from collections import Counter import math import os import gradio as gr from datasets import load_dataset from nltk.util import ngrams import pandas as pd import plotly.express as px import plotly.graph_objects as go from plotly.subplots import make_subplots from matplotlib import pyplot as plt from wordcloud import WordCloud from huggingface_hub import InferenceClient import matplotlib matplotlib.use("agg") def load_transform_dataset(): # Load the dataset and convert it to a Pandas dataframe sotu_dataset = "jsulz/state-of-the-union-addresses" dataset = load_dataset(sotu_dataset) _df = dataset["train"].to_pandas() # Do some on-the-fly calculations # calcualte the number of words in each address _df["word_count"] = _df["speech_html"].apply(lambda x: len(x.split())) # calculate the automated readibility index reading ease score for each address # automated readability index = 4.71 * (characters/words) + 0.5 * (words/sentences) - 21.43 _df["ari"] = _df["no-contractions"].apply( lambda x: (4.71 * (len(x.replace(" ", "")) / len(x.split()))) + (0.5 * (len(x.split()) / len(x.split(".")))) - 21.43 ) # create a column that is the year the speach was given from the date column _df["year"] = _df["date"].dt.year # create a column that is a concatenation of the president's name, year, and category _df["speech_key"] = ( _df["potus"] + " - " + _df["year"].astype(str) + " (" + _df["categories"] + ")" ) # Sort the dataframe by date because Plotly doesn't do any of this automatically _df = _df.sort_values(by="date") _written = _df[_df["categories"] == "Written"] _spoken = _df[_df["categories"] == "Spoken"] return _df, _written, _spoken """ Helper functions for Plotly charts """ def filter_potus(potus, _df): if potus != "All": # Filter on the potus potus_df = _df[_df["potus"] == potus] else: potus_df = _df return potus_df def plotly_ngrams(n_grams, potus, _df): if potus is not None: potus_df = filter_potus(potus, _df) # Create a counter generator for the n-grams trigrams = ( potus_df["tokens-nostop"] .apply(lambda x: list(ngrams(x, n_grams))) .apply(Counter) .sum() ) # get the most common trigrams common_trigrams = trigrams.most_common(10) # unzip the list of tuples and plot the trigrams and counts as a bar chart trigrams, counts = zip(*common_trigrams) # join the trigrams into a single string trigrams = [" ".join(trigram) for trigram in trigrams] # create a dataframe from the trigrams and counts trigrams_df = pd.DataFrame({"trigrams": trigrams, "counts": counts}) if potus == "All": potus = "All Presidents" fig4 = px.bar( trigrams_df, x="counts", y="trigrams", title=f"{potus}'s top {n_grams}-grams", orientation="h", height=400, ) return fig4 def plotly_word_and_ari(president, _df): potus_df = filter_potus(president, _df) fig5 = make_subplots(specs=[[{"secondary_y": True}]]) fig5.add_trace( go.Scatter( x=potus_df["date"], y=potus_df["word_count"], name="Word Count", ), secondary_y=False, ) fig5.add_trace( go.Scatter( x=potus_df["date"], y=potus_df["ari"], name="ARI", ), secondary_y=True, ) # Add figure title fig5.update_layout(title_text="Address Word Count and ARI") # Set x-axis title fig5.update_xaxes(title_text="Date of Address") # Set y-axes titles fig5.update_yaxes(title_text="Word Count", secondary_y=False) fig5.update_yaxes(title_text="ARI", secondary_y=True) return fig5 def plt_wordcloud(president, _df): potus_df = filter_potus(president, _df) lemmatized = potus_df["lemmatized"].apply(lambda x: " ".join(x)) # build a single string from lemmatized lemmatized = " ".join(lemmatized) # create a wordcloud from the lemmatized column of the dataframe wordcloud = WordCloud(background_color="white", width=800, height=400).generate( lemmatized ) # create a matplotlib figure fig6 = plt.figure(figsize=(8, 4)) # add the wordcloud to the figure plt.tight_layout() plt.imshow(wordcloud, interpolation="bilinear") plt.axis("off") return fig6 def summarization(speech_key, _df): client = InferenceClient(model="facebook/bart-large-cnn") chunk_len = 4000 speech = _df[_df["speech_key"] == speech_key]["speech_html"].values[0] sotu_chunks = int(math.ceil(len(speech) / chunk_len)) response = [] for chunk in range(1, sotu_chunks + 1): if chunk * 4000 < len(speech): chunk_text = speech[(chunk - 1) * chunk_len : chunk * chunk_len] else: chunk_text = speech[(chunk - 1) * chunk_len :] try: summarization_chunk = client.summarization( chunk_text, parameters={"truncation": "do_not_truncate"} ) except Exception as e: print(e) response.append(summarization_chunk.summary_text) return "\n\n".join(response) def streaming(speech_key, _df): client = InferenceClient(token=os.environ["HF_TOKEN"]) speech = _df[_df["speech_key"] == speech_key]["speech_html"].values[0] speech_info = speech_key.split(" - ") messages = [] for message in client.chat_completion( model="Qwen/Qwen2.5-72B-Instruct", messages=[ { "role": "system", "content": "You are a political scholar with a deep knowledge of State of the Union addresses. You are tasked with summarizing a speech from a given president. The speech is a mix of written and spoken addresses. The goal is to provide a concise summary of the speech with the proper historical and political context.", }, { "role": "user", "content": f"The following speech is a State of the Union address from {speech_info[0]} on {speech_info[1]}. Summarize it: {speech}", }, ], max_tokens=700, stream=True, ): # yield message.choices[0].delta.content # print(message) messages.append(message.choices[0].delta.content) return "".join(messages) # Create a Gradio interface with blocks with gr.Blocks() as demo: df, written, spoken = load_transform_dataset() # store the dataframe in a state object before passing to component functions df_state = gr.State(df) # Build out the top level static charts and content gr.Markdown( """ # An Interactive Dashboard for State of the Union Addresses This dashboard provides an analysis of all State of the Union (SOTU) addresses from 1790 to 2020 including written and spoken addresses. The data is sourced from the [State of the Union Addresses dataset](https://huggingface.co/datasets/jsulz/state-of-the-union-addresses) on the Hugging Face Datasets Hub. You can read more about how the data was gathered and cleaned on the dataset card. To read the speeches, you can visit the [The American Presidency Project's State of the Union page](https://www.presidency.ucsb.edu/documents/presidential-documents-archive-guidebook/annual-messages-congress-the-state-the-union) where this data was sourced. """ ) gr.Markdown( "In addition to analyzing the content, this space also leverages the [Qwen/2.5-72B-Instruct](https://deepinfra.com/Qwen/Qwen2.5-72B-Instruct) model to summarize a speech. The model is tasked with providing a concise summary of a speech from a given president. To get a summary, go to the 'Summarize a Speech' tab." ) with gr.Tab(label="Speech Data"): # Basic line chart showing the total number of words in each address gr.Markdown( """ ## The shape of words The line chart to the right shows the total number of words in each address. However, not all SOTUs are created equally. From 1801 to 1916, each address was a written message to Congress. In 1913, Woodrow Wilson broke with tradition and delivered his address in person. Since then, the addresses have been a mix of written and spoken (mostly spoken). The spikes you see in the early 1970's and early 1980's are from written addresses by Richard Nixon and Jimmy Carter respectively. Now that we have a little historical context, what does this data look like if we split things out by president? The bar chart below shows the average number of words in each address by president. The bars are grouped by written and spoken addresses. """ ) fig1 = px.line( df, x="date", y="word_count", title="Total Number of Words in Addresses", line_shape="spline", ) fig1.update_layout( xaxis=dict(title="Date of Address"), yaxis=dict(title="Word Count"), ) gr.Plot(fig1, scale=2) # group by president and category and calculate the average word count sort by date avg_word_count = ( df.groupby(["potus", "categories"])["word_count"].mean().reset_index() ) # Build a bar chart showing the average number of words in each address by president fig2 = px.bar( avg_word_count, x="potus", y="word_count", title="Average Number of Words in Addresses by