forensic_fasttext: Does 'forensics' on single FastText model

In jcgonzalez14/textwhiz: Support tools for Text Analytics

Description

Use to investigate where your choosen set of parameters for a fasttext model are lacking. Provides:

• investigation to the magnitude of the predicted score for misclassfied texts.

• comparison for the predicted label vs the actual label.

• exposure for mislabelled training data.

Usage

forensic_fasttext(k, texts, text_ids, labels, parameters, seed)

Arguments

k	The number of k-folds for each combination set of parameters to test. Defaults to 5.
texts	The texts given by the user to classify later.
labels	The labels for the texts given by the user train the FastText model.
parameters	A df that contains all the different combinations of paramters for a FastText model. Must include the following: lr = learning rate epoch = # of epochs dim = dimensions ws = window size wordNgrams = word n-grams minn = min of character n-grams maxn = max of character n-grams
seed	A number for set.seed when partioning data for have model reproducability.
texts_ids	The text_ids in the text to output nice clean format.

Value

A dataframe with the average and SD accuracy for each row of parameters.

Examples

fast.text.parameters <- expand.grid(
  lr = seq(4, 4.3, 0.5),
  epoch = seq(30, 33, 10),
  dim = seq(100,120, 25),
  ws = seq(4, 6, 2),
  wordNgrams = 2,
  minn = 2,
  maxn = 6
  )

tune_fasttext(k = 5,
              texts = df$mytext,
              text_ids = df$text_id,
              labels = df$topic,
              parameters = fast.text.parameters,
              seed = 123)

jcgonzalez14/textwhiz documentation built on Aug. 26, 2020, 9:39 a.m.