UnifiedABSA: Unified Framework for ABSA

• UnifiedABSA is a unified framework that encodes 11 ABSA subtasks into conditional natural-language instructions.
• It leverages the T5-Base model to transform diverse sentiment analysis tasks into a text-to-text generation format.
• Empirical results demonstrate improved data efficiency, strong cross-task transfer, and over 90% storage savings versus dedicated models.

UnifiedABSA is a general-purpose framework for Aspect-Based Sentiment Analysis (ABSA) based on multi-task instruction tuning. It is designed to unify diverse ABSA subtasks within a single pre-trained text-to-text LLM, leveraging explicit instruction-based prompting to guide generation. The framework addresses the limitations of task-specific models—such as redundant parameterization, lack of task transfer, and low data efficiency—by encoding each ABSA subtask as a conditional natural-language instruction. Empirical results on multipurpose datasets demonstrate that UnifiedABSA achieves superior data and storage efficiency compared to conventional dedicated models, supports joint modeling of 11 ABSA subtasks, and enables strong cross-task transfer effects (Wang et al., 2022).

1. UnifiedABSA Architecture

UnifiedABSA employs the T5-Base model as a universal backbone. For each ABSA subtask and instance, a unified structured input—termed the unified natural-language instruction (USI)—is constructed. The USI concatenates (with line breaks and explicit labels): Task Name (special token), Input sentence, Sentiment Options (e.g., “good, ok, bad”), Category Options (if required), and a Template that specifies the output paraphrase structure (e.g., “The aspect terms are <aspect>.”). Task-relevant options are omitted for subtasks where they are inapplicable.

During training, the tokenized USI is encoded, and T5's decoder learns, under teacher-forcing, to output summary text according to the subtask template, filling in placeholders such as <aspect>, <opinion>, <category>, and <sentiment>. Multiple labels are joined using a special separator “[SSEP]”. The final answer is extracted via postprocessing, parsing the generated summaries to yield task-specific outputs (e.g., sets of aspects, opinion terms, or sentiment tuples). Critically, no extra task-specific modules, heads, or parameters are introduced: only the natural-language instruction block differentiates subtasks (Wang et al., 2022).

2. Formalization of ABSA Subtasks and Instruction Templates

UnifiedABSA covers 11 representative ABSA subtasks, each expressed in terms of a template-driven conditional generation task. For each, the USI encodes all necessary information: task, input, label options, and desired output format. The subtasks and their USI formulations are as follows:

Task (Abbrev.)	Elements Extracted	USI Template
Aspect Term Extraction (ATE)	$\{a_i\}$	The aspect terms are <aspect>.
Aspect Category Detection (ACD)	$\{c_i\}$	The aspect categories are <category>. (all categories)
Aspect-Based Sentiment Classification (ABSC)	$\{(a_i,s_i)\}$	<aspect> is <sentiment>. (good, ok, bad)
Category-Oriented Sentiment Classification (COSC)	$\{(c_i,s_i)\}$	<category> is <sentiment>. (good, ok, bad, all categories)
Aspect-Oriented Opinion Extraction (AOOE)	$\{(a_i, \{o_{i,j}\})\}$	The opinions of <aspect> are <opinion>.
Aspect Sentiment Pair Extraction (ASPE)	$\{(a_i,s_i)\}$	The sentiment of <aspect> is <sentiment>. (good, ok, bad)
Aspect–Opinion Pair Extraction (AOPE)	$\{(a_i,o_i)\}$	The aspect-opinion pairs are <aspect>-><opinion>.
Category–Sentiment Pair Extraction (CSPE)	$\{(c_i,s_i)\}$	The category-sentiment pairs are <category>-><sentiment>. (good, ok, bad, all categories)
Aspect–Opinion–Sentiment Triplet Extraction (AOSTE)	$\{(a_i,o_i,s_i)\}$	The triplets are <aspect>–<opinion>–<sentiment>. (good, ok, bad)
Aspect–Category–Sentiment Triplet Extraction (ACSTE)	$\{(a_i,c_i,s_i)\}$	The triplets are <aspect>–<category>–<sentiment>. (good, ok, bad, all categories)
Aspect–Category–Opinion–Sentiment Quad Extraction (ACOSQE)	$\{c_i\}$0	<category> is <sentiment> because <aspect> is <opinion>. (good, ok, bad, all categories)

Additional prompt engineering, such as “The <aspect> is <extra_id_0>” or “What about the <aspect>?”, is applied to certain subtasks to better utilize T5's pretraining objectives (Wang et al., 2022).

