SimpleText: Automatic Simplification of Scientific Texts

Title SimpleText: Automatic Simplification of Scientific Texts
Summary Automatic Simplification of Scientific Texts
Keywords NLP
Supervisor Pablo Picazo-Sanchez
Level Master
Status Open

The general public tends to avoid reliable sources such as scientific literature due to their complex language and lacking background knowledge. Instead, they rely on shallow and derived sources on the web and in social media - often published for commercial or political incentives, rather than informational value. Can text simplification help to remove some of these access barriers? The SimpleText track is a part of the CLEF initiative which promotes the systematic evaluation of information access systems, primarily through experimentation on shared tasks. SimpleText addresses the challenges of text simplification approaches in the context of promoting scientific information access, by providing appropriate data and benchmarks. The track uses a corpus of scientific literature abstracts and popular science requests. Our overall use case is to create a simplified summary of multiple scientific documents based on a popular science query which provides a user with an accessible overview of this specific topic.  The track has the following four concrete tasks.

Task 1: Content Selection: Selecting passages to include in a simplified summary ( Given a popular science article from a major international newspaper, this task aims at retrieving all passages that would be relevant for this article, from a large corpus of scientific abstracts and bibliographic metadata. Relevant passages should relate to any of the topics in the source article. Relevant abstracts should relate to any of the topics in the source article. These passages can be complex and require further simplification to be carried out in tasks 2 and 3. Task 1 focuses on content retrieval. Corpus: DBLP + abstracts ( We use the Citation Network Dataset: DBLP+Citation ACM Citation network, the 12th version released in 2020:

An ElasticSearch index is provided to participants with access through an API. A JSON dump of the index is also available for participants. Queries ( Topics are a selection of press articles from the Science section of The Guardian and Tech Xplore, enriched with queries manually extracted from the content of the article. It has been checked that at least 5 relevant abstracts can be found for each query. Evaluation ( Topical relevance: Retrieval effectiveness will be evaluated on: Topic relevance: Not relevant (0) Relevant (1) Highly relevant (2) We will use traditional IR measures to evaluate the effectiveness (NDCG@10, MAP, …). Additional measures: We plan to assess additional (non-topical relevance) aspects: Text complexity: Easy (0) Difficult (1) Very difficult (2) Source credibility: Low (0) Medium (1) High credibility (2)

 Task 2: Complexity Spotting: Identifying and explaining difficult concepts for general audience ( The goal of this task is: to decide which terms (up to 5) require explanation and contextualization to help a reader to understand a complex scientific text – for example, with regard to a query, terms that need to be contextualized (with a definition, example and/or use-case) to provide short (one/two sentence) explanations/definitions for the detected difficult terms. For the abbreviations, the definition would be the extended abbreviation. For each passage, participants should provide a ranked list of difficult terms with corresponding scores on the scale 1-3 (3 to be the most difficult terms, while the meaning of terms scored 1 can be derived or guessed) and definitions. Passages (sentences) are considered to be independent, i.e. difficult term repetition is allowed. Term pooling and automatic metrics (accuracy of term binary classification, NDCG for term ranking, kappa statistics, similarity of the provided definitions to ground-truth definitions…) will be used to evaluate these results.  Task 3: Text Simplification: Scientific text simplification ( The goal of this task is to provide a simplified version of text passages. Participants will be provided with the popular science articles and queries and matching abstracts of scientific papers. The abstracts can be split into sentences. As in 2022, we will evaluate the complexity of the provided simplifications as well as the errors and information distortion which might occur in the simplification process. In 2023, we will expand the training and evaluation data and focus on large-scale automatic evaluation measures (SARI, ROUGE, compression, readability), supplemented with small-scale detailed human evaluation of other aspects.

Task 4: SOTA: Tracking the state-of-the-art in scholarly publications ( Leaderboards are like scoreboards that display top AI model results for specific tasks, datasets, and metrics. Traditionally community-curated, as seen on, text mining could speed up their creation. The goal is to develop systems that recognize if an incoming AI paper reports model performances on benchmark datasets. If it does, the model should extract all related (task, dataset, metric, score) tuples that are reported in the work.