Multilingual Semantic Search

Participation Guidelines

The overall organization will follow a CLEF-style evaluation process with a shared dataset composed of a: collection of documents, a set of topics, and a set relevance assessments.
The languages of the collections for the task are: English, French, Italian, German, Greek, Spanish, Swedish, Ukranian, and the queries will also be available in Chinese, Japanese.
In the first round, we plan to evaluate the systems in a classic multilingual lexical search fashion. In the subsequent rounds, the multilingual semantic search aspects will be more prominent by asking the systems, for example, to highlight technical concepts in the topics as well as the documents, to show the contextual meaning of the concept in order to improve the readability of the document and the effectiveness of the system.

Subtask 1 - High Precision:

In this subtask, participants are required to build systems that will help the general public to retrieve the most relevant documents on the Web concerning COVID-19 efficiently. The main focus of this subtask is on the top ranked documents; evaluation measures like Precision at 5 and 10 documents as well as Normalized Discounted Cumulative Gain will be used to compare systems.

Substask 2 - High recall:

In this subtask, the focus is more on the problem of finding as many relevant documents as possible with the least effort. Given a limited amount of resources, such as a time limit and expert availability in time of crisis, there will be a limit on the maximum number of documents that can be retrieved in order to build a set of relevant documents that should be delivered to the general public. Evaluation measures like Recall@k and Area Under ROC will be used to compare the systems. In the first round, the systems will work without relevant information. From the second round, the systems can use the information about the relevance assessments to optimize their systems.

Corpora:

The collection of documents used in this task is composed by the following corpora [...wait for the details...].

The Beautiful corpus contains over 1.8 million articles... The dataset is available here.

Topics:

The topics have been created by selecting 1) a subset of the queries created for the TREC-COVID Task (courtesy of TREC-COVID Task organizers) and 2) a selection of queries made available in the Bing search dataset for Coronavirus Intent which includes queries from all over the world that had an explicit/implicit intent related to the Coronavirus or Covid-19.
For the first round, we provide a set of 30 topics structured in the following way:

<topic>
	<num>query identifier</num>
	<language>ISO 639-1 code</language>
	<keyword>keyword based query</keyword>
	<conversational>the query as a question posed by the user</conversational>
	<explanation>a more detailed explanation of what the set of retrieved documents should look like</explanation>
</topic>
   						

The keyword field represents the “traditional” way a user performs the search on a Web search engine. It is basically a set of keywords, i.e. “surgical mask protection”.
The conversational field is more like a way of asking the same thing in a verbal way, i.e. “does a surgical mask protect from covid-19?”
The explanation field is used to provide information to the assessors when performing relevance assessments, i.e. “The documents retrieved should contain information about …”.
Please, find hereby the links to download the topic for each round.

• Round 1: topics.

Relevance Judgements:

After participants submit their runs, a subset of documents for each run will be pooled for each topic in order to get a sample of documents to judge.
Details of the pooling [...]
Please, find hereby the links to download the relevance judgements for each round.

• Round 1: qrels.

Participant Repository:

Participants are provided with a single repository for all the tasks they take part in. The repository contains the runs, resources, code, and report of each participant.

The repository is organised as follows:

• submission: this folder contains the runs submitted for the different tasks in the different evaluation rounds.
• score: this folder contains the performance scores of the submitted runs.
• code: this folder contains the source code of the developed system.
• resource: this folder contains (language) resources created during the participation.
• report: this folder contains the rolling technical report describing the techniques applied and insights gained during participation, round after round.

Covid-19 MLIA Eval consists of three tasks run in three rounds. Therefore, the submission and score folders are organized into sub-folders for each task and round as follows:

• submission/task1/round1: for the runs submitted to the first round of the first taks. Similar structure for the other tasks and rounds.
• score/task1/round1: for the performance scores of the runs submitted to the first round of the first taks. Similar structure for the other tasks and rounds.

Participants which do not take part in a given task or round can simply delete the corresponding sub-folders.

