The project contribution involves the creation and sharing of a configurable software workflow and a set of gold-standard datasets for training and evaluating supervised query refinement methods. We aim to empower researchers to construct gold-standard query refinement datasets more flexibly. Our software workflow, taking input datasets of queries with associated relevance judgments, an information retrieval method, and an evaluation metric, produces a gold standard dataset. This dataset includes refined queries for each input query, enhancing the performance of the specified retrieval method according to the chosen evaluation metric. Here we have generated and shared gold-standard datasets for different types of queries.

qe/: (query expander) source code for the expanders E={q}. Here is the readme for expanders.

+---qe
|   |   main.py
|   +---cmn
|   +---eval
|   +---expanders
|   |       abstractqexpander.py
|   |       backtranslation.py
|   |       thesaurus.py
|   |       ...
|   \---stemmers
|           abstractstemmer.py
|           krovetz.py
|           ...

qs/: (query suggester) source code from cair for the supervised query refinement methods (suggesters), including seq2seq, acg(seq2seq + attn.), and hred-qs.

+---qs
    |   main.py
    \---cair

The following source folders will be populated by the query dataset Q, judgment relevances Jq, and pre-trained models/embeddings.

pre/: (pre-trained models) source folder for pre-trained models and/or embeddings, including FastText and GloVe.

+---pre
|       # anchor_text_en.ttl                => https://downloads.dbpedia.org/2016-10/core-i18n/en/anchor_text_en.ttl.bz2
|       # glove.6B.300d.txt                 => https://nlp.stanford.edu/data/glove.6B.zip
|       # wiki-news-300d-1M.vec             => https://dl.fbaipublicfiles.com/fasttext/
|       # temp_model_Wiki                   => https://www.aclweb.org/anthology/C16-1252/
|       # temp_model_Wiki.vectors.npy       
|       # wiki-anchor-text-en-ttl-300d.vec  => https://github.com/hosseinfani/ReQue/blob/df3bcbdc3189a936f39ade0743450b7871e35517/qe/expanders/anchor.py#L22
|       # wiki-anchor-text-en-ttl-300d.vec.vectors.npy

ds/: (dataset) source folder for original query datasets, including Robust04, Gov2, ClueWeb09-B, ClueWeb12-B13, Antique, and DBPedia.

+---ds
|   +---robust04
|       \---lucene-index.robust04.pos+docvectors+rawdocs
|   +---clueweb09b
|   |   \---lucene-index.clueweb09b.pos+docvectors+rawdocs
|   +---clueweb12b13
|   |   \---lucene-index.clueweb12b13.pos+docvectors+rawdocs
|   +---gov2
|   |   \---lucene-index.gov2.pos+docvectors+rawdocs
|   +---antique
|   |   \---lucene-index-antique
|   +---dbpedia
|   |   \---lucene-index-dbpedia

The target folders are the output repo for the expanders, gold standard datasets, and benchmarks.

qe/output/: The output folder for all expanders and the gold standard datasets. qe/output/backtranslation/: The output directory for backtranslation expander yields results for two information retrieval (IR) metrics, BM25 and QLD, as well as two evaluation metrics, MAP and MRR. These results span across 10 different languages and are based on 5 distinct datasets.

+---qe
|   +---output
|   |   +---backtranslation
|   |   |   +---clueweb09b
|   |   |   +---gov2
|   |   |   +---robust04
|   |   |   +---antique
|   |   |   +---dbpedia
|   |   +---clueweb09b
|   |   +---clueweb12b13
|   |   +---gov2
|   |   +---robust04
|   |   +---antique
|   |   +---dbpedia

qe/output/eval/: output folder for the reports on the performance of expanders and statistics about the gold standard datasets.

+---qe
|   +---output
|   |   \---eval
|   |           overall.stat.csv

qs/output/: output folder for suggesters. This folder contains the benchmark results only and the trained models are ignored due to their sizes.

