[Dataset] MultiDoGO Dataset

[SaebyeolShin]

Multi-Domain Goal-Oriented Dialogues (MultiDoGO)

This corpus converts the Multi-Domain Goal-Oriented Dialogues (MultiDoGO) dataset into ConvoKit format across six domains: airline, fastfood, finance, insurance, media, and software. It provides three corpora: (1) unannotated dialogs, (2) sentence-level annotated splits, and (3) turn-level annotated splits.

Overview

This corpus adapts the MultiDoGO dataset into ConvoKit format. It includes:

• Unannotated human-to-human dialogs
• Annotated paper splits at sentence level and turn level (train/dev/test)

Dataset Details

Utterance

• text: original utterance text
• meta:
  • domain: one of {airline, fastfood, finance, insurance, media, software}
  • split: {train, dev, test} (empty for unannotated)
  • turnNumber: original turn index (string)
  • sentenceNumber: sentence index for sentence-level splits (string; empty otherwise)
  • raw_utteranceId: original utteranceId
  • slot_labels: space-separated labels
  • intent: single or multiple intents (if multiple, separated by <div>)

Speaker

• id: `domain__role`
• meta.role: original authorRole (fallback customer)
• meta.domain: domain

Conversation

• id: `domain__conversationId`
• meta:
  • raw_conversationId
  • domain
  • split
  • (optional) num_utterances

Conversion Choices

• ID scheme: Utterance.id = domain__conversationId__turnNumber__sentenceNumber__utteranceId
• Reply chain: reply_to links to the previous utterance in the same conversation.
• Speaker mapping: use authorRole (fallback customer), namespaced by domain.
• Label handling: intent uses <div> for multi-intent; slot_labels split by space.
• Robust parsing: tolerant TSV reading via pandas.read_csv(..., engine="python", on_bad_lines="skip"); coerce non-numeric turnNumber/sentenceNumber by regex for ordering.
• Conversation objects: created implicitly by Corpus(utterances=...), then conversation metadata (conv.meta) updated.

Statistics

Unannotated

• Utterances: 1,376,153
• Conversations: 86,716
• Speakers: 12
• Domains (by conversations):

  domain	convs
  media	33,322
  airline	15,098
  insurance	14,259
  fastfood	9,642
  finance	8,833
  software	5,562

• Splits: (none — unannotated)
• Top intents / slots: (not applicable in unannotated)

Sentence-level splits (annotated)

• Utterances: 121,636
• Conversations: 14,215
• Speakers: 6
• Domains (by conversations):

  domain	convs
  media	2,432
  airline	2,430
  finance	2,379
  software	2,355
  insurance	2,332
  fastfood	2,287

• Splits (by conversations): train 9,948 • test 2,848 • dev 1,419
• Top intents (k=10): contentonly 39,948; confirmation 18,009; openinggreeting 13,007; rejection 10,761; thankyou 10,403; outofdomain 6,448; closinggreeting 3,765; startserviceintent 2,224; orderpizzaintent 2,009; orderdrinkintent 1,916
• Top slot labels (k=10): O 318,446; name 12,373; address 12,284; booking_confirmation_number 6,883; food_item 6,596; ingredient 4,918; datacategoryvalues 4,563; ssn 4,010; quantity 3,670; email_address 3,473

Turn-level splits (annotated)

• Utterances: 116,597
• Conversations: 14,147
• Speakers: 6
• Domains (by conversations):

  domain	convs
  media	2,420
  finance	2,406
  airline	2,404
  software	2,362
  insurance	2,321
  fastfood	2,234

• Splits (by conversations): train 9,900 • test 2,834 • dev 1,413
• Top intents (k=10): contentonly 39,248; confirmation 16,666; openinggreeting 12,618; rejection 10,903; thankyou 9,712; outofdomain 4,810; closinggreeting 3,111; startserviceintent 2,450; orderdrinkintent 1,876; orderpizzaintent 1,780
• Top slot labels (k=10): O 317,541; name 12,683; address 11,901; food_item 7,413; booking_confirmation_number 6,853; datacategoryvalues 4,474; quantity 4,025; ssn 3,862; email_address 3,466; card_number 3,390

Aggregated (deduplicated across corpora)

• Utterances (unique): 1,614,386
• Conversations (unique): 101,111
• Speakers (unique): 18
• Domains (by conversations, normalized):

  domain	convs
  media	35,771 (= 33,322 unannotated + 2,449 annotated)
  airline	17,535 (= 15,098 + 2,437)
  insurance	16,606 (= 14,259 + 2,347)
  fastfood	11,986 (= 9,642 + 2,344)
  finance	11,271 (= 8,833 + 2,438)
  software	7,942 (= 5,562 + 2,380)

• Top intents (aggregate, k=10): contentonly 79,196; confirmation 34,675; openinggreeting 25,625; rejection 21,664; thankyou 20,115; outofdomain 11,258; closinggreeting 6,876; startserviceintent 4,674; orderdrinkintent 3,792; orderpizzaintent 3,789
• Top slot labels (aggregate, k=10): O 635,987; name 25,056; address 24,185; food_item 14,009; booking_confirmation_number 13,736; datacategoryvalues 9,037; ingredient 8,186; ssn 7,872; quantity 7,695; email_address 6,939

Contact

The original dataset was distributed in the EMNLP-2019 paper: Multi-Domain Goal-Oriented Dialogues (MultiDoGO): Strategies toward Curating and Annotating Large Scale Dialogue Data

The dataset was formatted for Convokit by Saebyeol Shin (ss4333@cornell.edu)

Dataset Access

Dataset with example script: https://drive.google.com/drive/folders/1BnFMFiGkVA1bUGY-IDzkvOXpZ-zYOVPn?usp=sharing