AutoJourn: Multi-Perspective Summarisation, Bias Detection and Bias Neutralisation for LLM-Generated News in Automated Journalism

# AutoJourn: Multi-Perspective Summarisation, Bias Detection and Bias Neutralisation for LLM-Generated News in Automated Journalism
Source: [https://arxiv.org/html/2607.18983](https://arxiv.org/html/2607.18983)
Himel Ghosh1,2,Ahmed Mosharafa1,Georg Groh1 1Technical University of Munich, Germany,2Sapienza University of Rome, Italy Correspondence:[himel\.ghosh@tum\.de](https://arxiv.org/html/2607.18983v1/mailto:[email protected])

###### Abstract

We presentAutoJourn, a demonstration system for multi\-perspective news generation and bias\-aware evaluation using large language models \(LLMs\)\. The system tackles three core challenges in responsible automated journalism: extracting diverse perspectives from unstructured social media discussions, generating summaries that preserve viewpoint diversity, and detecting or mitigating bias in AI\-generated news\. The pipeline integrates advanced prompt engineering with optional retrieval augmentation to produce semantically diverse perspective sets, a multi\-perspective summarisation module that merges conflicting viewpoints into balanced summaries, and a bias analysis suite supporting sentence\-level bias detection and type classification in the generated news article, and automatic neutralisation\. Users can inspect perspective clusters, compare stance\-specific summaries, generate news articles, and apply bias\-aware rewrites directly in the interface\. We evaluate each component with intrinsic metrics—semantic diversity, summary quality, and bias reduction and show improvements over strong baselines while maintaining content fidelity\. A live, publicly accessible demo accompanies the paper to facilitate reproducibility and further research on socially responsible automated journalism\.

AutoJourn: Multi\-Perspective Summarisation, Bias Detection and Bias Neutralisation for LLM\-Generated News in Automated Journalism

Himel Ghosh1,2, Ahmed Mosharafa1, Georg Groh11Technical University of Munich, Germany,2Sapienza University of Rome, ItalyCorrespondence:[himel\.ghosh@tum\.de](https://arxiv.org/html/2607.18983v1/mailto:[email protected])

## 1Introduction

Since the introduction of Transformer architecturesVaswani2017, large language models \(LLMs\)app15116110;brown2020language;bommasani2021foundationhave enabled fluent, domain\-general text generation and are now widely explored in automated journalismgraefe2016guide;diakopoulos2019automating\. While such systems offer scalability, they also risk flattening diverse viewpoints into a single narrative and reproducing subtle biases present in source data or model training signals\. This is particularly problematic when public discourse is increasingly shaped by social media discussionskwak2010twitter;zubiaga2018discourse, where opinions are fragmented, polarized, and influenced by algorithmic filtering and selective exposurepariser2011filter;gentzkow2006media;tversky1974judgment;cinelli2021echo;del2016echo;sunstein2017republic\.

![Refer to caption](https://arxiv.org/html/2607.18983v1/Figs/SystemDesign.png)Figure 1:System pipeline: multi\-perspective extraction, stance\-specific and merged summarisation, perspective\-conditioned news generation, and bias\-aware evaluation and neutralisation\.Existing NLP tools often target isolated components of this challenge\. Summarisation systems, even controllable onesfan2018controllable;goyal2021evaluating, typically merge competing viewpoints into one “neutral” account\. Work on bias detection and mitigationsheng2019woman;bender2021dangers;spinde2021babe;pryzant2020automaticallyidentifies problematic framing but is rarely integrated into end\-to\-end news generation\. As a result, current automated journalism pipelines lack support for \(i\) surfacing multiple perspectives, \(ii\) generating viewpoint\-aware articles, and \(iii\) evaluating the bias of the resulting text\.

