Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Multilingual Target-Stance Extraction

Ethan Mines, Bonnie Dorr

TL;DR

The paper addresses multilingual Target-Stance Extraction by extending English-only TSE to six languages using a unified two-stage pipeline: an mT5-based target generator and a BERTweet stance classifier, with a post-hoc mapping of generated targets to a fixed pool via English verbalizations. It reports a modest overall TSE F1 of $F_{ ext{TSE}}=12.78$, underscoring the added difficulty of multilingual target prediction and the bottleneck in generating accurate targets. It also reveals that F1 is sensitive to target verbalizations and provides a baseline resource for future multilingual TSE research. The authors propose three directions for improvement: multilingual keyphrase corpora, enhanced target-mapping embeddings, and broader, open-world evaluation datasets.

Abstract

Social media enables data-driven analysis of public opinion on contested issues. Target-Stance Extraction (TSE) is the task of identifying the target discussed in a document and the document's stance towards that target. Many works classify stance towards a given target in a multilingual setting, but all prior work in TSE is English-only. This work introduces the first multilingual TSE benchmark, spanning Catalan, Estonian, French, Italian, Mandarin, and Spanish corpora. It manages to extend the original TSE pipeline to a multilingual setting without requiring separate models for each language. Our model pipeline achieves a modest F1 score of 12.78, underscoring the increased difficulty of the multilingual task relative to English-only setups and highlighting target prediction as the primary bottleneck. We are also the first to demonstrate the sensitivity of TSE's F1 score to different target verbalizations. Together these serve as a much-needed baseline for resources, algorithms, and evaluation criteria in multilingual TSE.

Multilingual Target-Stance Extraction

TL;DR

The paper addresses multilingual Target-Stance Extraction by extending English-only TSE to six languages using a unified two-stage pipeline: an mT5-based target generator and a BERTweet stance classifier, with a post-hoc mapping of generated targets to a fixed pool via English verbalizations. It reports a modest overall TSE F1 of , underscoring the added difficulty of multilingual target prediction and the bottleneck in generating accurate targets. It also reveals that F1 is sensitive to target verbalizations and provides a baseline resource for future multilingual TSE research. The authors propose three directions for improvement: multilingual keyphrase corpora, enhanced target-mapping embeddings, and broader, open-world evaluation datasets.

Abstract

Social media enables data-driven analysis of public opinion on contested issues. Target-Stance Extraction (TSE) is the task of identifying the target discussed in a document and the document's stance towards that target. Many works classify stance towards a given target in a multilingual setting, but all prior work in TSE is English-only. This work introduces the first multilingual TSE benchmark, spanning Catalan, Estonian, French, Italian, Mandarin, and Spanish corpora. It manages to extend the original TSE pipeline to a multilingual setting without requiring separate models for each language. Our model pipeline achieves a modest F1 score of 12.78, underscoring the increased difficulty of the multilingual task relative to English-only setups and highlighting target prediction as the primary bottleneck. We are also the first to demonstrate the sensitivity of TSE's F1 score to different target verbalizations. Together these serve as a much-needed baseline for resources, algorithms, and evaluation criteria in multilingual TSE.
Paper Structure (12 sections, 2 equations, 2 figures, 7 tables)

This paper contains 12 sections, 2 equations, 2 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Stance Detection
  4. Multilingual Stance Detection
  5. Target-Stance Extraction
  6. Methodology
  7. Problem Definition
  8. TSE Pipeline
  9. Data
  10. Experiments
  11. Results and Analysis
  12. Conclusions and Future Directions

Figures (2)

  • Figure 1: Stance Detection versus Target-Stance Extraction
  • Figure 2: Overview of Multilingual Target-Stance Extraction. A machine-translated keyword corpus is used to fine-tune an mT5 sequence model for target prediction. A BERTweet model predicts the stance given the original document and the target prediction. Using FastText embedding similarity, the free-form predicted target is mapped to a set of fixed targets for evaluation purposes.