informagus/olddog-indexes

Inverted Indexes for Information Retrieval Experiments

This dataset contains inverted indexes for four IR test collections: TREC CAsT v1, ClueWeb09-B, ClueWeb12-B, and GOV2.

Representation

The indexes follow the "OldDog" schema by Mühleisen et al., by representing the index with three tables (dict, docs and postings). The easiest way to use the data is to load it in a relational database (like DuckDB) and perform retrieval through SQL. See the "Quickstart" section below.

The indexes are stemmed using the Porter stemmer, and stopwords have not been removed.

Each table is stored using one or more Parquet files, with the following schemas:

dict

Column	Type	Description
term	string	A stemmed term that appears in the collection
termid	int	A numeric identifier for each term, used in the postings table
df	int	The term's document frequency, i.e. the number of documents that contain this term
cf	int	The term's collection frequency, i.e. the total number of occurrences of this term

docs

Column	Type	Description
docid	int	A numeric identifier for each document, used in the postings table
collection_docid	string	The canonical identifier for each document (e.g. its TREC ID)
length	int	The document's length

postings

Column	Type	Description
termid	int	A term from the dictionary
docid	int	A document in which the term occurs
tf	int	The term frequency, i.e. the number of times the term appears in the document

Quickstart

The easiest way to run IR experiments with this collection is to load the data into DuckDB and perform retrieval through SQL.

Loading the index

First, download the index data for a collection:

CREATE TABLE dict AS FROM 'hf://datasets/informagus/olddog-indexes/{collection}/dict.parquet';
CREATE TABLE docs AS SELECT * RENAME (collection_docid AS name) FROM 'hf://datasets/informagus/olddog-indexes/{collection}/docs.parquet';
CREATE TABLE postings AS FROM 'hf://datasets/informagus/olddog-indexes/{collection}/postings/*.parquet';

Then, compute the statistics for the stats table:

CREATE TABLE stats AS SELECT count(*) AS num_docs, avg(length) AS avgdl FROM docs;

Create the stopwords table with the stopword list you want to use. For instance, to use the stopword list also used by DuckDB's FTS extension, issue:

CREATE TABLE stopwords AS FROM read_csv('https://raw.githubusercontent.com/igorbrigadir/stopwords/refs/heads/master/en/smart.txt', header=0) t(sw);

The stopword list can also be empty:

CREATE TABLE stopwords(sw VARCHAR);

Retrieval

We use some SQL macros to simplify retrieval:

CREATE MACRO tokenize(query) AS (
    string_split_regex(regexp_replace(lower(strip_accents(query::STRING)), '[0-9!@#$%^&*()_+={{}}\[\]:;<>,.?~\\/\|''"`-]+', ' ', 'g'), '\s+')
);

CREATE MACRO get_termids(query) AS TABLE (
    WITH tokens AS (
        SELECT unnest(tokenize(query)) AS token
    ),
    stemmed AS (
        SELECT stem(token, 'porter') AS term
        FROM tokens
        WHERE token NOT IN (FROM stopwords)
    ),
    qtfs AS (
        SELECT term, COUNT(*) AS qtf
        FROM stemmed
        GROUP BY term
    )
    SELECT termid, qtf
    FROM qtfs
    JOIN dict USING (term)
);

CREATE MACRO score_bm25(query, k1 := 0.9, b := 0.4) AS TABLE (
    SELECT
        docid,
        first(name) AS name,
        sum(
            qtf * 
            log(
                ((SELECT num_docs FROM stats) - df + 0.5) / (df + 0.5)
            ) * (tf * (k1 + 1)) / (
                tf + k1 * (1 - b + b * length / (SELECT avgdl FROM stats))
            )
        ) AS score
    FROM postings
    JOIN dict USING (termid)
    JOIN docs USING (docid)
    JOIN get_termids(query) USING (termid)
    GROUP BY docid
);

CREATE MACRO search_bm25(query, k := 1000, k1 := 0.9, b := 0.4) AS TABLE (
    SELECT *, row_number() OVER (ORDER BY score DESC) AS rank
    FROM score_bm25(query, k1 := k1, b := b)
    ORDER BY score DESC
    LIMIT k
);

A BM25 query can then simply be issued by using:

FROM search_bm25('this is a test query');

Selective Search

The indexes shared in this dataset have been used for experiments with selective search architectures. For more information on how run such experiments, refer to our code repositories on selective search for first-stage retrieval (where we use the ClueWeb and GOV2 indexes) and for session search (where we use the CAsT collection).

References

If you use our indexes for the ClueWeb or GOV2 collections, please cite the following paper:

@inproceedings{hendriksen_selective_2025,
  title = {Selective Search as a First-Stage Retriever},
  booktitle = {Experimental IR Meets Multilinguality, Multimodality, and Interaction: 16th International Conference of the {CLEF} Association, {CLEF} 2025, Madrid, Spain, September 9–12, 2025, Proceedings},
  author = {Gijs Hendriksen and Djoerd Hiemstra and Arjen P. de Vries},
  pages = {17--33},
  year = 2025,
  doi = {10.1007/978-3-032-04354-2_2}
}

If you use the TREC CAsT index, use the following reference:

@inproceedings{hendriksen_efficient_2025,
  title = {Efficient Session Retrieval Using Topical Index Shards},
  booktitle = {Advances in Information Retrieval: 47th European Conference on Information Retrieval, {ECIR} 2025, Lucca, Italy, April 6–10, 2025, Proceedings, Part {II}},
  author = {Gijs Hendriksen and Djoerd Hiemstra and Arjen P. de Vries},
  pages = {38--54},
  year = 2025,
  doi = {10.1007/978-3-031-88711-6_3}
}