In this demonstration paper, we present an integrated system that addresses these gaps by combiningmulti\-perspective extraction and summarisationwithbias evaluationfor automated news writing\. Given raw input text \(e\.g\., a social media conversation\), the system: \(i\) extracts topics and associated stances using LLM\-based perspective identification; \(ii\) produces perspective\-specific and merged multi\-perspective summaries; \(iii\) generates full news articles conditioned on a chosen topic and corresponding multi\-perspective summary; and \(iv\) evaluates the generated article using a neural bias detector and bias\-type classifier, and produces a bias\-neutralised version\.

A key novelty of this work lies in offering a unified, end\-to\-end framework that connects perspective extraction, multi\-perspective summarisation, controlled news generation, and bias evaluation within a single demonstrable system\. While prior research has examined these components in isolation, our prototype is, to our knowledge, the first to operationalize them jointly for automated journalism\. This integration enables users to trace how viewpoints propagate from source discussions into generated narratives and to assess the bias of those narratives through explicit, model\-driven analysis\. By combining interpretability, viewpoint diversity, and bias mitigation, the system illustrates a practical path toward more transparent and socially accountable applications of LLMs in news production\.

## 2Related Work

Automated journalism has been explored both commercially and academically, with systems such as Reuters’ Lynx Insight and The Washington Post’s Heliograf demonstrating how AI can support story drafting and data\-driven reportingreuters\_lynx\_insight\_2018;lynx\_insight\_platform;wapo\_heliograf\_2016;heliograf\_case\_study\_2025\. Research prototypes extend these ideas by generating summaries of political debates or producing “balanced” narratives, though these systems often rely on single\-perspective summaries that overlook viewpoint diversity\.

Topic modeling plays an important role in organising large conversational inputs\. Recently, prompt\-based topic modelingwang2023promptinglargelanguagemodels;mu2024largelanguagemodelsoffer;pham2024topicgptpromptbasedtopicmodelinghas emerged as a flexible alternative, leveraging LLMs to produce coherent, interpretable topics without corpus\-level training\. These methods are particularly suited to dynamic social\-media content and align with the prompt\-driven topic extraction used in our system\.

The rise of LLMs has intensified concerns about bias, fairness, and representational imbalance in AI\-generated news\. Recent studies show that LLM\-generated articles exhibit systematic ideological and topical biasesllm\_bias\_nature\_2024, reflecting both training data disparities and model reasoning patterns\. Surveys on fairness in LLM\-generated content highlight three core dimensions: representational, allocational, and procedural fairness, which are critical for evaluating automated news systemsbias\_fairness\_survey\_2024\. Complementary efforts in bias detection and mitigation, including neural media\-bias classifiers such as BABEspinde2021babeand neutrality style transfer approaches using the Wiki Neutrality Corpuspryzant2020automatically, provide tools for identifying or rewriting biased passages, but they are typically applied in isolation rather than integrated into generation workflows\.

Our work also draws on advances in stance detection and multi\-perspective summarisation \(MPS\) which aims to represent diverse or conflicting viewpoints rather than collapse them into a single narrativefabbri2021;yadav2022\. Frameworks such as PerSphereluo2024retrieve opposing documents and generate perspective\-specific summaries, while LLM\-driven methods leverage prompt engineering, fine\-tuning, and knowledge distillation to produce viewpoint\-aware summarieshayati2024;zhang2025;aly2025\. Multi\-agent debate systems further improve diversity and accuracy by combining complementary model outputski2025\. These approaches show strong promise for capturing distinct stances, but they seldom connect perspective extraction to downstream news writing or subsequent bias evaluation\.

Finally, work on bias mitigation in generated text investigates how to rewrite content into more neutral phrasing while preserving factual contentpryzant2019\. Despite progress, challenges remain due to the subjective and context\-dependent nature of biasghosh2025and the difficulty of balancing neutrality with informativeness\.

Our system bridges several of these lines of research by jointly integrating \(i\) multi\-perspective extraction, \(ii\) prompt\-based topic induction, \(iii\) perspective\-based summarisation, \(iv\) controlled news generation, and \(v\) sentence\-level bias analysis and neutralisation into one end\-to\-end demonstrable pipeline for automated journalism\.

