Transformers at SemEval-2020 Task 11: Propaganda Fragment Detection Using Diversified BERT Architectures Based Ensemble Learning

Proceedings of the 14th International Workshop on Semantic Evaluation, pages 194–200 Barcelona, Spain (Online), December 12, 2020. 194 DCC-Uchile at SemEval-2020 Task 1: Temporal Referencing Word Embeddings Frank D. Zamora-Reina Department of Computer Science University of Chile & IMFD fzamora@dcc.uchile.cl Felipe Bravo-Marquez Department of Computer Science University of Chile & IMFD fbravo@dcc.uchile.cl Abstract We present a system for the task of unsupervised lexical change detection. Given a target word and two corpora spanning different periods of time, automatically detects whether the word has lost or gained senses from one corpus to another. Our system employs the temporal referencing method to obtain compatible representations of target words in different periods of time. This is done by concatenating corpora of different periods and performing a temporal referencing of target words i.e., treating occurrences of target words in different periods as two independent tokens. Afterwards, we train word embeddings on the joint corpus and compare the referenced vectors of each target word using cosine similarity. Our submission was ranked 6th among 33 teams for subtask 1, obtaining an average accuracy of 0.637, only 0.050 points behind the first ranked system. 1 Introduction In the past decade we have seen a rise in academic work on the automatic detection of lexical semantic change (Tahmasebi et al., 2018; Kutuzov et al., 2018). However, many of these studies are difficult to compare because of differences in evaluation procedure, languages, corpora, and time periods. SemEval 2020 Task 1 aims to introduce a simple evaluation framework for unsupervised lexical semantic change detection (Schlechtweg et al., 2020). The task focuses on four languages: German, English, Swedish, and Latin. For each language, two corpora (C1 and C2) are provided, each of those spanning different periods of time (t1 and t2). The task is divided into two subtasks described below. 1. Subtask 1 - Binary classification: for a set of targets words, decide which words lost or gained senses between t1 and t2; as annotated by human judges. 2. Subtask 2 - Ranking: rank the same set of target words according to their degree of lexical semantic change between t1 and t2 (the stronger the change the higher the position in the ranking). Systems are evaluated against ground truth data annotated by human native speakers (except for Latin, which was annotated by scholars of Latin) (Schlechtweg et al., 2018). In this paper, we present our system for detecting semantic change of words in different periods of time. Our system was solely designed for subtask 1. The rationale of our approach is to obtain compatible representations of target words in different periods of time. This is done by concatenating corpora of different periods and performing a temporal disambiguation of target words. Afterwards, we train word embeddings on the joint corpus and compare the disambiguated vectors of each target word using cosine similarity. This process is known as temporal referencing and was proposed in (Dubossarsky et al., 2019). Our submission was ranked 6th among 33 teams for subtask 1, obtaining an average accuracy …

Content truncated to 500 words