EOT-DETECT: Emotion & Opinion Extraction

• EOT-DETECT is a structured prompting framework that jointly detects emotions and extracts exact opinion triggers from e-commerce reviews.
• It employs staged reasoning, strict extractive constraints, and a final self-check to maintain precise emotion-to-trigger mapping.
• Evaluated on the EOT-X benchmark, it demonstrates reliable inter-annotator agreement and improved performance over traditional methods.

Searching arXiv for the named topic and related work to ground the article in current literature. arxiv_search(query="EOT-DETECT emotion opinion trigger extraction e-commerce", max_results=5, sort_by="submittedDate") arxiv_search({"query":"EOT-DETECT emotion opinion trigger extraction e-commerce","max_results":5,"sort_by":"submittedDate"}) EOT-DETECT is a structured prompting framework for LLMs developed for the joint e-commerce NLP task of Emotion detection and Opinion Trigger extraction (EOT). Rather than predicting only affective labels, it produces a mapping from a review to one or more emotions and to exact extractive spans that explain why those emotions are present. Introduced in "Why We Feel What We Feel: Joint Detection of Emotions and Their Opinion Triggers in E-commerce" (Attri et al., 7 Jul 2025), the framework combines staged reasoning, extractive constraints, and a final self-check, and is evaluated on the human-annotated EOT-X benchmark of 2,400 reviews.

1. Scope, nomenclature, and problem setting

Within this line of work, EOT denotes Emotion detection and Opinion Trigger extraction: a joint formulation in which customer-review understanding is decomposed into affect recognition and extraction of the causal text spans that ground each affect label. The framework is designed for e-commerce reviews, where actionability depends not only on whether a review expresses joy, trust, disgust, or anger, but also on which exact review substrings encode the cause of that response (Attri et al., 7 Jul 2025).

The nomenclature is potentially ambiguous across arXiv-facing research cultures. The acronym EOT is also used for entropic optimal transport (Gushchin et al., 2022), equivalent oxide thickness (Gu et al., 2012), extraordinary optical transmission (Zhang et al., 2021), the Eocene–Oligocene transition (Weppe et al., 2021), and the Edge of Things (Yigit et al., 2023). In EOT-DETECT, however, the term is strictly an e-commerce NLP abbreviation.

The motivating claim is that emotion labels without explanatory evidence are incomplete for review mining. A single review may express multiple emotions, and a single emotion may have multiple causes. EOT-DETECT is therefore positioned against simpler formulations such as standard emotion classification, generic rationale extraction, and prior emotion-cause pair extraction systems that assume clause-level extraction in other domains. This suggests a shift from coarse affect labeling toward review-grounded, many-to-many emotion–trigger structure.

2. Formal task definition

The task operates on a review represented as

$\mathcal{R} = \{R_i\}_{i=1}^{N}.$

Its emotion inventory is Plutchik’s eight primary emotions,

$$\mathcal{E} = \{\texttt{Joy}, \texttt{Trust}, \texttt{Fear}, \texttt{Surprise}, \texttt{Sadness}, \texttt{Disgust}, \texttt{Anger}, \texttt{Anticipation}\},$$

with an added neutral label when no clear emotion is present.

For an emotion $e$, the corresponding trigger set is defined extractively as

$$\mathcal{T}_e = \{\, (i, j)\,\mid\, (w_i, \dots, w_j) \text{ is a substring explaining emotion } e\}.$$

The extractive constraint is fundamental: every trigger must be a verbatim contiguous span from the original review text. If no non-neutral emotion is detected, the output is

$$\mathcal{O}(\mathcal{R}) = \{(\texttt{Neutral}, \varnothing)\}.$$

This formulation is neither ordinary multi-label classification nor standalone extraction. A review can express multiple emotions, and each emotion can map to multiple triggers. The output is therefore a structured emotion-to-span assignment rather than a flat label set. The framework also constrains the label space to Plutchik’s eight emotions plus Neutral, motivated by four considerations recorded in the study: control of LLM-generated emotion proliferation, a psychologically grounded taxonomy, inclusion of trust and anticipation beyond Ekman’s six, and annotation feasibility (Attri et al., 7 Jul 2025).

