mohammed933/ConvRevise

The benchmark of MemRevise: A Revisable Memory for Conversational Search

TL;DR. In conversational search, users correct misunderstandings, revise their requests and return to earlier topics, so part of the history no longer reflects what they want. ConvRevise isolates such turns: each question is the last turn of a conversation, and a system must retrieve the passages that answer it for the topic the user means now, not for the topics they have left.

Overview

An illustrative conversation from the paper. The system answers turn 3 for the wrong astronaut; in turn 4 the user corrects it after first misspeaking, and turns 5–7 are a follow-up, a return and a value revision. To search well, a system must resolve every turn to the topic the user means (bold) and set aside the topics the user rejected (×).

Highlights

• 4,429 test questions in six tiers of returns and revisions, with 1,790 training and 1,783 development questions.
• Three source datasets: TopiOCQA, CORAL and OR-QuAC, each with its passage collection (670,821 passages in all).
• Gold and stale passages: every question lists the passages that answer it and the passages of the context the user has left.

Tiers

Tier	The question	Example	Test
Return-1	resumes the topic discussed just before the current one	when did the talk show begin? (The Panel, after six turns about John Bishop)	1,233
Return-2+	returns to a topic at least two topics earlier	Sorry for the detour — how did the one I asked about first evolve? (the example below)	815
Correction	rejects one or more topics and names the intended one	no, the one arthur b. rubinstein composed the music for. not the animated film about simba (WarGames, after an answer about The Lion King)	1,239
Follow-up	is a short question that depends on a preceding correction	can you name more muppet characters? (two turns after the user returns to Sesame Street)	300
Self-repair	corrects and re-corrects within the same turn	no, i meant the whatever it takes guy… sorry, no, the actor who named physical comedy as his big inspiration (Bryan Cranston)	608
Value revision	changes a value in the topic's name	hmm, i meant the one in 1863 (from War of 1812 to War of 1863)	234

The train and dev splits also contain natural CORAL questions (Natural) and, in CORAL's train questions, questions that stay on the current topic (Stay).

Example from the test set

The user asks about the red fox, digresses to polo and then to ancient Macedonia, and finally returns to the red fox without naming it:

Turn	User	Answer
1	how long does a red fox live in the wild	In the wild they typically do not survive past 5 years of age.
2	what is red fox?	It is the largest of the true foxes and one of the most widely distributed members of the order Carnivora.
3	where did it originate?	The species is Eurasian in origin.
4	How did polo likely begin, and where did its current form originate?	It most likely began as a simple game played by mounted Iranian and Turkic nomads in Central Asia, with the current …
5	What are some modern variants of polo?	Forms of arena polo include beach polo, played in many countries between teams of three riders on a sand surface, and …
6	Switching topics a bit, did ancient Macedonia contribute any notable inventions or early technologies?	Although it was perhaps not as prolific as other areas of Greece in technological innovations, some inventions may have …
7	Were there any early glass discoveries linked to Macedonia?	Yes. The first known clear, translucent glass pieces of the Greek world have been discovered in Macedonia and Rhodes, …
8	Sorry for the detour — how did the one I asked about first evolve?	(the system must retrieve the passage)

The last turn is the question of the record, and its tier is Return-2+: "the one I asked about first" is two topics back. The gold passage answers it for the red fox:

Red fox — Evolution. The red fox is considered a more specialised form of "Vulpes" than the Afghan, corsac and Bengal foxes in the direction of size and adaptation to carnivory; the skull displays far fewer neotenous traits than in other species, and its facial area is more developed. It is, however, not as adapted for a purely carnivorous diet as the Tibetan fox.

The stale passages belong to the topics the user has left. A system that takes "the one I asked about first" to be polo, where the detour began, retrieves passages such as:

Polo — Origins. Although the exact origins of the game are unknown, it most likely began as a simple game played by mounted Iranian and Turkic nomads in Central Asia, with the current form originating in Iran (Persia) and spreading east and west. In time polo became a …

The same record in the file (answers shortened here):

