Abstract Writer for Academic Papers

Purpose

Synthesize all completed section drafts into a well-structured abstract of 250 words (range: 225-275) following IMRAD format: Background, Objective, Method, Results, Conclusion. The abstract must be self-contained and readable without the full paper.

When to Trigger

• User invokes /aw-abstract
• User says "write abstract", "摘要", "generate abstract"
• Called automatically at end of Phase 2 (before aw-finalize)
• After all major sections (introduction, methodology, results, discussion, conclusion) have drafts

IMRAD Abstract Structure (250 words total)

Section	Word Count	Purpose
Background	40 words	Problem context + motivation
Objective	20 words	Paper aim + research question
Method	80 words	U-Net architecture, loss function, dataset
Results	80 words	SNR improvement, CCC metric, generalization
Conclusion	30 words	Significance + impact
Total	250 words

Workflow

All Section Drafts (sections/*.tex)
         │
         ▼
┌─────────────────────────┐
│  Read: All Sections     │
│  - introduction.tex     │
│  - methodology.tex      │
│  - results.tex          │
│  - discussion.tex       │
│  - conclusion.tex       │
└───────────┬─────────────┘
            │
            ▼
┌─────────────────────────┐
│  Read: Supporting Docs  │
│  - research-brief.json  │ → hypothesis, novelty
│  - methodology.md      │ → key technical claims, metrics
│  - project.yaml        │ → paper title, authors
└───────────┬─────────────┘
            │
            ▼
┌─────────────────────────┐
│  Draft Abstract         │
│  (IMRAD structure)      │
└───────────┬─────────────┘
            │
            ▼
┌─────────────────────────┐
│  Quality Checks         │
│  - Word count (250±10%) │
│  - Acronyms defined     │
│  - No \cite{}           │
│  - No \ref{}            │
│  - Self-contained       │
│  - Active voice         │
└───────────┬─────────────┘
            │
      ┌─────┴─────┐
      │  Pass?    │
      └─────┬─────┘
       Yes / \ No
        /     \
       ▼       ▼
  ┌──────── ┌──────────┐
  │ Output  │ Revise   │
  │ abstract│ abstract │
  └──────── └──────────┘
            │
            ▼
     sections/abstract.tex

Step-by-Step Procedure

Step 1: Read All Section Drafts

Read all available section drafts in order:

1. introduction.tex — Extract the research problem, motivation, and gap
2. methodology.tex — Extract technical approach, architecture, dataset info
3. results.tex — Extract key quantitative findings and metrics
4. discussion.tex — Extract implications and significance
5. conclusion.tex — Extract summary and contributions

Also read:

• research-brief.json — Research hypothesis, novelty claims, key contributions
• methodology.md — Key technical claims, evaluation metrics, dataset details
• project.yaml — Paper title, authors, journal/target

Step 2: Extract Key Information

For each section, extract and note:

From Introduction:

• The problem being addressed
• Gap in existing research
• Main research question or objective

From Methodology:

• Architecture: "U-Net with..." (encoder-decoder structure, skip connections)
• Loss function: "mixed loss combining..."
• Dataset: name, size, characteristics
• Key technical details for replication

From Results:

• SNR improvement (dB)
• CCC (concordance correlation coefficient)
• Statistical significance (p-values)
• Comparison with baselines
• Generalization performance

From Discussion/Conclusion:

• Main contributions
• Practical implications
• Limitations

Step 3: Draft Abstract by Section

Write each section following the word count targets:

Background (40 words):

Current deep learning methods for [task] often suffer from [problem].
Despite advances in [area], challenges remain in [specific gap].
This necessitates new approaches that [what is needed].

Objective (20 words):

This paper aims to develop a [method] that [what it does] to address [problem].
We validate our approach on [dataset] using [metrics].

Method (80 words):

We propose [method name], a [architecture] designed for [task].
The model employs [key technical features: encoder-decoder structure, skip connections, mixed loss function].
Training uses [optimizer] with [learning rate] on [dataset: size, characteristics].
Evaluation follows [standard protocol] using [metrics: SNR, CCC, accuracy].

Results (80 words):

Experiments on [dataset] demonstrate [metric] improvement of [value] compared to [baseline].
Our method achieves [metric] of [value], outperforming state-of-the-art by [percentage].
Statistical analysis confirms significance (p<0.05, n=[samples]).
The approach shows strong generalization across [conditions], validating its practical applicability.

Conclusion (30 words):

We contribute [method] that advances [task] with [key result].
This work provides a practical solution for [application area],
paving the way for [future direction].

