Text-to-Text Extraction and Verbalization of Biomedical Event Graphs

Proceedings of the 29th International Conference on Computational Linguistics, pages 2692–2710 October 12–17, 2022. 2692 Text-to-Text Extraction and Verbalization of Biomedical Event Graphs Giacomo Frisoni*, Gianluca Moro* and Lorenzo Balzani* Department of Computer Science and Engineering (DISI) University of Bologna, Via dell’Università 50, I-47522 Cesena, Italy {giacomo.frisoni,gianluca.moro}@unibo.it balzanilo@icloud.com Abstract Biomedical events represent complex, graph- ical, and semantically rich interactions ex- pressed in the scientific literature. Almost all contributions in the event realm orbit around se- mantic parsing, usually employing discrimina- tive architectures and cumbersome multi-step pipelines limited to a small number of target in- teraction types. We present the first lightweight framework to solve both event extraction and event verbalization with a unified text-to-text approach, allowing us to fuse all the resources so far designed for different tasks. To this end, we present a new event graph linearization tech- nique and release highly comprehensive event- text paired datasets, covering more than 150 event types from multiple biology subareas (En- glish language). By streamlining parsing and generation to translations, we propose baseline transformer model results according to multiple biomedical text mining benchmarks and natu- ral language generation metrics. Our extractive models achieve greater state-of-the-art perfor- mance than single-task competitors and show promising capabilities for the controlled gener- ation of coherent natural language utterances from structured data.1 1 Introduction In recent years, events have become an influential formalism for modeling complex relations men- tioned within the text as semantic graphs (Frisoni et al., 2021, 2022). In bioinformatics, an event generally refers to an interaction between one or more biomedical entities (e.g., proteins, genes, dis- eases, drugs), each contributing with a specific role (e.g., Theme, Cause, Site). For instance, biomedi- cal events include molecular reactions, organism- level outcomes, and adverse drug reactions. Their expressive power and flexibility have supported *Equal contribution. 1The data and the code to reproduce our baseline results are available at https://github.com/disi-unibo- nlp/bio-ee-egv Tumor tissue liberates VEGF into the blood liberates Localization Theme FromLoc tumor tissue Tissue VEGF Gene blood Organism substance ToLoc Event mention Event graph Event-to-Text Encoding Figure 1: Illustration of an event graph and its textual mention from our datasets. All text-event pairs refer to human-crafted annotations above the biomedical litera- ture (abstracts or full papers). many practical applications like literature-based knowledge discovery (Wang et al., 2021b), biologi- cal network construction (Björne et al., 2010), di- agnosis prediction (Zhang et al., 2020c), document summarization (Zhang et al., 2020b), and question answering (Berant et al., 2014). Text-to-event (or event extraction, EE) and event- to-text (or event graph verbalization, EGV) systems effectively bridge natural language and symbolic representations. They provide a step towards de- coupling concept units (what to say) from language competencies (how to say it) (Mel’ˇcuk, 1973). Strongly linked to natural language understanding, EE is a fundamental task to automatically identify, monitor, and aggregate the relational knowledge disseminated within life science papers, speeding up medical progress and promoting discoveries. Yet, although much attention has been paid to EE, no research efforts have been directed to its in- …

Content truncated to 500 words