Struct TextCnnLite 

pub struct TextCnnLite { /* private fields */ }

Lightweight text-CNN classifier.

Extracts n-gram features (bag-of-n-grams) for each filter size and then trains a multinomial logistic regression head using mini-batch stochastic gradient descent.

This is intentionally a “lite” variant — it approximates a CNN’s 1-D convolution by computing frequency counts of character/word n-grams and using the max-pooled (most-frequent) value as the feature. A proper convolutional network would require a full neural-network framework.

Example

use scirs2_text::text_classification::TextCnnLite;

let mut clf = TextCnnLite::new(vec![2, 3], 8);
let texts  = &["good movie fun", "bad film boring", "great show entertaining", "terrible awful waste"];
let labels = &["pos", "neg", "pos", "neg"];
clf.fit(texts, labels, 20).unwrap();

Implementations

impl TextCnnLite

pub fn new(filter_sizes: Vec<usize>, n_filters: usize) -> Self

Create a new TextCnnLite.

Parameters

• filter_sizes: n-gram window sizes, e.g. [2, 3, 4].
• n_filters: number of top n-gram features to keep per filter size.

pub fn fit( &mut self, texts: &[&str], labels: &[&str], epochs: usize, ) -> Result<()>

Train the classifier.

Internally:

1. Build the n-gram vocabulary from all documents.
2. Vectorise each document into a frequency vector.
3. Train multinomial logistic regression with SGD.

Errors

Returns TextError::InvalidInput when texts and labels lengths differ or when the corpus is empty.

pub fn predict(&self, text: &str) -> Result<String>

Predict the most probable class for text.

Errors

Returns TextError::ModelNotFitted when the model has not been trained.