There are currently few datasets appropriate for training and evaluating models for Conversational Information Seeking (CIS). The main aim of TREC CAsT is to advance research on conversational search systems. The goal of the track is to create a reusable benchmark for open-domain information centric conversational dialogues.
The track will run in 2021 and establish a concrete and standard collection of data with information needs to make systems directly comparable.
This is the third year of TREC CAsT, which will run as a track in TREC. This year we aim to focus on candidate information ranking in context:
Read the dialogue context: Track the evolution of the information need in the conversation, identifying salient information needed for the current turn in the conversation
Retrieve Candidate Response Information: Perform retrieval over a large collection of paragraphs (or knowledge base content) to identify relevant information
Year 3 (TREC 2021)
Important Dates
Test topic release: Available
Run submission deadline: August 18th, 2021
Data
Topics
NEWEvaluation topics for Year 3 V1.0 - 25 primary evaluation topics in JSON and Protocol Buffer format. There are two variants automatic and manual.
Corpora
Washington Post 2020 - Note: This is a new dump including January 2012 through December 2020.
MS MARCO (Documents) - MS MARCO document corpus used in DL 2019/2020. NOTE: This is NOT the recently release V2 corpus.
Baselines
Automatic and manual baselines are provided baseline runs. See README for details.
Interactive web UI - A web UI supporting BM25 document retrieval and T5 passage reranking. Rewriting is supported using query and result passage context using the T5-Canard model, that format is the “raw rewrite”. A manual that explains how to use the UI can be found here.
Duplicate handling
The provided tools support removing duplicate documents from collection, including both MARCO and WaPo duplicates.
Evaluation topics for Year 2 V1.0 - 25 primary evaluation topics in JSON and Protocol Buffer format. There are two variants automatic and manual.
Baselines
NEW - BM25 + BERT baseline - We provide a BM25 + BERT reranked baseline run for the raw utterances, automatically rewritten utterances, and the manually rewritten utterances.
NEW - Interactive web UI - A simple web UI with the BM25 + BERT model used to create the baseline runs. No rewriting is performed.
Collection
The corpus is a combination of two standard TREC collections: MARCO Ranking passages and Wikipedia (TREC CAR).
Training judgments - We provide limited (incomplete) training data for 5 topics (approximately 50 turns). These are judged from the baseline retrieval run (below). The judgments are graded on a three point scale (2 very relevant, 1 relevant, and 0 not relevant).
To facilitate work on passage ranking only we performed manual resolution of coreference as well as conversational ambiguity for topics. We make these available to participants who may not have access to automatic methods. Runs using this data manual runs. The annotations are provided in a tab separated format with the turn id (query id) and the rewritten query in text form.
EVALUTION: Complete annotations on the evaluation topics for the year 1 evaluation queries.
Baselines
Indri search interface - We provide an Indri index of the CAsT collection. See the help page for details on indexing parameters and statistics. It includes a standard batch search API limited to 50 queries per batch.)
Baseline retrieval - We provide the queries and run files in trec eval format: train queries, train run file, test queries, test run file - We provide an Indri baseline run with Query Likelihood run, including both the topics and run files. Queries are generated by running AllenNLP coreference resolution to perform rewriting and stopwords are removed using the Indri stopword list.
Collection
The corpus is a combination of three standard TREC collections: MARCO Ranking passages, Wikipedia (TREC CAR), and News (Washington Post)
The TREC Washington Post Corpus version 2: Note this is behind a password and requires an organizational agreement, to obtain it see: https://ir.nist.gov/wapo/
Document ID format
The document id format is [collection_id_paragraph_id] with collection id and paragraph id separated by an underscore.
The collection ids are in the set: {MARCO, CAR, WAPO}.
The paragraph ids are: standard provided by MARCO and CAR. For WAPO the paragraph ID is [article_id-paragraph_index] where the paragraph_index is the starting from 1-based index of the paragraph using the provided paragraph markup separated by a single dash.
Example WaPo combined document id: [WAPO_903cc1eab726b829294d1abdd755d5ab-1], or CAR: [CAR_6869dee46ab12f0f7060874f7fc7b1c57d53144a]
Duplicate handling
Early analysis found that both the MARCO and WaPo corpora both contain a significant number of near duplicate paragraphs. We have run near-dupliate detection to cluster results; only one result per duplicate cluster will be evaluated. It is suggested that you remove dupliates (keeping the canonical document) from your indices.
Note: The tools in the repository below require these files as input for processing the collection and perform deduplication when the data is generated.
Code and tools
TREC-CAsT Tools repository with code and scripts for processing data.
The tools contain scripts for parsing the collection into standard indexing formats. It also provides APIs for working with the topics (in text, json, and protocol buffer formats).
Note: This will evolve over time, it currently contains topic definition files and scripts for reading and loading topics.