## 3System Description

Given a social\-media discussion, our system identifies diverse viewpoints, produces stance\-specific and merged summaries, generates a full news article conditioned on these perspectives, and finally evaluates and neutralises bias in the generated text\. Figure[1](https://arxiv.org/html/2607.18983#S1.F1)illustrates the overall workflow\.

![Refer to caption](https://arxiv.org/html/2607.18983v1/Figs/AutoJournScreen1.png)Figure 2:Screenshot of the AutoJourn App from the website\. It shows the search results after a query search\.### 3\.1Application Demonstration

Users can access the system through a desktop browser at[https://autojourn\.cloud/](https://autojourn.cloud/)\. The homepage allows users to submit a query, which by default retrieves relevant Reddit discussions\. After selecting a thread, the user proceeds to the topic\-modelling interface by clickingGet Topics & Perspectives\. The system then displays the generated topics alongside their associated multiple perspectives\. Selecting a topic reveals the extractedAgreeandDisagreeperspectives\. By clickingSummarize, the system produces stance\-specific summaries and a merged, multi\-perspective summary\. ChoosingGenerate Newscreates a full news article conditioned on the selected topic and preferred writing style\. On the news page, theDetect Biasfunction highlights biased sentences, with hover tooltips showing predicted bias type and confidence\. TheNeutralise Biasoption provides rewritten neutral alternatives, which can be inserted directly into the article via theReplace Biasbutton\. Users may also export a comprehensive PDF report viaFull Report, containing the source conversation, extracted perspectives, summaries, generated article, bias annotations, and neutralised versions\. Figure[2](https://arxiv.org/html/2607.18983#S3.F2)provides an overview, with additional screenshots in Appendix[Appendix C: Application Screenshots](https://arxiv.org/html/2607.18983#Ax3)\. A full demonstration video is available at:[https://youtu\.be/XZWGlWqhJHQ](https://youtu.be/XZWGlWqhJHQ)\.

### 3\.2Topic Modeling

For topic extraction, AutoJourn implements a prompt\-based topic modeling pipeline inspired by recent LLM\-driven approacheswang2023promptinglargelanguagemodels;mu2024largelanguagemodelsoffer;pham2024topicgptpromptbasedtopicmodeling\. Given a conversation thread or document, a structured zero\-shot prompt instructs the LLM to identify the top five semantically coherent topics, estimate their relative proportions, and list 5 representative keywords per topic\. The Appendix[A](https://arxiv.org/html/2607.18983#A1)provides the exact template used\.

### 3\.3Multi\-Perspective Extraction

This stage identifies distinct viewpoints present in the source text\. Following prior work on perspective elicitationhayati2024, we use a GPT\-4o model prompted to generate multiple*Agree*and*Disagree*stances with concise, criteria\-based justifications\. This enables the model to produce opinions that differ not only in polarity but also in underlying value dimensions\.

To improve diversity and contextual grounding, the system optionally augments the input using a lightweight retrieval\-augmented generation \(RAG\) mechanismlima2025;logé2025\. Given a topic or a statement extracted from the text, we retrieve semantically related documents using TF–IDF similarity and sentence\-embedding reranking \(e\.g\., e5\-large\-v2\)\. Retrieved snippets provide additional argumentative context, supporting richer and more specific perspective generation, in line with evidence\-based prompting frameworks used in multi\-perspective summarisationluo2024\.

The final output of this stage is a structured dictionary of perspectives for each topic:

𝒫\[t\]=\{Agree:p1A,…,pkA;Disagree:p1D,…,pkD\}\.\\begin\{array\}\[\]\{c\}\\mathcal\{P\}\[t\]=\\\{\\texttt\{Agree\}:p\_\{1\}^\{A\},\\dots,p\_\{k\}^\{A\};\\\\ \\texttt\{Disagree\}:p\_\{1\}^\{D\},\\dots,p\_\{k\}^\{D\}\\\}\.\\end\{array\}
See Appendix\-[B](https://arxiv.org/html/2607.18983#A2),[C](https://arxiv.org/html/2607.18983#A3)for the prompt designs\.

