Algorithm ranking

1. Transformer ru or Transformer multi:

   • Offers good accuracy and high classification speed.
   • Uses a large language model to classify input, analyzing semantic relationships between words rather than just the words themselves.
   • Performs well on small datasets.
   • Can recognize more languages than are officially supported by JAICP.
   • Non-essential words in a phrase might reduce its intent weight. See a possible solution.
   • The Transformer multi might assign high weights to negative examples. See a possible solution.

2. Two algorithms are tied for second place:

   • Classic ML:

     • Analyzes dictionary forms (lemmas) and word stems.
     • Maintains good performance even under high load conditions.
     • Unlike Deep Learning, it works with negative examples.
     • Requires a large dataset.
     • Requires a similar number of phrases for each intent. See a possible solution.
     • Training is time-consuming.
     • Does not consider the semantics of words and synonyms. See a possible solution.

   • Deep Learning:

     • Considers the semantics of words when forming hypotheses.
     • Distinguishes similar intents better than Classic ML.
     • Requires a large dataset. See possible solution for small datasets.
     • Training is time-consuming.
     • Does not handle negative examples well. See a possible solution.

3. STS:

   • Performs well on small datasets. The algorithm can function even if each intent has only one training phrase.
   • Lets you use entities in training phrases.
   • Has the most flexible and interpretable settings.
   • Distinguishes semantically similar intents better than Classic ML.
   • Limited to 1,000 training phrases. It is suitable only for small projects not focused on NLU.
   • Classification is slow. Speed depends on the length of the request, as well as the number and length of the training phrases.