Overview

CHARP is diagnostic testbed, exclusively assess whether information-seeking dialogue systems effectively attend to and use the conversation history. CHARP is built by modifying examples from the FaithDial validation set to ensure maximum domain alignment with FaithDial and to minimize annotation costs. That is, we edit FaithDial examples to make their response dependent on the conversation history analogously to FaithDial's editing of WoW annotations to make them hallucination-free. CHARP consists of 2 subsets, where only the last seeker utterance differs: a self-contained easy version (eCHARP), and a hard (hCHARP) which requires reasoning over the conversation history and the provided knowledge that corresponds to the last seeker.

Data Splits

We create two variants of CHARP: hCHARP for examples where addressing the last seeker's inquiry requires reasoning over the conversation history, and eCHARP, where the last inquiry can be addressed without such reasoning. We annotate 42% of the FaithDial validation set (after excluding examples without conversation history) . CHARP consists of containing 2,160 examples, split equally between eCHARP and hCHARP:

eCHARP: 1080 samples
hCHARP: 1080 samples

Data Fields

Both eCHARP and hCHARP have the same data format:

row_idx: int. Index of the sample that is equivalent to the one in FaithDial validation (row enumeration).
history: List[string]. The dialogue history.
knowledge: string. The source knowkedge on which the bot should ground its response.
response: string. The expected model response

Data Instance

An example of eCHARP looks as follows:

{
  "row_idx": "1293",
  "history": [
    "I love watching and playing basketball.",
    "I see. Have you ever tried to describe basketball? I would say it is a low contact sport where the game is held in a rectangular court.",
    "Yeah I never though of that, can you repeat what you told me again so I can take notes?",
    "Yes I can, basketball is a sport with limited contact. It is held on a rectangular like court.",
    "What would you describe the sport is played like?",
    "The objective for basketball is shooting the ball into the hoops. The hoops are high and placed with a backboard on each side of the court.",
    "Oh yea, that's pretty simple. Do you know any famous basketball courts?"
  ],
  "knowledge": "Supreme Court in the USA is very famous to have well-known judges, while the Philippine Arena is popular due to the size of the basketball court.",
  "response": "Ah yeah, I heard that the Philippine Arena is popular because of the size of the basketball court."
}

An example of hCHARP looks as follows:

{
  "row_idx": "1293",
  "history": [
    "I love watching and playing basketball.",
    "I see. Have you ever tried to describe basketball? I would say it is a low contact sport where the game is held in a rectangular court.",
    "Yeah I never though of that, can you repeat what you told me again so I can take notes?",
    "Yes I can, basketball is a sport with limited contact. It is held on a rectangular like court.",
    "What would you describe the sport is played like?",
    "The objective for basketball is shooting the ball into the hoops. The hoops are high and placed with a backboard on each side of the court.",
    "Oh yea, that's pretty simple. Do you know any famous courts?"
  ],
  "knowledge": "Supreme Court in the USA is very famous to have well-known judges, while the Philippine Arena is popular due to the size of the basketball court.",
  "response": "Ah yeah, I heard that the Philippine Arena is popular because of the size of the basketball court."
}

Who are the annotators?

We would like to thank Imad Mousaoui, Ella Cho, Abdulmuizz Yusuf, and Parminder Singh Bharot, the professional annotators without whom this work would have not been possible.

Licensing Information

MIT

Citation

@article{ghaddar2024charp,
      title={CHARP: Conversation History AwaReness Probing for Knowledge-grounded Dialogue Systems}, 
      author={Abbas Ghaddar and David Alfonso-Hermelo and Philippe Langlais and Mehdi Rezagholizadeh and Boxing Chen and Prasanna Parthasarathi},
      year={2024},
      eprint={2405.15110},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}