### 3\.4Multi\-Perspective Summarisation

Building on the PerSphere frameworkluo2024, the next stage converts extracted perspectives into stance\-specific summaries\. For each topic, the system generates anAgreesummary and aDisagreesummary using prompt\-guided abstractive summarisation\.

After producing the two stance\-specific summaries, the system generates aMergedsummary that synthesises both viewpoints using neutral, balanced language\. Inspired by SubSummjiang2023and multi\-document summarisation researchkouris2021, the merging prompt emphasises equal representation, removal of emotionally charged phrases, and inclusion of evidence from both sides\. See Appendix\-[D](https://arxiv.org/html/2607.18983#A4)for the prompts\.

The result is a triplet\(A,D,M\)\(A,D,M\): an agree\-focused summary, a disagree\-focused summary, and a merged multi\-perspective summary\.

### 3\.5News Article Generation

The topic and merged summary then serve as conditioning signals for generating full news articles\. Using GPT\-4o, the system expands each summary into a coherent news article while preserving the intended perspective framing\. Check Appendix\-[E](https://arxiv.org/html/2607.18983#A5)for the prompt design\. Users can generate a*Balanced/Neutral*article based on the merged summary\.

This mirrors controllable generation approaches used in political and opinion\-aware summarisationfan2018controllable;goyal2021evaluating\.

### 3\.6Bias Detection

To evaluate the generated article, we incorporate a sentence\-level bias analysis moduleghosh2025grounded in prior work on media biasspinde2021babe;pryzant2020automatically\. The system uses a fine\-tuned RoBERTa classifier trained on the BABE datasetspinde2021babeto detect biased sentences\. Each sentence is assigned \(i\) a bias probability and \(ii\) a predicted bias type using the GusNet classifierpowers2025, covering political, demographic, or socioeconomic categories\.ghosh\-werner\-2026\-llmuses this pipeline for real\-time comparison of biases in LLM outputs\.

This produces a transparent, interpretable bias map over the generated article\.

### 3\.7Bias Neutralisation

When biased sentences are identified, the system applies a two\-stage mitigation pipeline inspired by the WNC neutrality workpryzant2019\. First, a fine\-tuned encoder–decoder model \(BART or T5\) rewrites biased sentences into neutral formulations while preserving meaning\. Second, if the model fails to sufficiently neutralise a sentence, a fallback GPT\-based rewrite is triggered using prompt shown in Appendix\-[G](https://arxiv.org/html/2607.18983#A7), following recent LLM\-based style\-transfer workreif2022;furniturewala2024\. This ensures that even difficult cases receive a neutralised variant\.

The result is a fully neutralised version of the generated news article, enabling direct comparison between the original output and a bias\-mitigated alternative\. See Fig\.[3](https://arxiv.org/html/2607.18983#S3.F3)for reference\.

![Refer to caption](https://arxiv.org/html/2607.18983v1/Figs/NeutralisedSentences.png)Figure 3:Application Screenshot: Demonstration of the highlighted biased sentences in the generated news and the neutralised alternatives\.While the individual components draw on existing techniques, the contribution of AutoJourn lies in their unified orchestration into a transparent, end\-to\-end workflow for multi\-perspective news generation\. Previous systems typically address only one stage \(e\.g\., summarization or bias detection\), whereas AutoJourn provides an integrated pipeline with actionable user\-facing controls and bias\-aware rewriting\.

### 3\.8Intended Audience

AutoJournis intended for journalists, NLP researchers, and practitioners in computational social science\. For researchers, it offers a modular testbed for studying multi\-perspective generation, bias detection, and controllable news rewriting\. For journalists and newsroom technologists, the system provides practical tools to inspect perspective diversity, compare stance\-specific summaries, and highlight biased sentences in AI\-generated drafts, supporting transparent and efficient editorial workflows\. Overall,AutoJournserves as a unified platform for exploring responsible automated journalism\.