3. Prompt architecture and reasoning procedure

EOT-DETECT is defined as a four-part prompt structure,

$\mathcal{P}= \langle S, T, I, R \rangle,$

where $S$ is the system message, $T$ the task description, $I$ the instructions, and $R$ the input review.

Component	Formal element	Function
System message	$$\mathcal{E} = \{\texttt{Joy}, \texttt{Trust}, \texttt{Fear}, \texttt{Surprise}, \texttt{Sadness}, \texttt{Disgust}, \texttt{Anger}, \texttt{Anticipation}\},$$0	establishes the model as an expert in emotion–trigger analysis
Task description	$$\mathcal{E} = \{\texttt{Joy}, \texttt{Trust}, \texttt{Fear}, \texttt{Surprise}, \texttt{Sadness}, \texttt{Disgust}, \texttt{Anger}, \texttt{Anticipation}\},$$1	defines task, trigger extraction, emotion scope, and guidelines
Instructions	$$\mathcal{E} = \{\texttt{Joy}, \texttt{Trust}, \texttt{Fear}, \texttt{Surprise}, \texttt{Sadness}, \texttt{Disgust}, \texttt{Anger}, \texttt{Anticipation}\},$$2	provides the staged reasoning sequence
Input review	$$\mathcal{E} = \{\texttt{Joy}, \texttt{Trust}, \texttt{Fear}, \texttt{Surprise}, \texttt{Sadness}, \texttt{Disgust}, \texttt{Anger}, \texttt{Anticipation}\},$$3	supplies the review text to analyze

The task description is itself structured as a tuple containing Task Definition, Opinion Trigger Definition, Emotion Scope, and Guidelines. The emphasis is that triggers must be exact substrings from the review, multiple emotions are allowed, and no text modification is permitted. This is a prompt-level mechanism for constraining hallucination and paraphrastic evidence generation.

The instruction sequence contains five reasoning steps. The model is asked to focus on key emotions, link each emotion to one or more opinion triggers, maintain balance so that minority emotions are not omitted, keep the final set clear and extractive, and then perform a final self-check. That self-check verifies four conditions: Emotion Coverage Check, Opinion Trigger Coverage Check, Emotion Faithfulness Check, and Opinion Trigger Verifiability Check. The self-check is the distinctive self-reflection stage in the framework and is presented as the main defense against LLM overgeneralization and unsupported trigger generation (Attri et al., 7 Jul 2025).

The execution abstraction is

$$\mathcal{E} = \{\texttt{Joy}, \texttt{Trust}, \texttt{Fear}, \texttt{Surprise}, \texttt{Sadness}, \texttt{Disgust}, \texttt{Anger}, \texttt{Anticipation}\},$$4

with the model $$\mathcal{E} = \{\texttt{Joy}, \texttt{Trust}, \texttt{Fear}, \texttt{Surprise}, \texttt{Sadness}, \texttt{Disgust}, \texttt{Anger}, \texttt{Anticipation}\},$$5 returning the structured emotion–trigger assignment. The study does not provide a full verbatim production prompt, a machine-enforced JSON schema for EOT-DETECT itself, or a component-wise ablation isolating the contribution of each prompt stage. A plausible implication is that the framework should be understood as a prompt design pattern rather than a formally specified decoding protocol.