+---qs
|   \---output
|   |       agg_results.csv

• anserini (Only for indexing purposes)

• cair (optional, needed for benchmark on suggesters)
• python 3.7 and the following packages:

pandas, scipy, numpy, requests, urllib
networkx, community, python-louvain
gensim, tagme, bs4, pywsd, 
nltk -> nltk.download(['averaged_perceptron_tagger','stopwords','wordnet', 'punkt'])
pyserini, pygaggle, pytorch

For the full list, refer to environment.yml. A conda environment, namely ReQue, can be created and activated by the following commands:

$> conda env create -f environment.yml
$> conda activate ReQue

• FastText
• GloVe
• Joint Embedding of Hierarchical Categories and Entities for Concept Categorization and Dataless Classification
• NLLB-200 as the machine translation

Supported Corpora: It should be noted that although ReQue supports the above-mentioned datasets, is not fully dependent on them. In addition to Trec, TrecWebCollection, Clueweb09Collection, Cluweb12Collection, TsvString and TsvInt structure which are already suported by ReQue and Anserini, ReQue can be applied to any TSV format collection as far as the corpus and the queries are in docid\tdocument and qid\query format, respectively. The TSV format collection can be converted to Jsoncollection and then indexed as per instructed in Anserini.

It is suggested to clone the repo and install a new conda environment along with the required packages using yaml configuration file by the following commands:

$> git clone https://github.com/hosseinfani/ReQue.git
$> cd ReQue
$> conda env create -f environment.yml
$> conda activate ReQue

Anserini must be installed in anserini/ for indexing, information retrieval and ranking, and evaluation on the original query datasets. The documents in the corpus must be indexed, e.g., by the following commands for Robust04 (already available here), Gov2, ClueWeb09-B, and ClueWeb12-B13.

For Antique dataset, we first need to convert the antique tsv files into tsv into Anserini's jsonl files as per instructed here,

For DBPedia dataset, you may use Nordlys toolkit in order to download MongoDB DBpedia cleaned collection. Later, you can index it using elastic search or convert it into tsv format and index it using Anserini.

Further, we can use the following commands to index the documents:

$> anserini/target/appassembler/bin/IndexCollection -collection TrecCollection -input Robust04-Corpus -index lucene-index.robust04.pos+docvectors+rawdocs -generator JsoupGenerator -threads 44 -storePositions -storeDocvectors -storeRawDocs 2>&1 | tee log.robust04.pos+docvectors+rawdocs &
$> anserini/target/appassembler/bin/IndexCollection -collection TrecwebCollection -input Gov2-Corpus -index lucene-index.gov2.pos+docvectors+rawdocs -generator JsoupGenerator -threads 44 -storePositions -storeDocvectors -storeRawDocs 2>&1 | tee log.gov2.pos+docvectors+rawdocs &
$> anserini/target/appassembler/bin/IndexCollection -collection ClueWeb09Collection -input ClueWeb09-B-Corpus -index lucene-index.cw09b.pos+docvectors+rawdocs -generator JsoupGenerator -threads 44 -storePositions -storeDocvectors -storeRawDocs 2>&1 | tee  log.cw09b.pos+docvectors+rawdocs &
$> anserini/target/appassembler/bin/IndexCollection -collection ClueWeb12Collection -input ClueWeb12-B-Corpus -index lucene-index.cw12b13.pos+docvectors+rawdocs -generator JsoupGenerator -threads 44 -storePositions -storeDocvectors -storeRawDocs 2>&1 | tee  log.cw12b13.pos+docvectors+rawdocs &
$> anserini/target/appassembler/bin/IndexCollection -collection JsonCollection -input antique-Corpus -index lucene-index-antique -generator DefaultLuceneDocumentGenerator -threads 1 -storePositions -storeDocvectors -storeRaw 2>&1 | tee  log.antique.pos+docvectors+rawdocs &
$> anserini/target/appassembler/bin/IndexCollection -collection JsonCollection -input dbpedia-collection -index lucene-index-dbpedia -generator DefaultLuceneDocumentGenerator -threads 1 -storePositions -storeDocvectors -storeRaw 2>&1 | tee  log.dbpedia.pos+docvectors+rawdocs &
 