## 4Evaluation

We evaluate the system along four axes that mirror its main components: \(i\) multi\-perspective extraction, \(ii\) multi\-perspective summarisation, \(iii\) sentence\-level bias detection, and \(iv\) bias neutralisation\. Our goal is not to exhaustively benchmark each sub\-module, but to show that the integrated system produces diverse perspectives, balanced summaries, and effectively detects and reduces bias in generated news\. Detailed metrics, plots, and screenshots are provided in the appendices\.

### 4\.1Multi\-Perspective Extraction

Following prior work on diversity extraction from LLMshayati2024, we evaluate our perspective extraction module on two subsets: \(i\)SocialChem\-101socialchem\), and \(ii\)CMVcmv\. For each input, GPT\-4o generates 6 opinions \(Agree/Disagree\) using our final criteria\-based prompt, optionally with RAG context\.

We adopt the semantic diversity metric of Hayati et al\.hayati2024, computing pairwise cosine distances between sentence embeddings of generated reasons\. Across both datasets, our GPT\-4o variants achieve the highest semantic diversity scores among all models reported in their study\. See Fig[4](https://arxiv.org/html/2607.18983#S4.F4)for the results\. Further evaluation details are in Appendix[Appendix B: Evaluation Results](https://arxiv.org/html/2607.18983#Ax2)Table[3](https://arxiv.org/html/2607.18983#Ax2.T3)\.

Figure 4:Semantic diversity \(criteria\-based prompting\)\. Top: radar plots comparing all models for Social\-Chem\-101 \(left\) and CMV \(right\)\. Bottom: line plot \(left\) and grouped bar chart \(right\) provide complementary views\. Our variants are GPT\-4o with and without RAG\.![Refer to caption](https://arxiv.org/html/2607.18983v1/Figs/radar_semdiv_socialchem_1.png)

![Refer to caption](https://arxiv.org/html/2607.18983v1/Figs/radar_semdiv_cmv_1.png)

![Refer to caption](https://arxiv.org/html/2607.18983v1/Figs/semdiv_lineplot_blue_1.png)

![Refer to caption](https://arxiv.org/html/2607.18983v1/Figs/semdiv_grouped_bars_blue_1.png)

Interestingly, the RAG\-augmented variant shows slightly lower raw diversity than our designed prompting\. Hence, in practice, we use the GPT\-4o prompt in our demo system\.

### 4\.2Multi\-Perspective Summarisation

For each topic, the system generates \(i\) anAgreesummary, \(ii\) aDisagreesummary, and \(iii\) aMergedsummary intended to concisely represent both sides\. Since no established benchmark exists for multi\-perspective summarisation with stance preservation, our evaluation focuses on demonstrating how summaries can be inspected within the system, rather than establishing new state\-of\-the\-art results\.

To obtain a structured, model\-based assessment of summary quality, we employ an LLM\-as\-judge protocol inspired by Luo et al\.luo2024\. The prompt \(Appendix[F](https://arxiv.org/html/2607.18983#A6)\) guides the evaluator to examine faithfulness to the input, coverage of perspectives, and distinctiveness between stances\. This provides a qualitative check on whether the summaries adhere to the intended task definition\.

We additionally report surface\-level fluency scores using Grammarly—not as a proxy for content accuracy or perspective preservation, but as an auxiliary readability indicator for users of the demo system\. These fluency scores should not be interpreted as a comprehensive evaluation metric for summarisation quality\. Table[1](https://arxiv.org/html/2607.18983#S4.T1)reports these values for completeness\.

