FrameNet Brasil (UFJF)

Mentors: Student: Aleksey Dorkin

<p>The proposed project is concerned with the implementation <em>Automatic Generation of Qualia Relations between Lexical Units in FrameNet</em>. The main point of the project is to enrich the data in the existing <em>FrameNet-BR</em> database with some of the information available in <em>BabelNet</em>, more specifically the relations between various word senses.</p> <p>The project is going to be composed of two individual components encompassing two independent processes: an extraction tool and an annotation tool.</p> <p>The extraction tool is meant to utilize the available information in <em>FrameNet-BR</em> and <em>BabelNet</em> and generate a set of <em>hypotheses</em>, i.e. possible qualia relations between lexical units.</p> <p>The annotation tool is meant to provide a visual interface to a human annotator, so that they would be able to evaluate the quality of hypotheses generated by the extraction tool. Hypotheses of acceptable quality are expected to be added to the <em>FrameNet-BR</em> database.</p>

Mentors: Student: Prishita Ray

<p>The main motivation for this project is to create a more automated and simplified video annotation pipeline using both image and textual data for the FrameNet webtool. The existing version relies on manual annotation which is a highly tedious task. In order to annotate multimodal corpora, individual frames or ideas depicted within the video need to be extracted using corresponding audio transcriptions and identified objects. This is important to obtain fine grained semantic representations of events and entities. Moreover, the video may be presented in multiple languages that need to be detected and translated to Portuguese. Also individual objects within a video need to be tracked and identified to generate corresponding textual frame elements that can speed up the annotation process.</p>

Mentors: Student: Zheng Xin Yong

<p>My proposal aims to accelerate future semantic annotation projection research by creating two main deliverables. First, before the GSoC 2020 starts, I intend to complete a Python API for retrieving the annotation dataset. Throughout the GSoC 2020, I will modify existing parsers and create neural networks to project semantic frames and frame elements to unannotated sentences. I mainly work with projections on the manually aligned sentences and aligned sentences generated through transitivity. In addition to suggesting techniques for tackling m:n alignment, I also propose metrics for evaluating the projections in binary and graded manners as well as a phraseological method for analyzing projections for metaphorical phrases. If time permits, I would create sentence alignment models for aligning EN-JP, EN-FR, EN-UR, and EN-HI sentences.</p>