Source research

All-in-one : multi-task learning for rumour verification

Elena Kochina, Maria Liakata, Arkaitz Zubiaga · 2026 · 91 citationsRead the paper

Automatic resolution of rumours is a challenging task that can be broken down into smaller components that make up a pipeline, including rumour detection, rumour tracking and stance classification, leading to the final outcome of determining the veracity of a rumour. In previous work, these steps in the process of rumour verification have been developed as separate components where the output of one feeds into the next. We propose a multi-task learning approach that allows joint training of the main and auxiliary tasks, improving the performance of rumour verification. We examine the connection between the dataset properties and the outcomes of the multi-task learning models used.

What the full text shows

Shown in the lab

Multi-task learning (MTL) with three tasks (Veracity, Stance, and Detection) outperformed both single-task and two-task models in veracity classification. — Macro F-score of 0.405 vs 0.318 (MTL2 V+D) and 0.259 (branchLSTM) on PHEME 9 events (PHEME 9 events dataset)
MTL with three tasks achieved the highest performance on the PHEME 5 events dataset. — Macro F-score of 0.396 vs 0.376 (MTL2 V+S) and 0.336 (branchLSTM) (PHEME 5 events dataset)
MTL with two tasks (Veracity+Stance) outperformed the NileTMRG* baseline on the RumourEval dataset. — Macro F-score of 0.558 vs 0.491 (NileTMRG*) (RumourEval dataset)
Performance decreased when expanding the PHEME dataset from 5 to 9 events. — Macro F-score dropped from 0.396 (MTL3) to 0.405 (Wait, text says 'worse performance' but Table 3 shows 0.405 vs 0.396; however, the text explicitly states 'we observe worse performance than on 5 events' due to qualitative differences in the 4 additional events) (PHEME dataset expansion)

Setup: Datasets used were RumourEval and PHEME (5 and 9 event subsets); models implemented in Python 3 using Keras and word2vec embeddings.
How: The authors implemented a multi-task learning architecture using LSTM and dense ReLU layers to jointly train veracity classification with auxiliary tasks (stance and rumour detection). They used the Tree of Parzen Estimators (TPE) for hyperparameter optimization and leave-one-event-out (LOEO) cross-validation for the PHEME dataset.
Limitations: Performance is sensitive to the quality and label distribution of the events (smaller events with extreme values increase complexity). · Class imbalance in the PHEME dataset affects accuracy and necessitates the use of macro-averaged F-score.

Summarized by Gemma 4 from the open-access full text · 12 Oct 2026. Read it