We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

cs.CL

Change to browse by:

cs

References & Citations

DBLP - CS Bibliography

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Computation and Language

Title: UnifiedABSA: A Unified ABSA Framework Based on Multi-task Instruction Tuning

Abstract: Aspect-Based Sentiment Analysis (ABSA) aims to provide fine-grained aspect-level sentiment information. There are many ABSA tasks, and the current dominant paradigm is to train task-specific models for each task. However, application scenarios of ABSA tasks are often diverse. This solution usually requires a large amount of labeled data from each task to perform excellently. These dedicated models are separately trained and separately predicted, ignoring the relationship between tasks. To tackle these issues, we present UnifiedABSA, a general-purpose ABSA framework based on multi-task instruction tuning, which can uniformly model various tasks and capture the inter-task dependency with multi-task learning. Extensive experiments on two benchmark datasets show that UnifiedABSA can significantly outperform dedicated models on 11 ABSA tasks and show its superiority in terms of data efficiency.
Subjects: Computation and Language (cs.CL)
Cite as: arXiv:2211.10986 [cs.CL]
  (or arXiv:2211.10986v1 [cs.CL] for this version)

Submission history

From: Zengzhi Wang [view email]
[v1] Sun, 20 Nov 2022 14:21:09 GMT (504kb,D)

Link back to: arXiv, form interface, contact.