Table 1:Grammarly evaluation results for summaries generated by the system \(mean±\\pm95% CI\)\. Concat is a naive concatenation of the agree and disagree summary compared with the merged summaryAs expected, merging opposing viewpoints into a single compact paragraph slightly reduces surface fluency relative to stance\-specific summaries\. However, this trade\-off is acceptable in our demonstration context, where the goal is to provide users with accessible mechanisms for inspecting how multiple perspectives are distilled into a single narrative\.

### 4\.3Bias Detection

Our system requires a reliable sentence\-level bias detection to flag potentially biased spans in generated news articles\. We integrate an existing RoBERTa\-based bias detector introduced in prior workghosh2025, publicly released on Hugging Face\.111[https://huggingface\.co/himel7/bias\-detector](https://huggingface.co/himel7/bias-detector)

This detector was originally trained and evaluated on the BABE datasetspinde2021babeunder the standard protocol of Krieger et al\.Krieger2022, and has been evaluated against the DA\-Roberta Baseline of Krieger et al\.Krieger2022\. Our contribution to demonstrate how such a detector can be embedded into an interactive multi\-perspective journalism pipeline\.

In our system, the detector provides sentence\-level bias probabilities that drive several user\-facing features: highlighted biased spans in generated articles, click\-to\-inspect bias types and confidence\. For bias\-type explanations, we integrate the complementary classifier ofpowers2025, allowing users to view the predicted category \(e\.g\., political, racial, socioeconomic\)\.

We are demonstrating how such a detector can be integrated into a multi\-perspective news generation pipeline and how users can interact with the resulting bias signals during the authoring and editing process\.

### 4\.4Bias Neutralisation

To support users in editing biased or subjective passages in generated news, our system incorporates existing bias\-neutralisation models trained on the WNC corpuspryzant2019\. We use the neutralisers \(BART222[https://huggingface\.co/himel7/bias\-neutralizer\-bart](https://huggingface.co/himel7/bias-neutralizer-bart)and T5\-small333[https://huggingface\.co/himel7/bias\-neutralizer\-t5s](https://huggingface.co/himel7/bias-neutralizer-t5s)\) and a pipeline that combines BART with an LLM fallback neutraliser, as described in §[3](https://arxiv.org/html/2607.18983#S3)\. The neutralisation of the biased sentences is triggered when the user clicks on the "Neutralise Bias" button\. Evaluation is performed on a held\-out WNC split plus a small set of biased sentences from BABE and our generated news\. We measure \(i\) content preservation via BLEU/ROUGE, and \(ii\)*bias reduction*as the relative decrease in bias probability from the detector before and after rewriting\.

Table 2:Bias neutralization performance on combined WNC and test set\.Both encoder–decoder models achieve high similarity with neutral references \(ROUGE\-L≈\\approx0\.96\) while substantially decreasing bias probabilities \(82–85% relative reduction\)\. The BART \+ LLM fallback pipeline yields the strongest bias reduction \(about 92%\) with only a minor drop in lexical similarity, confirming that the fallback is beneficial for difficult cases where the fine\-tuned model fails to sufficiently edit the sentence\. See Table[2](https://arxiv.org/html/2607.18983#S4.T2)for results and Appendix[Appendix B: Evaluation Results](https://arxiv.org/html/2607.18983#Ax2)Table[4](https://arxiv.org/html/2607.18983#Ax2.T4)for more detailed results\. In the demo, users can toggle between the original and neutralised article, visually inspect highlighted changes, and export a bias report combining detection, bias typespowers2025, and neutralised text\.

These results show that our system \(i\) extracts diverse, grounded perspectives, \(ii\) produces stance\-aware summaries, and \(iii\) reliably detects and reduces bias in generated news articles\.

As a final step in the bias analysis workflow, users may download a complete Bias Report that compiles all bias\-related outputs into a single, audit\-ready document\. The report aggregates: the original extracted conversation, detected biased sentences with confidence scores, predicted bias types and classifier probabilities, neutralised rewrites for each biased sentence, bias distribution visualisations \(pie charts and statistics\), and the fully neutralised version of the generated article\. A screenshot of the report is given in Fig\.[5](https://arxiv.org/html/2607.18983#S4.F5)\.