Step 4: Quality Checks

After drafting, perform each check:

Check	Requirement	If Failed
Word count	250 ± 10% (225-275 words)	Revise to fit target
Acronyms defined	All acronyms defined on first use	Add full term before acronym
No \cite{}	Abstract must not contain citations	Remove or paraphrase
No \ref{}	Abstract must not reference figures/tables	Remove or inline the reference
Self-contained	Readable without reading the paper	Rewrite to be standalone
Active voice	Use active voice where possible	Rewrite passive sentences

Self-contained check:

• Does the abstract explain WHAT was done, HOW it was done, and WHY it matters?
• Can a reader understand the paper's contribution from the abstract alone?
• Are all technical terms either common knowledge or defined?

Step 5: Output Final Abstract

Write to manuscripts/{slug}/sections/abstract.tex:

\begin{abstract}
% [Background - 40 words]
Current deep learning methods for medical image segmentation often suffer from poor generalization under domain shift. Despite advances in convolutional architectures, challenges remain in preserving fine-grained details while maintaining robustness across datasets. This necessitates new approaches that explicitly model domain-invariant features.

% [Objective - 20 words]
This paper aims to develop a U-Net variant with mixed loss training that addresses the domain generalization problem in medical imaging. We validate our approach on the MM-WHS dataset using SNR and CCC metrics.

% [Method - 80 words]
We propose GSD-UNet, an encoder-decoder network with domain-adaptive attention mechanisms designed for cross-site cardiac MR image segmentation. The model employs a mixed loss combining Dice and focal loss, with additional domain discriminator regularization. Training uses Adam optimizer with learning rate 1e-4 on 200 annotated subjects from multiple sites. Evaluation follows five-fold cross-validation using Signal-to-Noise Ratio (SNR) and Concordance Correlation Coefficient (CCC) metrics.

% [Results - 80 words]
Experiments on the MM-WHS dataset demonstrate SNR improvement of 4.7 dB compared to baseline U-Net. Our method achieves CCC of 0.92, outperforming state-of-the-art methods by 8.3\%. Statistical analysis confirms significance (p<0.01, n=200). The approach shows strong generalization across four external validation sites, with CCC dropping less than 5\%, validating its practical applicability in multi-center studies.

% [Conclusion - 30 words]
We contribute GSD-UNet that advances cardiac MR segmentation with improved domain generalization. This work provides a practical solution for multi-center clinical deployment, paving the way for real-time diagnostic assistance.

\PACS{PACS code}
\keywords{domain adaptation; medical image segmentation; U-Net; cardiac MRI; deep learning}
\end{abstract}

Abstract Quality Checklist

• Word count is 250 ± 10% (225-275 words)
• All acronyms defined on first use (SNR, CCC, etc.)
• No \cite{} commands present
• No \ref{} commands present
• No figure/table references that require reading the paper
• Abstract is self-contained and understandable alone
• Active voice used where possible
• Each IMRAD section has appropriate word count
• Key quantitative results are included (metrics, percentages)
• \PACS and \keywords fields are populated
• LaTeX compiles without errors

Common Errors to Avoid

1. Including citations — The abstract must stand alone. Remove all \cite{}.
2. Cross-references — Don't write "as shown in Fig. 1" — describe the finding instead.
3. Vague claims — Include specific metrics: "improved by 4.7 dB" not just "significantly improved".
4. Background too long — The problem statement should be concise; save space for results.
5. Methods too detailed — Focus on what makes your approach novel, not every parameter.
6. No results — This is the most common abstract failure. Include specific numbers.
7. Conclusions over claims — Don't overstate; be precise about what was demonstrated.

Acronym Handling

Define all acronyms on first use in the abstract:

Acronym	Full Term	When to Define
SNR	Signal-to-Noise Ratio	Method or Results section
CCC	Concordance Correlation Coefficient	Method or Results section
MRI	Magnetic Resonance Imaging	Background
CT	Computed Tomography	Background
U-Net	U-Net (architecture name, not expanded)	Method
CNN	Convolutional Neural Network	Background
Dice	Dice similarity coefficient	Method
IoU	Intersection over Union	Method

Trigger Variations

• /aw-abstract — Full abstract generation
• /aw-abstract draft — Generate draft only, no quality checks
• /aw-abstract check — Check existing abstract quality
• /aw-abstract revise — Revise existing abstract based on feedback

Error Handling

Error	Response
Section files missing	List which sections are missing; generate abstract from available content
research-brief.json not found	Proceed without hypothesis/novelty; note this in output
methodology.md not found	Extract method details from methodology.tex
Word count far off (>300 or <200)	Show breakdown by section; suggest specific revisions
Contains citations	Remove all \cite{} and rewrite affected sentences
Not self-contained	Identify unclear references; rewrite to inline the information