aies-nlp-preprocessing-tk

CSV Format Documentation

Single Label Format

The single label format is designed for datasets where each document is associated with a single class. The CSV file must adhere to the following structure:

• CLASS: Represents the label or category of the document.
• Text of document: The actual text content of the document.

Each row in the CSV file corresponds to a single document and its associated class. The CSV must contain exactly two columns: the first column for the class and the second column for the text of the document.

Multi Label Format

The multi label format is intended for datasets where each document can be associated with multiple classes. The CSV file must follow this structure:

• CLASS: Represents a single label or category of the document.
• CLASS|CLASS|CLASS: Represents multiple labels or categories separated by the | character.
• Text of document: The actual text content of the document.

Each row in the CSV file corresponds to a single document and its associated classes. The CSV must contain exactly two columns: the first column for the classes and the second column for the text of the document. For multi-label documents, multiple classes must be separated by the | character with no leading or trailing | characters.

Example CSV Files

Single Label Example

greeting,"Hello!" 
greeting,"Good morning!" 
question,"How are you?"

Multi Label Example

greeting,"Hello!" 
greeting|question,"Hi, how are you?" 
question|feedback,"What do you think of this service?"

Validation Rules

• Single Label: The tag column must contain only one class per row. No | character should be present.
• Multi Label: The tag column can contain multiple classes separated by |. Ensure there are no empty classes and no leading or trailing | characters.

raw_tokenization

Tokenizes text data and prepares it for training a neural network model.

Text Tokenization:

The input textual data is tokenized, i.e., split into individual words or tokens. This is done using the spaCy library, which provides robust linguistic annotations.

Text Cleaning:

• Removing Punctuation: If specified, punctuation marks are removed from the text. This helps in simplifying the text and reducing noise in the data.
• Removing Stop Words: Optionally, common stop words (e.g., 'and', 'the', 'is') can be removed from the text. Stop words often carry little semantic meaning and can be safely excluded from the analysis.

Padding Sequences:

Text sequences are padded to ensure uniform length. This is necessary for feeding the data into a neural network, as they typically require fixed-size inputs. Padding is done using the Keras pad_sequences function.

Label Encoding:

If the labels are categorical, they are encoded using either LabelEncoder or MultiLabelBinarizer from the scikit-learn library. This step converts textual labels into numerical representations, which are easier for the neural network to process.

Data Splitting:

Optionally, the preprocessed data can be split into training and testing sets using train_test_split from scikit-learn. This facilitates model evaluation by providing a separate dataset for testing.

Parameters

• data_frame (pandas.DataFrame): The input DataFrame containing 'text' and 'tag' columns.
• max_length (int): Maximum length of sequences after padding. (Important! Critical variable, depending on the value entered, can cause erroneous operation, disrupting tokenization, occurs mainly when using low values, to avoid using values ​​above 100)
• split_test_size (float, optional): Size of the test dataset if splitting is needed. Defaults to None.
• remove_stop_words (bool, optional): Whether to remove stop words. Defaults to False.
• remove_punctuation (bool, optional): Whether to remove punctuation. Defaults to False.
• language (str, optional): Language to be used for tokenization. Defaults to "portuguese".

Returns

• tuple: If the split_test_size parameter is passed, the separation into training and test masses will be done and it will return the following results:

  X_train, X_test, y_train, y_test, word_index
  

  If split_test_size=False, it would return the datasets only from the tokenization of the texts, the labels will be hidden and the word index will be returned:

  X, y, word_index