DialogZoo

Data construction

To replicate data construction, three steps are required:

• Download data: bash scripts/download.sh
• Convert origin data into our unified format: bash scripts/convert_to_unified.sh

{
    # Optional values: `single` or `multi`. Indicates whether it is a single-turn or multi-turn dialogue.
    "turn": str,

    # The domains involved in the dialogue (a list because some dialogues involve multiple domains).
    "domain": [],

    # The language of the dialogue, based on the original dataset annotations (e.g., en, fr, etc.).
    "locale": str,

    # The dialogue, represented as a list where each element is a dictionary for a single turn.
    "dialog": [
        {
            # The roles involved in each turn. Some datasets may have multiple roles per turn, so it's a list.
            # For datasets without role annotations:
            #   * Use `ROLE` for single-turn data.
            #   * Use `ROLE1`, `ROLE2`, etc., for multi-turn data.
            "roles": [str, ...],

            # The text of the current turn.
            "utterance": str,
            
            # Used for the "answer" in QA tasks.
            "start": int,
            "end": int,
            "dialog_turn": int

            # Rewritten text corresponding to the current turn.
            "rewritten": str,

            # Dialogue state, represented as a list where each element includes:
            #   Domain: Some datasets constrain slot-value pairs within specific domains.
            #   Intent: Some datasets constrain slot-value pairs within specific intents.
            #   Slot-value pairs: A list where each element includes a slot and its corresponding values.
            #     Slot name: A string.
            #     Values: A list where a slot may have multiple values.
            #       Each value includes four parts: the value itself, the normalized value, 
            #       the character index in the current turn's text, and more.
            #     Relation: Some slots are equal to a value, while others are greater than a value. 
            #       Defaults to "equal" if not specified.
            #   Requested slots: A list of slots that need to be queried but are not filled in the current state.
            "belief_state": [
                {
                    # Intent
                    "intent": str,
                    # Slot-value pairs
                    "informed_slot_value_table": [
                        {
                            # Slot name
                            "slot": str,
                            # Values
                            "values": [{
                                # Actual value
                                "value": str,
                                # Normalized value
                                "cononical_value": str
                            }, ...],
                            # Slot-value relation
                            "relation": str,
                        },
                        ...
                    ],
                    # Requested slots
                    "requested_slots": [],
                    # Domain
                    "domain": str,
                }, ...
            ],

            # Dialogue actions, represented as a list where each element includes:
            #   Domain: Some datasets constrain slot-value pairs within specific domains.
            #   Action: The actions involved in the current turn.
            #   Slot-value pairs: Same as in dialogue state.
            "dialog_acts": [
                {
                    # Action
                    "act": str,
                    # Slot-value pairs
                    "slot_value_table": [
                        {
                            # Slot name
                            "slot": str,
                            # Slot-value relation
                            "relation": str,
                            # Values
                            "values": [
                                {
                                    # Actual value
                                    "value": str,
                                    # Normalized value
                                    "cononical_value": str,
                                    # Start position
                                    "start": int,
                                    # End position
                                    "end": int,
                                },...
                            ]
                        },
                        ...
                    ],
                    # Domain
                    "domain": str,
                },
                ...
            ],
            
            # Slot filling
            "slots_to_fill": {
                "intent": str,
                "slot_value_table": [
                    {
                        "slot": str,
                        "values": [
                            {
                                "value": str,
                                "start": int,
                                "end": int
                            }
                        ],
                        "relation": str, # '=', '<=', and so on
                    }
                ]
            },
            
            # Named entity recognition
            "named_entity_recognition": [
                {
                    "type": str,
                    "values": [
                        {
                            "value": str,
                            "start": int,
                            "end": int
                        }, ...
                    ]
                }, ...
            ],

            "characters": [
                {
                    "value": str,
                    "start": int,
                    "end": int
                }
            ]
            
            # Intent detection
            "active_intents": [str],

            # Query
            "query" {
                ...
            },
            
            # Query result
            "querying_result": {
               ...
            },

            # Recorded satisfied main items
            "main_items": [],

            # Aspect Sentiment Triplet Extraction task, represented as a list where each element includes three parts:
            #   Target entity.
            #   Related sentiment.
            #   Words reflecting the sentiment.
            "aspects": [
                {
                    # Target entity
                    "target": {
                        # Entity value
                        "value": str,
                        # Start position in the current turn's text
                        "start": int,
                        # End position in the current turn's text
                        "end": int
                    },
                    
                    # Category of the target entity
                    "category": str,
                    
                    # Words reflecting the sentiment
                    "opinion": {
                        # Sentiment word
                        "value": str,
                        # Start position in the current turn's text
                        "start": int,
                        # End position in the current turn's text
                        "end": int
                    },
                    # Related sentiment
                    "sentiment": str
                }
            ],

            "emotions": [
                {
                    "emotion": str,
                    "sentiment": "positive", "negative", or "ambiguous",
                    "evidences": [
                        {
                            "turn": int,
                            "span": str,
                            "start": int,
                            "end": int
                        }
                    ],
                    "evidence_types": [str]
                }
            ],

            "kg_label": str,
            
            # Knowledge that may be required for each turn, used to select knowledge.
            "knowledge_to_select": str,
            
            # SQL
            "sql": str,
            
            # Rewritten text
            "rewritten": str,
            
            "roles_to_select": [str],
        },

    ],

    # Summary derived from the entire dialogue.
    "summary": str,

    # Entity relations determined from the entire dialogue.
    "instance_relations": [
        {
            "instance1": str,
            "instance2": str,
            "relations": [
                {
                    "relation": str,
                    "trigger": str
                }, ...
            ]
        }, ...
    ]
    
    # Role relations determined from the entire dialogue.
    "role_relations": [
        {
            "turn": int,
            "relation": str
        }
    ],
    
    # Used in FriendsPersona to determine a character's persona based on the entire dialogue.
    "role_personas": [
        {
            "name": str,
            "personas": [
                {
                    "persona": str,
                    "sentiment": int
                }, ...
            ]
        }
    ],

    # External knowledge required for the dialogue.
    "knowledge": {
        # `text`, `persona`, `kg`, or `schema`.
        "type": str,

        # For `text`.
        "value": str,

        # For `persona`, persona of all roles, used for personachat.
        "value": [
            {
                # Role name, matching the dialogue turn.
                "role": str,

                # Persona description, which may include several sentences.
                "description": []
            },
            ...
        ]

        # For `kg`.
        "value": {
            # `directed` or `undirected`.
            "direction": str,

            # Graph.
            "graph": [
                {
                    # Source node.
                    "source": str,

                    # Target node.
                    "target": str,

                    # Relation.
                    "relation": str
                },
                ...
            ]
        }

        # For `schema`.
        "value": {
            ...
        }

        # For `dialogue`.
        "value": {
            "dialog": [],
            "relations": []
        }
        
        # For `wiki`.
        "value": {
            ...
        }
        
        # For `sql`.
        "value": [
            {
                "turn": int, 
                "sql": str,
                "result": ...
            }, ...
        ],
        
        # For dialogues based on specific article excerpts, this field indicates the article and section titles.
        "value": {
            "article title": str,
            "section title": str
        },
    }
}

• Linearize: bash scripts/convert_to_seq.sh

The processed data is located at DialogZoo.tar.

Data statistics

ID	MRC	ER	MCQA	QCR	RRR	CI	SF	DCRG	CC	ABSA	T2S	DST	DT	DS	SP	NLI	Total
34,963	368,490	135,356	196,620	33,192	5,037	36,385	104,100	390,463	262,876	17,328	30,220	298,358	60,563	27,192	31,279	169,654	2,202,076