LEGO-ABSA: A Prompt-based Task Assemblable Unified Generative Framework for Multi-task Aspect-based Sentiment Analysis

​

gaotianhao@pku.edu.cn

{junfang8, liuhanyu11, liuzhiyuan8, liuchao397,

liupengzhang, baoyongjun, Paul.yan}@jd.com

​

Introduction

• ABSA is a fine-grained sentiment analysis task, aims to extract different elements including:1) the aspect term (a); 2) opinion term(o); 3) the aspect category(c); 4) the sentiment polarity(s)

binary

triplet

quadruple

• ABSA has multiple sub-tasks, As shown in left figure^[1], We mainly focus on binary, triplet, quadruple extraction. From bianry to quadruple, the difficulty is gradually increases.

The target of different sub-tasks are shown in below table.

[1] Zhang, Wenxuan, et al. "A Survey on Aspect-Based Sentiment Analysis: Tasks, Methods, and Challenges." arXiv preprint arXiv:2203.01054 (2022).

Previous generative Method

Recently, large-scale generative language models have become increasingly powerful, and any ABSA task can be converted to a generative problem. Many generative methods have been proposed But they:

•  usually only one subtask is solved at a time, and if you want to solve multiple subtasks, you need to train a model for each subtask.
• treating the output as a whole string rather than a combination of elements, which have different meanings. Intuitively, It should be treated differently.
• have poor transferability from simple task to difficult task.

​

A generative AOPE model

“Pizza is delicious”

“Pizza, delicious”

WHY NOT

“Pizza, delicious”

“Pizza, POS”

“Pizza, food ”

Related Work

• Towards generative aspect-based sentiment analysis (zhang et al., 2021a)

- Applied extractive-style and annotated-style generative method to several independent sub-tasks, and the SoTA effect is achieved.

• A unified generative framework for aspect-based sentiment analysis . (Yan et al., 2021)

-  Instead of generating text directly, a span index of text is generated, and although multiple subtasks are explored, multiple subtasks cannot be solved using a single model

• Aspect sentiment quad prediction as paraphrase generation . (zhang et al., 2021b)

- Explore the generative approach to the ASQP task and add task specific prompt after text

​

​

Our Contribution

1. For single task training, task prompt is split into Element prompt[1] to achieve the purpose of transforming the output into a combination of different elements .
2. Multiple tasks can be trained under a set of training framework, so as to solve multiple sub-tasks with one model.
3.  The first one proposes the Task Transfer scenario and completes the first exploration of this scenario through Task Prompt Assemble method that we defined .

​

[1]element prompt

Problem Formulation of LEGO-ABSA

Single Task Training

​

•   In Single task training, taking AOPE as an example (the elements to be extracted are Aspect term and Opinion term), the task prompt can be obtained by combining the corresponding elements prompt. Inputs and outputs of the model could be constructed according to task prompt.�

• The GAS^[1] model simply views the output as a string, rather than a combination of Elemenets

• [1]Towards generative aspect-based sentiment analysis (zhang et al., 2021a)

Multitask Training

• For multi-task training, the global mapping of Sentinel token and Element category is constructed on the basis of single task training setting.
1. Sentinel token maintains global mapping with element. This disambiguates what sentinel token refers to.
2. Multiple subtasks can be trained in the same framework and performance is better than simply adding a text prefix such as “ASQP” for ASQP task after the input text ^[1].

​

[1] Aspect sentiment quad prediction as paraphrase generation . (zhang et al., 2021b)

​

Task Transfer Scenario

Task Transfer Scenario：complete Advanced task（such as ASTE task） by training only on basic tasks（such as AOPE and E2E-ABSA tasks）.

core：

1. By fusing element prompt shared across multiple task prompts(of basic task) and multiple task prompts are combined into one(of advanced task), like piecing together lego blocks.
2. Maintain the order of the Element Prompt arrangement

training：Basic Tasks each added their own task prompt and mixed the data for training�

inference： Use the task prompt corresponding to the Advanced Task for inference

Experiments

Analysis

Table 8: Lego split case for text "tech support would not fix the problem unless I bought your plan for $