Supporting Development of COVID-19 Assistants

Announcements Important Updates
3

Hola from Córdoba, Argentina!

We have an assistant who responds to COVID-19 questions and questions about the measures the goverment is taking with respect to the pandemic. check it out: asistocovid.com.ar The (not-for-profit) project is going really well we have a huge team of volunteers, journalists, epidemiologists and computer scientists. I’m leading the backend development part with Rasa.

I have done other task-oriented bots with Rasa that worked really well, but this one is different from a design point of view. Because the stories are all short, one question, one answer, it is much simplier than, say, a scheduling bot. However, the challenge is that we have such a high number of intents, more than 50, that it is getting hard to train for a model that knows the difference between “Are the hardware store open?” - “¿Están abiertas las ferreterías?” and “Are the grocery stores open?” “¿Están abiertos los supermercados?” And more unanswered questions come in everyday that we need to incorporate into the system.

Any advice on how to maximize the NLU performance, even if it means ignoring the entity extraction and/ or taking focus away from rasa core (story) training?

Here’s my current pipeline:

language: es

pipeline:

  • name: SpacyNLP
  • name: SpacyTokenizer
  • name: SpacyFeaturizer
  • name: RegexFeaturizer
  • name: CRFEntityExtractor
  • name: EntitySynonymMapper
  • name: SklearnIntentClassifier intent_tokenization_flag: true intent_split_symbol: + policies:
  • name: MemoizationPolicy
  • name: KerasPolicy
  • name: MappingPolicy
  • name: EmbeddingPolicy
  • name: FormPolicy
  • name: TwoStageFallbackPolicy nlu_threshold: 0.2 core_threshold: 0.2 fallback_core_action_name: action_default_fallback fallback_nlu_action_name: action_default_ask_affirmation

Any feedback would be amazing!

Abrazos!

1 Like