Handle sentences with multiple intents

[hdh112]

• Right now, each sentence has a single label that represents single intent.
  • Ex: Label 0 for fruit_order
• Yet, as sentences get long, a single sentence may contain multiple intents
  • Ex: fruit_order, fruit_type, fruit_quantity

Needs to consider labelling method

[hdh112]

Labelling method

• Multi-label classification
• Update labels as the conversation continues
• When all required intents are met, order & end conversation

Required intents to complete an order

• (Group 0) fruit_order
• (Group 1) fruit_type
• (Group 2) fruit_quantity

Whether or not to use the following:

(Reference: Diagram / URL )

• Define a chunk grammar, or chunking rules, then chunk
  • Ex: NP(Noun phrase), VP(Verb phrase), AP(Adjective phrase)
• Construct graph
  • Vertex: Nouns
  • Edge: Verbs
• Binary relevance between labels
  • Implementing this may help faster classification

Helpful resources

• Semantic role labelling (He et al., 2017)
• Semantic role labelling, multiple layers (Zhou and Xu, 2015)

[hdh112]

Intent groups

Group 0 : Basic intent

• 0 undefined
• 1 fruit_order; fruit order service

• 2 fruit_advice; fruit recommendation service

  Used for fruit_order

  Group 1 : fruit_type

• 0 undefined
• 1 Blueberries
• 2 Cherry tomatoes
• 3 Cherries
• 4 Grapes
• 5 Strawberries
• 6 Apples
• 7 Peaches
• 8 Grapefruits
• 9 Fruit cups
• 10 Kiwis

Group 2 : fruit_quantity

• 0 undefined
• 1 specific
• 2 few
• 3 enough

• 4 many

  Used for fruit_advice

  Group 3: fruit_price

  Used to decide Group 2(fruit_quantity) May need to utilize currency method

• 0 undefined
• 1 cheap
• 2 average
• 3 sale

Group 4: fruit_taste

Used to decide Group 1(fruit_type)

• 0 undefined
• 1 simple (5, 7)
• 2 rich (0, 2, 4, 6)

Group 5: fruit_easy

Used to decide Group 1(fruit_type)

• 0 undefined
• 1 clean (0, 1, 2, 3, 4, 8)
• 2 handy (9)
• 3 fast (0, 1, 8)

Group 6: fruit_size

Used to decide Group 1(fruit_type)

• 0 undefined
• 1 small (0, 1, 2, 3, 4)
• 2 medium (5, 6, 7, 8)

Group 7: user_status

Used for empathy, a more friendly conversation scenario, and understand user's needs TODO: may have multiple user status

• 0 undefined
• 1 happy
• 2 sad
• 3 unsure
• 4 tired
• 5 recommend; Used to decide Group 0(fruit_advice)
• 6 sick
• 7 order; Used to decide Group 0(fruit_order)

[hdh112]

TODO

• ~Predict sample sentence in the current model(only fruit_order)~
• ~Change data structure of label into a vector~
• Decrease loss by adjusting layers of model
• Train fruit_advice dataset & predict sample(fruit_order or fruit_advice)
• Save current conversation status of the current lifespan, by adding up label vectors
  • Ex) [1 0 0 0 0 0 0 0] && [0 1 0 0 0 0 0 0]: fruit_order and fruit_type == "Blueberries"
  • Ex) [1 1 0 0 0 0 0 0] && [0 5 2 0 0 0 0 0] (when user changed its mind about fruit_type)
    • fruit_order
    • For fruit_type, ask user "Which fruit would you like to order: Blueberries or strawberries?"
    • fruit_quantity == "few"