4. EOT-X benchmark, sampling, and annotation protocol

EOT-DETECT is evaluated on EOT-X, a human-annotated benchmark containing 2,400 reviews and 7,200 raw annotations from three expert raters per review. The corpus spans Amazon, TripAdvisor, and Yelp, with Amazon further divided into Beauty, Home, Electronics, and Clothing. Sampling uses Simple Random Sampling Without Replacement (SRSWOR), and review length is filtered to

$$\mathcal{E} = \{\texttt{Joy}, \texttt{Trust}, \texttt{Fear}, \texttt{Surprise}, \texttt{Sadness}, \texttt{Disgust}, \texttt{Anger}, \texttt{Anticipation}\},$$6

The annotation protocol requires raters to identify all emotions present from the Plutchik inventory or Neutral, and all corresponding opinion triggers as exact text spans. Trigger guidelines require extractive, clearly linked, self-contained spans; allow multiple emotions per review and multiple triggers per emotion; disallow discontinuous spans; prefer the most informative coherent span in nested cases; and instruct raters not to speculate. Gold labels are aggregated by majority vote for emotions, while trigger spans retain all unique triggers and resolve overlap by keeping the longest span.

Agreement is reported as high. Pairwise emotion agreement ranges from 0.85 to 0.91, with overall average reported as 0.88/0.89 depending on table or text wording. Trigger agreement, computed with token-level partial-boundary overlap, ranges from 0.81 to 0.87, with overall average 0.84. These values are used to support the claim that both emotion labels and trigger spans are reliably annotatable in customer reviews (Attri et al., 7 Jul 2025).

The dataset statistics reinforce the task’s genuinely multi-emotion character. Average emotions per review are 2.39 for Beauty, 2.24 for Clothing, 2.29 for Home, 2.28 for Electronics, 2.96 for TripAdvisor, and 2.71 for Yelp. Joy is the most frequent emotion across all domains, while Fear is rare. This distribution makes the “maintain balance” stage of EOT-DETECT operationally important, because dominant emotions can otherwise crowd out minority ones.

5. Evaluation protocol and empirical performance

The study reports Precision, Recall, and F1 for emotion detection, and Exact Match (EM), Partial Match (PM), ROUGE-1 (R1), and ROUGE-L (RL) for opinion trigger extraction. It does not report a single end-to-end joint metric combining correct emotion prediction with correct trigger extraction. Inference settings are standardized as

$$\mathcal{E} = \{\texttt{Joy}, \texttt{Trust}, \texttt{Fear}, \texttt{Surprise}, \texttt{Sadness}, \texttt{Disgust}, \texttt{Anger}, \texttt{Anticipation}\},$$7

The reported model inventory is not fully internally consistent. The abstract and contributions emphasize 23 LLMs, whereas the experiments section mentions 26 LLMs including 3 PLMs, 3 proprietary, and 20 open-source models. The main prompting baselines are Zero-shot (ZS), Zero-shot Chain-of-Thought (ZS-CoT), and EOT-DETECT (Attri et al., 7 Jul 2025).

Model	Setting	Key reported scores
Claude Sonnet 3.5	EOT-DETECT	Emotion F1 0.81; Trigger EM 0.35; R1 0.83; RL 0.68
Mistral-7B-Instruct-v0.2	EOT-DETECT	Emotion F1 0.86; Trigger EM 0.55; R1 0.82; RL 0.68
Phi-4	EOT-DETECT	Emotion F1 0.86; Trigger EM 0.54; R1 0.82; RL 0.67
Qwen2.5-32B-Instruct	EOT-DETECT	Emotion F1 0.80; Trigger EM 0.42; R1 0.78; RL 0.64

Among proprietary models, the strongest reported result is Claude Sonnet 3.5 + EOT-DETECT, with emotion F1 = 0.81, trigger EM = 0.35, R1 = 0.83, and RL = 0.68. Relative to Claude Sonnet 3.5 in plain zero-shot mode, this corresponds to gains of +0.10 in emotion F1, +0.17 in trigger EM, +0.18 in ROUGE-1, and +0.12 in ROUGE-L. Among open models, Mistral-7B-Instruct-v0.2 and Phi-4 reach emotion F1 = 0.86 under EOT-DETECT, with Mistral achieving EM = 0.55 and Phi-4 EM = 0.54.

