Headlines Analysis

Mix.install(
  [
    {:req, "~> 0.4.4"},
    {:kino_bumblebee, "~> 0.4.0"},
    {:exla, ">= 0.0.0"}
  ],
  config: [nx: [default_backend: EXLA.Backend]]
)

Section

defmodule Article do
  defstruct [
    :author,
    :content,
    :description,
    :published_at,
    :source,
    :title,
    :url,
    :url_to_image,
    :entities,
    :sentiment
  ]

  def parse(body) do
    %Article{
      author: body["author"],
      content: body["content"],
      description: body["description"],
      published_at: body["publishedAt"],
      source: get_in(body, ["source", "id"]),
      title: body["title"],
      url: body["url"],
      url_to_image: body["urlToImage"]
    }
  end
end

{:ok, %{status: 200, body: body}} =
  "https://newsapi.org/v2/top-headlines"
  |> Req.get(params: %{country: "us", apiKey: System.get_env("LB_NEWS_API_KEY")})

articles =
  body["articles"]
  |> Enum.reject(&(&1["content"] == nil))
  |> Enum.map(&Article.parse/1)

{:ok, model_info} = Bumblebee.load_model({:hf, "dslim/bert-base-NER"})
{:ok, tokenizer} = Bumblebee.load_tokenizer({:hf, "bert-base-cased"})

serving =
  Bumblebee.Text.token_classification(model_info, tokenizer,
    aggregation: :same,
    compile: [batch_size: 1, sequence_length: 128],
    defn_options: [compiler: EXLA]
  )

entities = Nx.Serving.run(serving, Enum.map(articles, & &1.title))

articles =
  articles
  |> Enum.zip_with(entities, fn article, %{entities: entities} ->
    %{article | entities: entities}
  end)

{:ok, model_info} = Bumblebee.load_model({:hf, "ProsusAI/finbert"})
{:ok, tokenizer} = Bumblebee.load_tokenizer({:hf, "bert-base-uncased"})

finbert_serving =
  Bumblebee.Text.text_classification(model_info, tokenizer,
    compile: [batch_size: 1, sequence_length: 128],
    defn_options: [compiler: EXLA]
  )

predictions = Nx.Serving.run(finbert_serving, Enum.map(articles, & &1.title))

articles =
  articles
  |> Enum.zip_with(predictions, fn article, %{predictions: prediction} ->
    %{label: max_label} = Enum.max_by(prediction, & &1.score)
    %{article | sentiment: max_label}
  end)

frame = Kino.Frame.new()

articles
|> Enum.each(fn article ->
  Kino.Frame.append(frame, Kino.Bumblebee.HighlightedText.new(article.title, article.entities))
  Kino.Frame.append(frame, Kino.Text.new(article.sentiment))
end)

Kino.Layout.grid([frame], boxed: true, gap: 16)