President", color="categories", barmode="group", ) fig2.update_layout( xaxis=dict( title="President", tickangle=-45, # Rotate labels 45 degrees counterclockwise ), yaxis=dict( title="Average Word Count", tickangle=0, # Default label angle (horizontal) ), legend=dict( orientation="h", yanchor="bottom", y=1.02, xanchor="right", x=1 ), ) gr.Plot(fig2) # Create a line chart showing the Automated Readability Index in each address with gr.Row(): ari = df[["potus", "date", "ari", "categories"]] fig3 = px.line( ari, x="date", y="ari", title="Automated Readability Index in each Address", line_shape="spline", ) fig3.update_layout( xaxis=dict(title="Date of Address"), yaxis=dict(title="ARI Score"), ) gr.Plot(fig3, scale=2) gr.Markdown( """ The line chart to the left shows the Automated Redibility Index (ARI) for each speech by year. The ARI is calculated using the formula: 4.71 * (characters/words) + 0.5 * (words/sentences) - 21.43. In general, ARI scores correspond to U.S. grade levels. For example, an ARI of 8.0 corresponds to an 8th grade reading level. While there are other scores that are more representative of attributes we might want to measure, they require values like syllables. The ARI is a simple score to compute with our data. The drop off is quite noticeable, don't you think? ;) """ ) gr.Markdown( """ ## Dive Deeper on Each President Use the dropdown to select a president a go a little deeper. To begin with, there is an [n-gram](https://en.wikipedia.org/wiki/N-gram) bar chart built from all of the given president's addresses. An n-gram is a contiguous sequence of n items from a given sample of text or speech. Because written and spoken speech is littered with so-called "stop words" such as "and", "the", and "but", they've been removed to provide a more rich (albeit sometimes more difficult to read) view of the text. The slider only goes up to 4-grams because the data is sparse beyond that. I personally found the n-grams from our last three presidents to be less than inspiring and full of platitudes. Earlier presidents have more interesting n-grams. Next up is a word cloud of the lemmatized text from the president's addresses. [Lemmatization](https://en.wikipedia.org/wiki/Lemmatization) is the process of grouping together the inflected forms of a word so they can be analyzed as a single item. Think of this as a more advanced version of [stemming](https://en.wikipedia.org/wiki/Stemming) where we can establish novel links between words like "better" and "good" that might otherwise be overlooked in stemming. You can also see a line chart of word count and ARI for each address. """ ) # get all unique president names presidents = df["potus"].unique() presidents = presidents.tolist() presidents.append("All") # create a dropdown to select a president president = gr.Dropdown( label="Select a President", choices=presidents, value="All" ) # create a text area to display the summarized speech # create a slider for number of word grams grams = gr.Slider( minimum=1, maximum=4, step=1, label="N-grams", interactive=True, value=1 ) # show a bar chart of the top n-grams for a selected president gr.Plot(plotly_ngrams, inputs=[grams, president, df_state]) gr.Plot(plt_wordcloud, scale=2, inputs=[president, df_state]) # show a line chart of word count and ARI for a selected president gr.Plot(plotly_word_and_ari, inputs=[president, df_state]) with gr.Tab(label="Summarize a Speech"): gr.Markdown("## Summarize a Speech") gr.Markdown( """ Context is king; get a summary of a State of the Union now that you've seen a bit more. Use the dropdown to select a speech from a president and click the button to summarize the speech. [Qwen/2.5-72B-Instruct](https://deepinfra.com/Qwen/Qwen2.5-72B-Instruct) will provide a concise summary of the speech with the proper historical and political context. """ ) speeches = df["speech_key"].unique() speeches = speeches.tolist() speech = gr.Dropdown(label="Select a Speech", choices=speeches) # create a dropdown to select a speech from a president run_summarization = gr.Button(value="Summarize") fin_speech = gr.Textbox(label="Summarized Speech", type="text", lines=10) run_summarization.click( streaming, inputs=[speech, df_state], outputs=[fin_speech] ) demo.launch()