3. Multi-task Instruction Tuning Methodology

All ABSA subtasks are converted into an instruction-conditioned text generation format using the USI schema. Training aggregates all data from the $\{c_i\}$1 subtasks: $\{c_i\}$2, where $\{c_i\}$3 is the dataset for the $\{c_i\}$4-th subtask. Each training sample is paired $\{c_i\}$5, where $\{c_i\}$6 is the USI plus review sentence, and $\{c_i\}$7 is the formatted summary with filled slots.

Fine-tuning is performed by minimizing the negative log likelihood over all tokens in $\{c_i\}$8 conditioned on $\{c_i\}$9:

$\{(a_i,s_i)\}$0

Batch construction defaults to random sampling across all tasks, with experiments also performed for uniform undersampling and oversampling. Multi-task parameter sharing is complete: no adapters or heads are added for any subtask. The sole task-differentiating signal arises from the natural-language instruction (Wang et al., 2022).

4. Datasets and Label Projection

UnifiedABSA experiments utilize two datasets (Restaurant-ACOS and Laptop-ACOS), both with original ACOSQE (aspect, category, opinion, sentiment quadruple) annotation. Each quadruple is programmatically projected onto the other 10 subtasks by discarding or rearranging elements and deduplicating as necessary. Dataset statistics are:

Dataset	#Categories	#Quadruples	#Train	#Dev	#Test
Restaurant-ACOS	13	3,658	1,531	170	585
Laptop-ACOS	121	5,758	2,934	326	816

Implicit aspects or opinions are included only when their paired category or sentiment is available within the label set (Wang et al., 2022).

5. Empirical Results and Ablation Analyses

5.1 Full-Supervised Learning

UnifiedABSA exhibits improved data efficiency and transfer relative to dedicated T5 models. In the full data regime, UnifiedABSA's average F1 is 79.99% on Restaurant-ACOS (vs. 77.82% for single-task) and 72.02% on Laptop-ACOS (vs. 70.78%), with consistent per-task improvements of 1–2% F1 across most subtasks.

5.2 Few-Shot and Data-Efficient Learning

In low-resource settings (32 and 64 training examples per task), UnifiedABSA's advantage grows: on Restaurant-ACOS, 32-shot average F1 is 57.29% (vs. 51.34% for single-task); on Laptop-ACOS, 51.05% (vs. 45.50%). Multi-element extraction tasks (e.g., AOPE, AOSTE) benefit most, with F1 improvements >10 points. Average F1 data efficiency curves indicate that UnifiedABSA with half the data matches or exceeds the single-task baseline with full data (Wang et al., 2022).

5.3 Storage Efficiency

UnifiedABSA maintains a single T5-Base instance for all 11 subtasks, yielding >90% storage savings compared to the 11-task-specific model baseline.

5.4 Ablation on Instruction Components

Removing any USI component (task name, label options, template) degrades performance, particularly in few-shot settings. Complete removal of USI collapses average F1 from 57.29% to 26.87% (Restaurant 32-shot), indicating that all three components—explicit instruction, label options, and output template—are vital for generalization and transfer.

5.5 Sampling Strategy and Task Complementarity

Random task sampling per mini-batch outperforms or matches uniform oversampling and undersampling in both few-shot and full regimes. Training on the full set of 11 tasks yields higher generalization than "aspect only," "category only," or other subtask groupings, confirming task complementarity and transfer (Wang et al., 2022).

6. Strengths, Limitations, and Future Directions

UnifiedABSA demonstrates strong joint modeling capability across heterogeneous ABSA subtasks, achieving efficiency and transfer through a simple instruction-based mechanism. It outperforms task-specific T5 in both full and few-shot learning and requires only manual instruction design for extensibility to new tasks. However, the USI schema is currently crafted manually; automated instruction design or learnable soft prompts remain open for exploration. The present work focuses on English and in-domain data; multidomain, cross-lingual, or multimodal extension is an open area for future research. Additional ABSA subtasks or expanding beyond ABSA would further test scalability and generalization (Wang et al., 2022).