Robot-Assisted minimally invasive robotic surgery is the gold standard for the surgical treatment of many pathological conditions, and several manuals and academic papers describe how to perform these interventions. These high-quality, often peer-reviewed texts are the main study resource for medical personnel and consequently contain essential procedural domain-specific knowledge. The procedural knowledge therein described could be extracted, e.g., on the basis of semantic parsing models, and used to develop clinical decision support systems or even automation methods for some procedure’s steps. However, natural language understanding algorithms such as, for instance, semantic role labelers have lower efficacy and coverage issues when applied to domain others than those they are typically trained on (i.e., newswire text). To overcome this problem, starting from PropBank frames, we propose a new linguistic resource specific to the robotic-surgery domain, named Robotic Surgery Procedural Framebank (RSPF). We extract from robotic-surgical texts verbs and nouns that describe surgical actions and extend PropBank frames by adding any of new lemmas, frames or role sets required to cover missing lemmas, specific frames describing the surgical significance, or new semantic roles used in procedural surgical language. Our resource is publicly available and can be used to annotate corpora in the surgical domain to train and evaluate Semantic Role Labeling (SRL) systems in a challenging fine-grained domain setting.

This paper introduces our proposed system for the MIA Shared Task on Cross-lingual Openretrieval Question Answering (COQA). In this challenging scenario, given an input question the system has to gather evidence documents from a multilingual pool and generate from them an answer in the language of the question. We devised several approaches combining different model variants for three main components: Data Augmentation, Passage Retrieval, and Answer Generation. For passage retrieval, we evaluated the monolingual BM25 ranker against the ensemble of re-rankers based on multilingual pretrained language models (PLMs) and also variants of the shared task baseline, re-training it from scratch using a recently introduced contrastive loss that maintains a strong gradient signal throughout training by means of mixed negative samples. For answer generation, we focused on languageand domain-specialization by means of continued language model (LM) pretraining of existing multilingual encoders. Additionally, for both passage retrieval and answer generation, we augmented the training data provided by the task organizers with automatically generated question-answer pairs created from Wikipedia passages to mitigate the issue of data scarcity, particularly for the low-resource languages for which no training data were provided. Our results show that language- and domain-specialization as well as data augmentation help, especially for low-resource languages.