The most pronounced improvement is reported for Mistral-7B-Instruct-v0.2. Its emotion F1 rises from 0.54 in ZS and 0.69 in ZS-CoT to 0.86 under EOT-DETECT. Trigger metrics also improve from EM = 0.41, R1 = 0.59, and RL = 0.50 in ZS to 0.55, 0.82, and 0.68, respectively. Phi-4 improves from 0.74 and 0.70 in ZS and ZS-CoT to 0.86 emotion F1, with smaller but still positive trigger gains. Qwen2.5-32B-Instruct improves from 0.63 and 0.69 to 0.80 emotion F1, and from EM = 0.33 to 0.42.

The gains are not universal. For GPT-4o, the reported zero-shot setting outperforms EOT-DETECT on emotion F1 and on most trigger metrics except EM. The same non-monotonic pattern appears for several Gemma variants, Llama-3.1-8B, Llama-3.3-70B, and some smaller Qwen variants. The study therefore supports EOT-DETECT most strongly as a broad trend rather than a strict model-independent law.

Fine-tuning results are materially weaker than the best prompted large models. The strongest listed fine-tuned system, Llama-3.2-1B-Instruct-FT (EOT-LLaMA), reaches emotion F1 = 0.49, trigger EM = 0.20, PM = 0.33, R1 = 0.43, and RL = 0.37. This improves substantially over the base 1B Llama zero-shot settings, but remains below top prompted proprietary and mid-sized open models.

6. Error profile, limitations, adjacent tasks, and significance

The reported error modes are concentrated in four areas: trigger boundary detection, emotion conflation, context sensitivity, and trigger multiplicity. Boundary errors arise from over-extended or under-extended spans in multi-clause reviews. Emotion conflation is especially salient for Trust versus Joy. Performance degrades when emotions are implicit, sarcastic, or context-dependent. Models also miss secondary or tertiary triggers when one emotion has several causes. These failure modes align closely with the framework’s design: staged reasoning and self-check are meant to improve coverage and faithfulness, but they do not eliminate boundary ambiguity or subtle pragmatic inference problems (Attri et al., 7 Jul 2025).

Several methodological limitations are explicit. The study includes no formal ablation table isolating the contribution of the self-check, extractive constraints, or individual reasoning steps. It also provides no exact production prompt and no single machine-enforced output schema beyond the conceptual emotion-to-trigger mapping. Closed-model opacity prevents inspection of internal reasoning traces, and some proprietary reasoning models could not be tested because of budget or API-tier constraints. The benchmark itself contains only 2,400 reviews, and trigger annotation is inherently boundary-sensitive even with token-level agreement and longest-overlap aggregation.

In relation to adjacent tasks, EOT-DETECT is positioned against traditional emotion detection, aspect-based sentiment analysis, rationale extraction, and emotion-cause pair extraction. The distinction is structural. Standard emotion detection predicts labels without evidence spans; ABSA targets aspect polarity rather than fine-grained affect categories; generic rationale extraction does not enforce explicit emotion-to-trigger alignment; and prior ECPE systems often assume clause-level extraction in newswire or other non-e-commerce settings. EOT-DETECT therefore occupies a more constrained but also more interpretable niche: multi-emotion, multi-trigger, strictly extractive, and domain-specific.

Its practical significance lies in explainable review analytics. By returning

$$\mathcal{E} = \{\texttt{Joy}, \texttt{Trust}, \texttt{Fear}, \texttt{Surprise}, \texttt{Sadness}, \texttt{Disgust}, \texttt{Anger}, \texttt{Anticipation}\},$$8

the framework supports auditable customer-feedback mining, more actionable business intelligence, and evidence-grounded emotion summaries. At the same time, the reported non-universal gains, absent joint metric, moderate dataset size, and missing component ablations indicate that EOT-DETECT is best understood as a promising prompt-structured baseline for joint emotion–cause analysis rather than a settled endpoint for the task.