Why rasa bert cannot handle singular and plural words?

I used rasa bert to train a model, but it can’t handle the singular and plural of words, as in the following two sentences: A:How to search in notes B:How to search in note These two sentences are only different from “note”, but the predicted intention is completely different. I got the vectors of A and B by debugging the code. As follows:

A-batch_sequence_features.0.txt (74.7 KB) A-batch_sentence_features.txt (8.5 KB)

B-batch_sentence_features.txt (8.6 KB) B-batch_sequence_features.0.txt (74.7 KB)

I compared the similarity between A-batch_sentence_features and B-batch_sentence_features, and it can reach a similarity of more than 95%, but why is the recognized intention different? I need to solve this problem urgently, please help me, thank you!

Without knowing your pipeline and the intents combined with some examples, nobody will be able to give you a good answer.

Hi @harloc My pipeline is:

# Configuration for Rasa NLU.
# https://rasa.com/docs/rasa/nlu/components/
language: en

# # No configuration for the NLU pipeline was provided. The following default pipeline was used to train your model.
# # If you'd like to customize it, uncomment and adjust the pipeline.
# # See https://rasa.com/docs/rasa/tuning-your-model for more information.
   - name: WhitespaceTokenizer
   - name: RegexFeaturizer
   - name: LexicalSyntacticFeaturizer
   - name: LanguageModelFeaturizer
     model_name: "bert"
     model_weights: "bert-base-uncased"
     cache_dir: null
   - name: DIETClassifier
     epochs: 100
   - name: EntitySynonymMapper
   - name: ResponseSelector
     epochs: 100
   - name: FallbackClassifier
     threshold: 0.8
     ambiguity_threshold: 0.1

# Configuration for Rasa Core.
# https://rasa.com/docs/rasa/core/policies/
# # No configuration for policies was provided. The following default policies were used to train your model.
# # If you'd like to customize them, uncomment and adjust the policies.
# # See https://rasa.com/docs/rasa/policies for more information.
   - name: AugmentedMemoizationPolicy
   - name: TEDPolicy
     max_history: 8
     epochs: 200
       dialogue: [256, 128]
   - name: RulePolicy
     core_fallback_threshold: 0.3
     core_fallback_action_name: "action_default_fallback"
     enable_fallback_prediction: True

@gongshaojie12 We will also need the intents and some training examples for each

Hi @tyd We have 51 intentions, and it is data on the production environment. I’m sorry that we can’t make it public.

B sentence is in the training data, A is not in the training data, the result of the current model is that B can be correctly classified, A can not be correctly classified, and a is recognized as other_intent.The intention of a and B should be consistent,. And the semantic similarity between the training data in this “other_intent” and A is less than 60% . So now I don’t understand why dietclassifier mistakenly classifies A.