如何用purrr提取textmodel_nb输出的类别概率并转为data_frame？

Let's break down how to convert that PcGw matrix into the tidy data frame you need, using purrr and other tidyverse tools (since purrr works seamlessly with them):

Step-by-Step Code #

First, make sure you have the tidyverse package loaded alongside quanteda:

library(quanteda)
library(tidyverse)

Then run your original example code to get the nb_test object, then use this code to reshape the PcGw matrix:

# Reshape the PcGw matrix into the desired data frame
tidy_probabilities <- nb_test$PcGw %>%
  # Convert matrix to data frame, keeping row names (your classes)
  as.data.frame() %>%
  # Extract row names into a dedicated "class" column
  rownames_to_column(var = "class") %>%
  # Convert from wide to long format: one row per class-feature pair
  pivot_longer(cols = -class, names_to = "variable", values_to = "probability") %>%
  # Convert back to wide format: one column per class (with P_ prefix)
  pivot_wider(names_from = class, values_from = probability, names_prefix = "P_")

What This Does #

Let's walk through each step with your example:

1. as.data.frame(): Converts the PcGw matrix (rows = classes, columns = features) into a data frame where row names are your class labels (Y/N in your example, TRUE/FALSE in your actual data).
2. rownames_to_column(): Moves those class labels from row names into a proper column called class—this makes it easy to reshape later.
3. pivot_longer(): Flattens the data so each row represents a single class-feature-probability combination (e.g., one row for Y + Chinese + its probability, another for N + Chinese + its probability).
4. pivot_wider(): Pivots the data back to wide format, creating separate columns for each class (prefixed with P_ to match your P_TRUE/P_FALSE naming).

If You Want a More purrr-Focused Approach #

If you prefer to lean more heavily on purrr instead of tidyr's pivot functions, you can use imap_dfr to iterate over the matrix columns directly:

tidy_probabilities <- imap_dfr(as.data.frame(nb_test$PcGw), 
                               ~ tibble(variable = .y, 
                                        class = rownames(nb_test$PcGw), 
                                        probability = .x)) %>%
  pivot_wider(names_from = class, values_from = probability, names_prefix = "P_")

Here, imap_dfr loops through each column (feature) of the matrix:

• .y gives the column name (your feature/variable)
• .x gives the values in the column (probabilities for each class)
  We combine these into a tibble, then pivot to wide format as before.

Result for Your Example #

Running either code on your sample data will produce a data frame like this:

# A tibble: 6 × 3
  variable   P_Y      P_N     
  <chr>      <dbl>    <dbl>
1 Chinese    0.768    0.232
2 Beijing    0.113    0     
3 Shanghai   0.113    0     
4 Macao      0.113    0     
5 Tokyo      0.0377   0.384
6 Japan      0.0377   0.384

In your actual data with TRUE/FALSE classes, the columns will automatically be P_TRUE and P_FALSE instead of P_Y/P_N.

内容的提问来源于stack exchange，提问作者ℕʘʘḆḽḘ