ICNLSP 2023: Deep Learning-Based Claim Matching with Multiple Negatives Training, Anna Neumann.

Title of the presentation: Deep Learning-Based Claim Matching with Multiple Negatives Training.
By: Anna Neumann, Dorothea Kolossa, Robert M Nickel.
Ruhr-Universität Bochum, Germany.
Technische Universität Berlin, Germany.
Bucknell University, Lewisburg, PA, USA.
6th International Conference on Natural Language and Speech Processing.
icnlsp.org/2023welcome
Abstract:
Numerous approaches for the implementation of automated fact-checking pipelines have been
proposed and reviewed recently (Guo et al., 2022). A key part in these pipelines is a claim
matching module that seeks to match new incoming claims with potentially existing, verified claims in a database of completed fact checks. To that end, we propose a modification of the two-stage deep learning-based approach for claim matching which won the CLEF CheckThat! 2022 Subtask 2A Challenge (Shliselberg and Dori-Hacohen, 2022). With our modification, we were able to reduce the error rate of the winning algorithm by more than 20%. This was accomplished by employing a loss function that fuses information from not only a single, but from multiple non-matching (i.e. negative) examples into the training process at each iteration.