Refining queries is done using all expanders by qe/main.py that accept the following arguments:

--corpus: The name of the original query dataset whose queries are to be expanded and paired with the refined queries, if any, which could be one of {robust04, gov2, clueweb09b, clueweb12b13, antique, dbpedia}. Required;

--output: The output path for the gold standard dataset, e.g., qe/output/robust04/. Required;

--ranker: The ranker name which could be any of the available ranker models in anserini(SearchCollection). ReQue has been tested for {bm25,qld} (default: bm25);

--metric: The evaluation metric name which could be any metric from trec_eval. Currently, ReQue has been tested for mean average precision, (default: map);

There are other required parameters that should be set in params.py for each collection, including:

anserini: The path to the anserini library e.g., ../anserini/; Required.

index: The corpus index, e.g. ../ds/robust04/lucene-index.robust04.pos+docvectors+rawdocs; Required

size: Number of documents in the collection e.g., 528155; Required for Onfields and AdapOnFields query expander.

w_t: Weight for title field. Required for Onfields and AdapOnFields query expander.

w_a: Weight for anchor field. Required for Onfields and AdapOnFields query expander.

tokens: Total number of tokens in the collection e.g., 148000000; Required for Onfields and AdapOnFields query expander.

The sample running for robust04 is:

$> python -u main.py --corpus robust04 --output ./output/ --ranker bm25 --metric map 2>&1 | tee robust04.bm25.log &
$> python -u main.py --corpus robust04 --output ./output/ --ranker qld --metric map 2>&1 | tee robust04.qld.log &

The gold standard dataset for each original query dataset is generated in qe/output/{original query dataset}/*.{ranker}.{metric}.dataset.csv.

The columns in the gold standard dataset are:

• qid: the original query id in the input query dataset;

• abstractqueryexpansion: the original query q;

• abstractqueryexpansion.{ranker}.{metric}: the original evaluation value r_m(q, Jq) for a ranker r in terms of an evaluation metric m;

• star_model_count: number of refined queries 0 <= |R_qrm| <= |E| for the original query q that improve the original evaluation value;

• expander.{i}: the name of the expander e_i that revised the original query and improved the original evaluation value. Information about various expanders can be located here.;

• metric.{i}: the evaluation value r_m(q'_i, Jq) of the refined query q'_i.

• query.{i}: the refined query q'_i;

and 0 <= i <= star_model_count. The refined queries are listed in order of descending values.

The golden standard dataset for Robust04 using the ranker bm25 and based on the evaluation metric map (mean average precision) is qe/output/robust04/topics.robust04.bm25.map.dataset.csv and, for instance, includes:

311,Industrial Espionage,0.4382,1,relevancefeedback.topn10.bm25,0.489,industrial espionage compani bnd mr foreign intellig samsung vw mossad

where there is only 1 refined query for the query# 311. The original query Industrial Espionage is revised to industrial espionage compani bnd mr foreign intellig samsung vw mossad by relevancefeedback, and the map is improved from 0.4382 (original map) to 0.489.

Another instance is:

306,African Civilian Deaths,0.1196,0

that is no available expander was able to revise the query# 306 such that the revise query improve the performance of ranker bm25 in terms of map.

Cair by Ahmad et al. sigir2019 has been used to benchmark the gold standard datasets for seq2seq, acg(seq2seq + attn.), hred-qs.

The qs/main.py accepts a positive integer n, for considering the topn refined queries, and the name of the original query dataset. n may be chosen to be a large number, e.g., n=100 to consider all the refined queries.