![Refer to caption](https://arxiv.org/html/2607.18983v1/Figs/BiasStats.png)Figure 5:Screenshot of the Final Report downloaded from the application\. It demonstrates the Bias Statistics present in a generated news article of this applicationAvailability & Licensing\.The AutoJourn system is available as an online demonstration at[https://autojourn\.cloud](https://autojourn.cloud/)for research and evaluation purposes\. The underlying source code is not publicly released and can be distributed under a research\-only usage agreement\.

## References

## Ethics and Broader Impact Statement

AutoJournis designed to support responsible automated journalism by making the generation process transparent, interpretable, and sensitive to news bias\. Nonetheless, the system operates on large language models and social media data, both of which carry ethical considerations\. First, extracted perspectives may reflect harmful or extreme viewpoints present in online discussions; while these are surfaced for analysis, they should not be interpreted as endorsements of these views\. Second, automated summarisation and news generation systems risk propagating factual errors or subtle framing biases\. To mitigate this,AutoJournincludes explicit bias detection and neutralisation components, and allows users to inspect, revise, or reject generated content\.

No personal user data is collected by the system, and the demonstration uses publicly accessible web content\. We strongly caution against deploying the system in high\-stakes editorial contexts without human oversight\. Automated news\-writing tools can influence public discourse, and their outputs must be critically reviewed to avoid reinforcing misinformation, polarization, or representational harm\. We release this system as a research demonstrator to encourage further work on transparency, viewpoint diversity, and ethical safeguards in AI\-assisted journalism\.

## Appendix A: Prompt Designs

This appendix contains all prompt templates used in the system, including perspective extraction, RAG\-augmented prompting, stance\-aware summarisation, and bias\-neutralisation prompts\.

## Appendix ADynamic Topic Modeling Prompt

1Analyzethefollowingtextandprovideatleast\{self\.num\_topics\}mostsuitabletopics,

2alongwiththeirpercentagesandtherelevant15keywords\.

3Text:

4\{preprocessed\_doc\}

5FormattheoutputasaJSONobject:

6\{

7"topics":\[

8\{

9"name":"Topic1",

10"percentage":25\.0,

11"keywords":\["keyword1","keyword2",\.\.\.\]

12\},

13\.\.\.

14\]\}

## Appendix BMulti\-Perspective Extraction Prompts

### B\.1Free\-Form Zero\-Shot Prompting \(Hayati et al\.hayati2024\)

1Statement:\{statement\}

2Tellmeopinionsaboutthestatementasmanyaspossible

3fromdifferentpeoplewith"Agree"or"Disagree"and

4explainhowtheyhavedifferentopinions\.

5

6OutputshouldbeintheJSONformat:

7\{

81:\{"Stance":"Agree","Reason":"Writeyourreasonhere"\},

92:\{"Stance":"Disagree","Reason":"Writeyourreasonhere"\}

10\}

### B\.2Criteria\-Based Zero\-Shot Prompting \(Hayati et al\.hayati2024\)

1Statement:\{statement\}

2Tellmeopinionsaboutthestatementasmanyaspossible

3fromdifferentpeoplewith"Agree"or"Disagree",

4one\-wordorone\-phrasecriteriathatisimportant

5fortheiropinions,andexplainwhy\.

