compute_flagged_words_ratio( personal_doc, self.sentencepiece_model_tok, self.param["strip_characters"], self.param["cond_words_augmentation"], self.param["words_augmentation_group_sizes"], self.param["words_augmentation_join_char"], self.flagged_words, ) flagged_words_ratio = round(flagged_words_ratio, 3) st.markdown(f"Flagged words ratio: {flagged_words_ratio}") if is_doc_discarded(key, flagged_words_ratio): is_discarded = True elif key[0] == "lang_id_score": ( lang_pred_dataset_id, lang_id_score, ) = Filtering.compute_lang_id_pred_score( personal_doc, self.model_lang_id ) lang_id_score = round(lang_id_score, 3) st.markdown( f"Language identification confidence score: {lang_id_score}" ) if is_doc_discarded(key, flagged_words_ratio) or ( self.lang_dataset_id != lang_pred_dataset_id ): is_discarded = True elif key[0] == "perplexity_score": perplexity_score = Filtering.compute_perplexity_score( personal_doc, self.sentencepiece_model, self.kenlm_model, ) perplexity_score = round(perplexity_score, 3) st.markdown(f"Perplexity score: {perplexity_score}") if is_doc_discarded(key, perplexity_score): is_discarded = True is_discarded = "" if is_discarded else "not " st.markdown( f"With the current filtering parameters, this document **is {is_discarded}discarded**." ) def visualization(self): self.warning_preamble() self.preamble() self.open_data() self.set_title() self.filtering_of_docs() self.filtering_of_words() self.download_parameters() self.analyse_personal_doc() path_instructions = "./explanation_filtering_pipeline.pdf" path_data = "./en_examples_with_stats.json" lang = "English" num_docs = 5000 num_docs_for_words = 500 max_len_text_display = 10000 # Only useful for analyse_personal_doc lang_dataset_id = "en" path_fasttext_model = "./lid.176.bin" path_sentencepiece_model = "./en.sp.model" path_kenlm_model = "./en.arpa.bin" visualization = Visualization( path_instructions, path_data, lang, num_docs, num_docs_for_words, max_len_text_display, lang_dataset_id, path_fasttext_model, path_sentencepiece_model, path_kenlm_model, ) visualization.visualization()