Optimizing Intent

Pretrained Embeddings: Intent Classifier Sklearn

What is it?

• Uses spacy

• Uses word embeddings (vector representations of words)

• Similar words get converted to similar numeric matrices

• Trains linear SVM - optimized with gridsearch (hyperparameter tuning to determine optimal values for model)

• Research paper: word2vec

• Depends on langauge (choose spacy's "en" for english)

  • Can customize word embeddings (look into Facebook's fastText)

  • Then link model (spaCy guide)

# Linking spaCy model with word embedding
python -m spacy link <converted model> <language code>

Pros and Cons

• Not much data needed

• Pretrained word embeddings

• Not available outside of English

• Does not cover domain specific words (product names, acrynoms)

Supervised Embeddings: Intent Classifier TensorFlow Embedding

What is it?

Pipeline name in Rasa: EmbeddingIntentClassifier

• trains word embeddings from scratch

• Typically used with CountVectorsFeaturizer

  • Counts how often distinct words in training data appear in message

  • Then feeds this vector to intent classifier

• Another count vector created for intent label (for multiple intent support)

# Example:
Training data: ["best", "bot", "great", "is", "my", "the"]
"My bot is the best bot." = [1, 2, 0, 1, 1, 1]
"My bot is great." = [0, 1, 1, 1, 1, 0]

Configuring CountVectorsFeaturizer (recommended featurizer for this classifier):

• change analyzer property to char (using ngram counts instead of word token counts)

  • More robust against typos

Note: Bag-of-words is unigram, n-grams lets you detect multiple words together (if its bigrams you can detect United States, not just United)

Pros and Cons

• Supports multiple intents per message

• Needs more training data (learning from scratch)

• Language independent

Common Problems

Lack of Training Data

• Try Chatito

• Try Rasa Core Interactive

Out-of-Vocabulary Words

• Typos, words you did not think of

  • If using EmbeddingIntentClassifier:

    • More training data

    • Add examples that include OOV_token (out-of-vocab token)

Similar Intents

• Eg. You have 2 intents: provide_name and provide_data

• provide_name: It is Laura

• provide_date: It is on Monday

• These 2 are similar from NLU perspective - comine into one intent called inform

Skewed data

• Maintain equal # of examples per intent