# Report

## I. Module in convlab v2:

### 1. DST:

* DST is a dialogue-level task that map partial dialogues into dialogues states.
* Approach:
* Input: a dialogue/a turn with its previous state
* Output: dialogue state (i.e slot-value pair)

#### 1.1. TRADE:

* TRADE model takes \{*dialogue, domain_i, slot_j*\} as input and generates the value corresponding to \{*domain_i, slot_j*\} and the slot gates in {*ptr*, *dontcare*, *none*}.

##### a. Ontology:

* Ontology domain-slot has predefined, such as:

- hotel-price range
  + cheap
  + do n't care
   ...
- hotel-type
  + hotel
  + guest house
   ...
- hotel-parking
  + yes
  + do n't care
   ...
- hotel-book stay
  + 2
  + 1
   ...
- hotel-book day
  + tuesday
  + monday
   ...
- hotel-book people
  + 6
  + six
   ...
- hotel-area
  + east
  + north
   ...
- hotel-stars
  + 4
  + 2
   ...
- hotel-internet
  + yes
  + do n't care
   ...
- train-destination
  + bishops stortford
  + cambridge
   ...
- train-day
  + friday
  + we
   ...
- train-departure
  + cambridge
  + birmingham new street
   ...
- train-arrive by
  + 23:27
  + 19:51
   ...
- train-book people
  + 1
  + 15
   ...
- train-leave at
  + 21:15
  + 15:01
   ...
- attraction-area
  + centre
  + west
   ...
- restaurant-food
  + turkish
  + chinese
   ...
- restaurant-price range
  + expensive
  + moderate
   ...
- restaurant-area
  + centre
  + north
   ...
- attraction-name
  + all saints church
  + funky fun house
   ...
- restaurant-name
  + meze bar restaurant
  + pizza hut city centre
   ...
- attraction-type
  + museum
  + college
   ...
- hospital-department
  + paediatric clinic
  + transitional care
   ...
- hotel-name
  + city centre north b and b
  + el shaddai
   ...
- taxi-leave at
  + 18:45
  + after 15:45
   ...
- taxi-destination
  + express by holiday inn cambridge
  + dojo noodle bar
   ...
- taxi-departure
  + christ's college
  + kirkwood house
   ...
- restaurant-book time
  + not given
  + 13:10
   ...
- restaurant-book day
  + tuesday
  + not given
   ...
- restaurant-book people
  + 2
  + 3
   ...
- taxi-arrive by
  + 19:15
  + 15:45
   ...
- bus-departure
  + cambridge
   ...
- bus-destination
  + london kings cross
  + bishops stortford
   ...
- bus-leaveAt
  + 21:45
   ...
- bus-day
  + wednesday
   ...

- Example:
- Prev_state:
- Usr_action:

    "I need to book a hotel in the east that has 4 stars ."

- NLU output:

    ['hotel-area-east', 'hotel-stars-4']

- Updated state:

- Update state algorithm:
- Replace value from NLU output to current state.

##### b. Non-Ontology
* The zero-shot DST model should be able to generate the correct value of $Y_{target}^{value}$ given dialogue history $X$, domain $D_{target}$ and slot $S_{target}$ where the pair ( $D_{target}$, $S_{target}$) unseen in training pharse.

#### 1.2. Sumbt:
* SUMBT considers a domain-slot type (e.g., ‘restaurant-food’) as a question andfinds the corresponding slot-value in a pair of userand system utterances, assuming the desirable answer exists in the utterance. Given domain and slot-types (e.g., ‘restaurant’ domain and ‘food’ slot-type), it stimates the probability of corresponding slot-value candidates (e.g., ‘Korean’ and European’) that are predefined in a domain ontology.


#### 1.3. Comer
* Using the hierarchical decoder. The hierarchical decoder, a hierarchical stacked decoder, decodes domains, slots, and values in a hierarchical manner, and generates the current round of dialogue state itself as a target sequence.
* Taking the encoder-decoder network as the skeleton, but using hierarchically stacked decoders, first generate slot sequences in the dialogue state , and then generate corresponding value sequences for each slot.
* Representation state: (Domain1- Slot1, Value1; Slot2, Value2;… Domain2- Slot1, Value1; …).
* Single value assumption:

### 2. Policy

* Dialogue policy decides the system action for interacting with users based on dialogue states.
* Approach:
* Input: dialogue state + KB results (database)
* Output: system action (speech-act + slot-value pairs)

#### 2.1. Rule

* From current state, take items randomly from database to recommend. For example:
* Usr action:

    {'Hotel-Inform': [['Area', 'east'], ['Stars', '4']]}

* Constrain:

    [['area', 'east'], ['stars', '4']]

* From database:

    first_item = '''- address : 124 tenison road
     - area : east
     - internet : yes
     - parking : no
     - id : 0
     - location : [52.1963733, 0.1987426]
     - name : a and b guest house
     - phone : [phone]
     - postcode : cb12dp
     - price : {'double': '70', 'family': '90', 'single': '50'}
     - pricerange : moderate
     - stars : 4
     - takesbookings : yes
     - type : guesthouse
     - Ref : 00000000'''
    second_item = ...

* Recommend:

    ['Recommend', 'Hotel', 'Name', 'a and b guest house']

## II. ParlAI

* ParlAI (pronounced “par-lay”) is a python framework for sharing, training and testing dialogue models, from open-domain chitchat, to task-oriented dialogue, to visual question answering.
* 100+ popular datasets available all in one place, with the same API

### 1. Feature:
#### 1.1. Test public model

#### 1.2. Obtain data:
* Follow [this colab tutorial](https://colab.research.google.com/drive/1bRMvN0lGXaTF5fuTidgvlAl-Lb41F7AD#scrollTo=AqckSXqlmWuT), we can use both CMD or python script to download data.

#### 1.3. Train model
* Ex: train *seq2seq model* with *empathetic_dialogues dataset* and save model in `/from_scratch_model`:


#### 1.4. Bringing our dataset

#### 1.5. Create our agent model

* Read more in [this colab tutorial](https://colab.research.google.com/drive/1bRMvN0lGXaTF5fuTidgvlAl-Lb41F7AD#scrollTo=AqckSXqlmWuT)

#### 1.6. Test model