{
  "id": "v32t_c801525_t8",
  "eval_set": "test_topiocqa",
  "source": "topiocqa",
  "tier": "Return-2+",
  "conversation_id": "topiocqa_dev:99",
  "history": [
    {"turn": 1, "question": "how long does a red fox live in the wild", "answer": "In the wild they typically do not survive past 5 years of …"},
    {"turn": 2, "question": "what is red fox?", "answer": "It is the largest of the true foxes and one of the most …"},
    {"turn": 3, "question": "where did it originate?", "answer": "The species is Eurasian in origin."},
    {"turn": 4, "question": "How did polo likely begin, and where did its current form originate?", "answer": "It most likely began as a simple game played by mounted …"},
    {"turn": 5, "question": "What are some modern variants of polo?", "answer": "Forms of arena polo include beach polo, played in many …"},
    {"turn": 6, "question": "Switching topics a bit, did ancient Macedonia contribute any notable inventions or early technologies?", "answer": "Although it was perhaps not as prolific as other areas of …"},
    {"turn": 7, "question": "Were there any early glass discoveries linked to Macedonia?", "answer": "Yes. The first known clear, translucent glass pieces of the …"}
  ],
  "question": "Sorry for the detour — how did the one I asked about first evolve?",
  "gold_passages": ["wiki:8691032"],
  "stale_passages": ["wiki:11546321", "wiki:25639413", "wiki:25639431", "wiki:25639436"]
}

Dataset structure

Every record is one question: history holds the earlier turns of its conversation with their answers, and tier, gold_passages and stale_passages belong to question, the last turn. The passages are looked up by doc_id in the corpus split named by source. Each training question also comes with four mined negative passages, given in data/competitor_training/train.jsonl on GitHub.

Config	Split	Rows	Content
benchmark	train	1,790	train_topiocqa 839, train_coral 951
	dev	1,783	dev_topiocqa 431, dev_coral_natural 300, dev_coral_corrections_returns 450, dev_coral_returns 300, dev_orquac_corrections_returns 218, dev_orquac_value_revisions 84
	test	4,429	test_topiocqa 3,500, test_coral 617, test_orquac 312
corpus	topiocqa	434,179	passages with page title and section
	coral	162,650	passages without titles
	orquac	73,992	passages with page title

The names in the content column are the values of eval_set; the paper reports results per test set. Train and test share no conversation.

Field	Description
id	question id
eval_set	the evaluation set (see above)
source	the corpus split to search: topiocqa, coral or orquac
tier	what the question does (see below)
conversation_id	questions with the same id share their history
history	the earlier turns: turn, question, answer
question	the question to answer, the last turn of the conversation
gold_passages	doc_ids of the passages that answer the question
stale_passages	doc_ids of passages about topics the user has left
doc_id, title, section, text	a passage in corpus

Usage

from datasets import load_dataset

test = load_dataset("mohammed933/ConvRevise", "benchmark", split="test")
coral = test.filter(lambda q: q["eval_set"] == "test_coral")
passages = load_dataset("mohammed933/ConvRevise", "corpus", split="coral")

The paper reports MRR@10 per test set and the mean over the three. The TopiOCQA and OR-QuAC collections are subsets built around the benchmark's articles, so scores are not comparable with results on the full collections. Code to run MemRevise and score a system: GitHub.

Licence and citation

CC BY-NC-SA 4.0, following TopiOCQA (CC BY-NC-SA 4.0); CORAL is CC BY-SA 4.0. Please also cite the source datasets.

The BibTeX entry will be added when the paper is published.

BibTeX of the source datasets

@article{topiocqa,
  title     = {{TopiOCQA}: Open-domain Conversational Question Answering with Topic Switching},
  pages     = {468--483},
  author    = {Adlakha, Vaibhav and Dhuliawala, Shehzaad and Suleman, Kaheer and de Vries, Harm and Reddy, Siva},
  journal   = {Transactions of the Association for Computational Linguistics},
  volume    = {10},
  year      = {2022}
}

@inproceedings{coral,
  title     = {{CORAL}: Benchmarking Multi-turn Conversational Retrieval-Augmented Generation},
  author    = {Cheng, Yiruo and Mao, Kelong and Zhao, Ziliang and Dong, Guanting and Qian, Hongjin and Wu, Yongkang and Sakai, Tetsuya and Wen, Ji-Rong and Dou, Zhicheng},
  booktitle = {Findings of the Association for Computational Linguistics: NAACL 2025},
  pages     = {1308--1330},
  year      = {2025}
}

@inproceedings{orquac,
  title     = {Open-Retrieval Conversational Question Answering},
  author    = {Qu, Chen and Yang, Liu and Chen, Cen and Qiu, Minghui and Croft, W. Bruce and Iyyer, Mohit},
  booktitle = {Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval},
  pages     = {539--548},
  year      = {2020}
}