|
defmodule MedicalTranscriptionWeb.Components.TranscriptionTextComponent do |
|
@moduledoc """ |
|
Represents a portion of transcribed text and its codes and starts a task to determine keywords within the text. |
|
""" |
|
|
|
use MedicalTranscriptionWeb, :live_component |
|
import MedicalTranscriptionWeb.Components |
|
import MedicalTranscriptionWeb.Components.KeywordHighlighter |
|
alias MedicalTranscription.Coding.CodeVectorMatch |
|
alias AudioTagger.KeywordFinder |
|
|
|
@impl Phoenix.LiveComponent |
|
def update(assigns, socket) do |
|
socket = assign_for_row(assigns, socket) |
|
|
|
{:ok, socket} |
|
end |
|
|
|
defp assign_for_row(%{row: row}, socket) do |
|
self_pid = self() |
|
|
|
socket |
|
|> assign(:row, row) |
|
|> assign_async(:keywords, fn -> find_keywords(self_pid, row.text) end) |
|
end |
|
|
|
defp assign_for_row(_assigns, socket), do: socket |
|
|
|
@impl Phoenix.LiveComponent |
|
def render(assigns) do |
|
~H""" |
|
<div class="flex gap-12 pb-10 border-b border-[ |
|
<div class="flex-1 flex flex-col gap-4"> |
|
<p class="text-[32px] leading-normal font-semibold"> |
|
<%= if !is_nil(@row.start_mark) && !is_nil(@row.end_mark) do %> |
|
<%= @row.start_mark %> - <%= @row.end_mark %> |
|
<% end %> |
|
</p> |
|
<p class="text-[28px] leading-normal text-type-black-tertiary"> |
|
<.async_result :let={keywords} assign={@keywords}> |
|
<:loading><%= @row.text %></:loading> |
|
<.highlight text={@row.text} keywords={keywords} /> |
|
</.async_result> |
|
</p> |
|
|
|
<div class="flex items-center gap-2 text-sm italic"> |
|
<.async_result assign={@keywords}> |
|
<:loading> |
|
<img src={~p"/images/loading.svg"} width="16" class="animate-spin" /> |
|
Finding keywords... |
|
</:loading> |
|
<:failed :let={reason}>There was an error finding keywords: <%= reason %></:failed> |
|
</.async_result> |
|
</div> |
|
</div> |
|
|
|
<div class="flex-1 flex flex-col items-stretch gap-3"> |
|
<%= for %CodeVectorMatch{id: id, code: code, description: label, cosine_similarity: score, weighting: weighting} <- @row.tags do %> |
|
<.tag_result |
|
code_vector_id={id} |
|
code={code} |
|
label={label} |
|
score={score} |
|
text={@row.text} |
|
weighting={weighting} |
|
/> |
|
<% end %> |
|
</div> |
|
</div> |
|
""" |
|
end |
|
|
|
defp find_keywords(_live_view_pid, ""), do: {:ok, %{keywords: []}} |
|
|
|
defp find_keywords(live_view_pid, text) do |
|
# First, we use token classification to determine parts of speech and then retrieve the verb and adjective+noun phrases. |
|
%{entities: entities} = |
|
Nx.Serving.batched_run(MedicalTranscription.TokenClassificationServing, text) |
|
|
|
phrases = KeywordFinder.cleanup_phrases(entities) |
|
|
|
# Then, we use one of two processes to determine which to show as keywords: |
|
# 1. A slower process that looks to classify the text by the extracted phrases. |
|
# serving = KeywordFinder.prepare_zero_shot_classification_serving(phrases) |
|
# %{predictions: predictions} = Nx.Serving.run(serving, text) |
|
|
|
# 2. A fast process finding the phrase closest in vector space to the whole text. |
|
predictions = KeywordFinder.find_most_similar_label(text, phrases, 2) |
|
|
|
# For now, retrieve the top three keywords that have a score of more than 0.25 |
|
keywords = |
|
predictions |
|
|> Enum.filter(fn keyword -> keyword.score > 0.25 end) |
|
|> Enum.take(3) |
|
|
|
send(live_view_pid, {:new_keywords, predictions}) |
|
|
|
{:ok, %{keywords: keywords}} |
|
end |
|
end |
|
|