Sample commands for top-5 are:

$> python -u qs/main.py 5 robust04 2>&1 | tee robust04.topn5.log &
$> python -u qs/main.py 5 gov2 2>&1 | tee gov2.topn5.log &
$> python -u qs/main.py 5 clueweb09b 2>&1 | tee clueweb09b.topn5.log &
$> python -u qs/main.py 5 clueweb12b13 2>&1 | tee clueweb12b13.topn5.log &
$> python -u qs/main.py 5 antique 2>&1 | tee antique.topn5.log &
$> python -u qs/main.py 5 dbpedia 2>&1 | tee dbpedia.topn5.log &

By passing all as the name of the original query dataset, it is also possible to merge all the gold standard datasets and do the benchmark:

$> python -u qs/main.py 5 all 2>&1 | tee all.topn5.log &

Statistics shows that for all the rankers, at least 1.44 refined queries exists on average for an original query while the best performance is for Robust04 over bm25 with 4.24.

The average map improvement rate is also reported, given the best refined query for each original query. As shown, the minimum value of map improvement is greater than 100% for all the gold standard datasets which means that even in the worst case, the best refined query for an original query almost doubled the performance of the ranker in terms of map while the gold standard datasets for clueweb09b have improvement rate close or greater than 1000%, meaning that, on average, the best refined query improved each original query by a factor of 10 in that datasets.

For each gold standard dataset belonging to ReQue, given the pairs {(q, q') | q ∈ Q, q' ∈ Rqrm}, the performance of three state-of-the-art supervised query refinement methods including seq2seq, acg(seq2seq + attn.), hred-qs after running the models for 100 epochs has been reported in terms of rouge-l and bleu. As shown, in all the cases, acg(seq2seq + attn.) and hred-qs outperform seq2seq in terms of all the evaluation metric, in most cases, acg(seq2seq + attn.) that uses the attention mechanism outperform hred-qs. The similar observation is also reported by Ahmad et al. Context Attentive Document Ranking and Query Suggestion and Dehghani et al. Learning to Attend, Copy, and Generate for Session-Based Query Suggestion.

Given a gold standard dataset which is built for an original query dataset for a ranker and a metric, the full report can be found here: ds/qs/{original query dataset}.topn5/topics.{original query dataset}.{ranker}.{metric}.results.csv. For instance, for qe/output/robust04/topics.robust04.bm25.map/dataset.csv the benchmarks are in ds/qs/robust04.topn5/topics.robust04.bm25.map/results.csv

Hossein Fani¹, Negar Arabzadeh², Mahtab Tamannaee², Fattane Zarrinkalam², Jamil Samouh², Samad Paydar² and Ebrahim Bagheri²

^{1School of Computer Science, Faculty of Science, University of Windsor, ON, Canada.}

^{2Laboratory for Systems, Software and Semantics (LS3), Ryerson University, ON, Canada.}

@inproceedings{ReQue,
  author    = {Mahtab Tamannaee and
               Hossein Fani and
               Fattane Zarrinkalam and
               Jamil Samouh and
               Samad Paydar and
               Ebrahim Bagheri},
  title     = {ReQue: A Configurable Workflow and Dataset Collection for Query Refinement},
  booktitle = {{CIKM} '20: The 29th {ACM} International Conference on Information and Knowledge Management, Virtual Event, Ireland, October 19-23, 2020},
  pages     = {3165--3172},
  publisher = {{ACM}},
  year      = {2020},
  url       = {https://doi.org/10.1145/3340531.3412775},
  doi       = {10.1145/3340531.3412775}
  }

@inproceedings{ReQueDemo,
  author    = {Hossein Fani and
               Mahtab Tamannaee and
               Fattane Zarrinkalam and
               Jamil Samouh and
               Samad Paydar and
               Ebrahim Bagheri},
  title     = {An Extensible Toolkit of Query Refinement Methods and Gold Standard Dataset Generation},
  booktitle = {{ECIR} '21: Advances in Information Retrieval - 43rd European Conference on IR Research},
  series    = {Lecture Notes in Computer Science},
  volume    = {12657},
  year      = {2021},
  pages     = {498--503},
  publisher = {{Springer}},
  url       = {https://doi.org/10.1007/978-3-030-72240-1_54},
  doi       = {10.1007/978-3-030-72240-1_54}
  }

©2020. This work is licensed under a CC BY-NC-SA 4.0 license.

We benefited from codebases of cair and anserini. We would like to expresse our gratitdue for authors of these repositeries.