The goal of Covid-19 MLIA Eval is to speed up the creation of multilingual information acces systems and (language) resources for Covid-19 as well as openly share these systems and resources as much as possible. Therefore, participants are more than encouraged to share their code and any additional (language) resources they have used or created.

All the contents of these repositories are realeased under the Creative Commons Attribution-ShareAlike 4.0 International License.

Task Repository:

Organizers share contents common to all participants through the Multilingual Semantic Search task repository.

The repository is organised as follows:

• topics: this folder contains the topics to be used for task.
• ground-truth: this folder contains the ground-truth, i.e. the qrels, for the task.
• report: this folder contains the rolling technical report describing the overall outcomes of the task, round after round.

Covid-19 MLIA Eval runs in three rounds. Therefore, the topics and ground-truth folders are organized into sub-folders for each round, i.e. round1, round2, and round3.

All the contents of this repository are realeased under the Creative Commons Attribution-ShareAlike 4.0 International License.

Rolling Technical Report:

The rolling technical report should be formatted according to the Springer LNCS format, using either the LaTeX or the Word template. LaTeX is the preferred format.

Submission Guidelines

Participating teams should satisfy the following guidelines:

• The runs should be submitted in TREC format (described below);
• Each group can submit a maximum of five runs for each subtask;
• The code used to produce the runs should be uploaded in a Bitbucket repository provided by the organizers upon the registration to Covid-19 MLIA Eval.

Submission for subtask 1 - High precision

The run must have a limit of 1,000 documents retrieved per topic. Therefore, the file of the run submitted for this task must contain no more than 30,000 lines. Any additional retrieved document will not be considered in the evaluation.

Submission for subtask 2 - High recall

For this task, there is no limit for the number of documents retrieved per topic. However, the maximum number of retrieved documents allowed in a run is 6,000. Therefore, each run can have a variable number of documents retrieved per topic (on average 200 documents per topic), but the file of the run submitted for this task must contain no more than 6,000 lines. Any additional retrieved document will not be considered in the evaluation.

Trec Format:

Runs should be submitted with the following format:

30 Q0 ZF08-175-870  0 4238 prise1
30 Q0 ZF08-306-044  1 4223 prise1
30 Q0 ZF09-477-757  2 4207 prise1
30 Q0 ZF08-312-422  3 4194 prise1
30 Q0 ZF08-013-262  4 4189 prise1
...
   						

where:

• Columns are separated by a white space;
• The first column is the topic number;
• The second column is the query number within that topic, which is currently unused and should always be Q0;
• The third column is the official document number of the retrieved document, which is the number found in the "docno" field of the document;
• The fourth column is the rank at which the document is retrieved;
• The fifth column shows the score (integer or floating point) that generated the ranking. This score must be in descending (non-increasing) order. It is important to include the score so that we can handle tied scores (for a given run) in a uniform fashion (trec_eval sorts documents by these scores, not your ranks);
• The sixth column is called the "run tag" and should be an unique identifier for your group AND for the method used. That is, each run should have a different tag that identifies the group and the method that produced the run.

It is important to include all the columns and have a white space delimiter between the columns.

Submission Upload:

Runs should be uploaded in the repository provided by the organizers. Following the repository structure discussed above, for example, a run submitted for the first round of the Multilingual Semantic Search task should be included in submission/task2/round1.

Runs should be uploaded with the following name convention: <teamname>_task2_<freefield> where teamname is the name of the participating team, task2 is the identifier of the Multilingual Semantic Search task, and freefield is a free field that participants can use as they prefer.

Performance scores for the submitted runs will be returned by the organizers in the score folder, which follows the same structure as the submission folder.

The rolling technical report has to be uploaded and kept update in the report folder.

Here, you can find a sample participant repository to get a better idea of its layout.

Evaluation:

The quality of the submitted runs will be evaluated with the following measures:

• Precision at 5 (P@5)
• Average Precision (AP)
• normalized Discounted Cumulated Gain (nDCG)
• R-Precision (RPrec)