Advanced NLP: training data building

Avoid missleading intent names

Keeping the intent name understandable is a part of success. There's always someone, who'll continue in your work.

bad

# (intent: utterance)
advice: how much make a chatbot
advice: how to build a bot
duration: how much time it takes to build a bot
duration: how does it take to make a chatbot
long: how long the message should be
long: what is the best length of text message

Also categorization of intents helps a lot.

good

# (intent: utterance)
faq-howToMakeBot: how much make a chatbot
faq-howToMakeBot: how to build a bot
faq-buildDuration: how much time it takes to build a bot
faq-buildDuration: how does it take to make a chatbot
conversation-bestTextLength: how long the message should be
conversation-bestTextLength: what is the best length of text message

Avoid general intents

Putting all utterances which leads into the same answer under a same intent is a very common mistake.

bad

# (intent: utterance)
aboutCats: how much a cat eats
aboutCats: how much food a cat consumes
aboutCats: how much a cat weights
aboutCats: how many kilos a cat has

good

# (intent: utterance)
howMuchCatEats: how much a cat eats
howMuchCatEats: how much food a cat consumes
howMuchCatWeights: how much a cat weights
howMuchCatWeights: how many kilos a cat has

Entities: Provide as much synonyms as possible

Recognition of entities can be little bit tricky. Add as much synonyms as possible. Also common misspels helps a lot.

bad

# (@entity: value [synonyms, ...])
@salutation: bro [brother]

good

# (@entity: value [synonyms, ...])
@salutation: bro [brother, broo, broo, brotha]

Try to mix long and short utterances

Some people are more expressive, some not. Make the NLP usable for the both groups.

bad

# (intent: utterance)
howMuchCatEats: how much a cat eats
howMuchCatEats: how much food a cat consumes
howMuchCatEats: what quantity of food the cat eats
howMuchCatWeights: cat weight
howMuchCatWeights: cat kilos

good

# (intent: utterance)
howMuchCatEats: how much a cat eats
howMuchCatEats: how much food a cat consumes
howMuchCatEats: what quantity of food the cat eats
howMuchCatEats: cat food quantity
howMuchCatEats: cat food consupmtion
howMuchCatWeights: how much a cat weights
howMuchCatWeights: how many kilos a cat has
howMuchCatWeights: cat weight
howMuchCatWeights: cat kilos

Avoid missleading words in intent utterances

Misleading words in combination with a disproportion between intents results in suprising false-positive intent recognition.

bad

# (intent: utterance)
whatTimeIsIt: please tell me, what time is
whatTimeIsIt: hello, can you tell me, what time do you have
whatTimeIsIt: I`am wondering, what could be hours
whatTimeIsIt: where is my watch
whatTimeIsIt: oh no, I completely lost track of time

good

# (intent: utterance)
whatTimeIsIt: what time is it
whatTimeIsIt: tell me, what time is
whatTimeIsIt: what time do you have
whatTimeIsIt: what could be hours

Avoid disproportion between intents

Disproportions makes a single intent much more likely to be false-positive matched.

  • all intents should have similar number of utterances
    • largest intent should have twice as many utterances than the smallest
    • good numbers of utterances are 6-12 for simple bots, 12-24 for more sophisticated bots
  • distribute common phrases over all intents

bad

the intent whereIsATiger has just short utterances and the intent whereIsFoodCourt only long utterances

# (intent: utterance)
whereIsATiger: where is a tiger
whereIsATiger: tiger where
whereIsFoodCourt: where can i find a foodcourt
whereIsFoodCourt: how to get to foodcourt
whereIsFoodCourt: where the food court is?
whereIsFoodCourt: where is a food court?
whereIsFoodCourt: foodcourt where?

good

# (intent: utterance)
whereIsATiger: where is a tiger
whereIsATiger: tiger where
whereIsATiger: where can i find a tiger
whereIsATiger: how to get to tigers cage
whereIsATiger: where the tiger is
whereIsFoodCourt: where can i find a foodcourt
whereIsFoodCourt: how to get to foodcourt
whereIsFoodCourt: where the food court is
whereIsFoodCourt: where is a food court
whereIsFoodCourt: foodcourt where

Solving a problem with lot of synonyms for a single meaning

If there is a word with many synonyms spread all around the bot's training data, it's good to identify the word soon and replace it with an entity.

bad

# (intent: utterance)
whatIsCorona: what is a corona virus
whatIsCorona: what is a covid-19
whatIsCorona: more info about covid-19
treatment: how the virus is treated
treatment: how the covid19 is treated
treatment: treatment of corona
  • by intoducing an entity called @virus, we can simplify all training data.
  • the entity does not need to be included in interaction conditions

good

# (@entity: value [synonyms, ...])
@virus: covid19 [corona virus, covid-19, corona, the virus, virus, covid virus]

# (intent: utterance)
whatIsCorona: what is a @virus
whatIsCorona: more info about @virus
treatment: how the @virus is treated
treatment: treatment of @virus

Solving a problem with similar intents - Merging them

If there are many similar intents and

  • the only difference is a meaning of single word
  • the answer differs depending on that word

bad

# (intent: utterance)
whatIsScrewdriver: what is a screwdriver
whatIsScrewdriver: tell me about screwdrivers
whatIsHammer: what is a hammer
whatIsHammer: tell me about hammers
whatIsWarranty: what is a warranty
whatIsWarranty: tell me about warranty

The solution is to introduce an entity or more entities

good

# (@entity: value [synonyms, ...])
@product: screwdriver [screwdrivers, turnscrew, worm-screw]
          hammer [hammers]
@service: warranty [warranties, return policy]

# (intent: utterance)
whatIs: what is a @product
whatIs: what is the @product
whatIs: tell me about @product

And then you have to use entity condition at each interaction.

condition

Adding a new utterance or intent

Let's have a new utterance say hello and the bot now does not understand it.

  1. find an answer interaction, and check if there is already an intent

    condition

    If the right interaction does not exist, you'll have to create it.

  2. if the intent exists, ensure your utterance is similar to other utterances

    If the intent has a different meaning, but the answer is same - add a new intent - the answer can be distinguished later.

    matches

  3. use the NLP trainer to look for possible conflicting intents

    There already can be an intent you're looking for. Scroll down on discover all matched intents.

    trainer

  4. if there is a intent recognized, check the connected interaction is the answer for your question

    matching interactions

results matching ""

    No results matching ""