Guidance and Best Practices
To build the best possible intelligence, meaning a dataset that is accurate in its predictions, we must follow certain best practices when creating training phrases.
In this article, we will learn about some training methods and best practices in Weni.
Main Guidelines
Within best practices, there are some key guidelines we need to follow:
- Number of Phrases
- Balanced Number of Phrases
- Vocabulary Specificity
- Variation in Sentence Structures
Each of these topics is explained below.
Number of Phrases
Most NLP algorithm models rely on the number of training examples to increase the prediction rate for each intent. Therefore, to achieve high accuracy, we need to balance the relationship between the number of phrases and the number of intents in your dataset.
Below are some classifications of dataset quality based on the number of training phrases per intent, for an example with five intents or fewer:
- Minimum: 10 phrases per intent;
- Good: 25 phrases per intent;
- Excellent: 40 phrases per intent.
Some factors may influence these numbers, such as the total number of intents in the intelligence (which can affect the number of false positives). The more intents there are, the more phrases are required per intent.
The chosen algorithm also impacts this number. For instance, the algorithm using BERT, due to its use of a pre-trained model, tends to need fewer phrases to achieve good results.
Balancing
Using a balanced number of phrases across all intents of your intelligence reduces the chances of bias toward a specific intent.
For example, if the intelligence has an intent X with 50 phrases and an intent Y with 200 phrases, the probability that the algorithm classifies inputs as belonging to intent Y might be higher because it has more examples (considering that the input is a new phrase never seen during training).
Thus, a good practice is to have an approximately equal number of phrases for all intents in your dataset, if possible.
Vocabulary Specificity
To reduce the number of false positives in the dataset and increase precision, we recommend that the training phrases generated respect the topic specificity rule.
This rule dictates that all specific words related to an intent should only be added to the phrases for that intent, and words that should not be interpreted as belonging to any intent should be distributed across all intents, so the algorithm does not associate those words with any specific topic.
For example, if I have intelligence that identifies orders in a fast-food restaurant, with the intents "food" and "drinks," I need to associate words related to each intent, such as "sandwich" for the first and "juice" for the second.
Thus, we would generate training phrases like "I would like to buy a sandwich" for the intent "food" and "I want to buy a juice" for the intent "drinks."
Note that specific words like "sandwich" and "juice" are associated with individual intents, while words like "I would like," "want," "to," "buy," and "a" are distributed between the two intents, ensuring that if someone types just "I would like to buy," the intelligence will not associate it with either intent due to very low confidence.
Variation in Sentence Structures
Sentence structure is also an important factor in interpreting user input. For example, if the phrase "I would like to eat a pizza" is trained for the intent "food," the algorithm would classify the phrase "I would love to eat a pizza" as the same intent, given that the sentence structure is similar (assuming there is a sufficient number of trained phrases with this structure).
This means that the more varied the example phrases are, both in terms of structures and words, the greater the probability that the intelligence will accurately predict more words related to that intent.