6

7Output:

8\{

91:\{"Stance":"Agree","Criteria":\["example1","example2"\],

10"Reason":"Writeyourreasonhere"\},

112:\{"Stance":"Disagree","Criteria":\["example1","example2"\],

12"Reason":"Writeyourreasonhere"\}

13\}

### B\.3Final Unified Perspective Extraction Prompt \(Our Version\)

1Giventhestatement:"\{statement\}",generateatleast6

2opinionswith"Agree"or"Disagree"stances,including:

3\-Criteria\(1\-2words\)importantforthestance

4\-Asingle\-sentencereason

5

6Outputformat:

7\{

81:\{"Stance":"Agree",

9"Criteria":\["c1","c2"\],

10"Reason":"Someexplanation"\},

112:\{"Stance":"Disagree",

12"Criteria":\["c1","c2"\],

13"Reason":"Someexplanation"\}

14\}

## Appendix CRAG\-Augmented Prompts

### C\.1Retrieval\-Augmented Perspective Generation Prompt

1Giventhestatement:"\{statement\}",

2andthefollowingsupportingdocuments:

3

4

5\{docs\}

6

7

8Generateatleast6opinionswith"Agree"or"Disagree",

9includingcriteriaandone\-sentencereasons\.

10

11Output:

12\{

131:\{"Stance":"Agree",

14"Criteria":\["example1","example2"\],

15"Reason":"explanation"\},

162:\{"Stance":"Disagree",

17"Criteria":\["example1","example2"\],

18"Reason":"explanation"\}

19\}

## Appendix DSummarisation Prompts

### D\.1PerSphere Summarisation Prompt \(Luo et al\.luo2024\)

1Giventhequeryanddocuments,summarizetheperspectives\.

2

3Requirements:

41\.Includebothpositiveandnegativeclaims\.

52\.Perspectivesmustnotoverlap\.

63\.Summarymustbecloselytiedtothequery\.

74\.Referencesmayexceedone\.

8OutputformatisXML\.

9

10Query:\{query\}

11Documents:\{doc\}

### D\.2Our Prompt for Dual\-Perspective Summarisation

1Givenastatementandasetofperspectives,generate:

2

31\.Asummaryofagreeingperspectives

42\.Asummaryofdisagreeingperspectives

5

6Requirements:

7\-Nooverlapbetweensummaries

8\-Contentmustremaintiedtotheoriginalstatement

9

10Output\(JSON\):

11\{

12"statement":"\{statement\}",

13"summaries":\{

14"agree":"\.\.\.",

15"disagree":"\.\.\."

16\}

17\}

### D\.3PerSphere Summary Merging Prompt

1Giventhesummarizations,mergethemintoonesummary\.

2Donotuseoutsideknowledge\.

3

4Requirements:

51\.Includebothpositiveandnegativeclaims\.

62\.Mergesimilarperspectives\.

73\.UseanXMLoutputformat\.

8

9Summary:\{summaries\}

### D\.4Our Final Summary Merging Prompt

1Giventhestatementandsummaries,mergethemintoonesummary\.

2

3Requirements:

41\.Identifycommonthemes\.

52\.Ensurebothperspectivesareequallyrepresented\.

63\.Useneutral,objectivelanguage\.

74\.Keepthemergedsummaryconcise\.

85\.Includeevidencefrombothsummaries\.

9

10Output\(JSON\):

11\{

12"statement":"\{statement\}",

13"merged\_summary":"\.\.\."

14\}

## Appendix ENews Generation Prompt

1Youareaprofessionaljournalistwritingforamajornewsoutlet\.

2Yourgoalistocraftacompellinganddetailednewsarticle\.

3

4Topic:\{topic\_name\}

5Summary:\{summary\}

6KeyPointsandKeywords:\{keywords\}

7Style:\{tone\_instructions\}

8

9Requirements:

10Writealong,engagingnewsarticle\.

11Includeanattention\-grabbingheadlineatthebeginning\.

12Expandupontheprovidedsummaryusingthelistedkeywords\.

13Usethemnaturallythroughoutthearticle\.

14Includehistoricalcontext,background,oranalysiswhererelevant\.

15Usethespecifiedstyleandtonethroughoutthearticle\.

16

17Instructions:

18Startyourresponsewiththeheadline,followedbythefullarticle\.

## Appendix FGPT\-4o Evaluation Prompt

1Youareanimpartialjudgeevaluatingmulti\-perspectivesummaries\.

2

3Criteria:

41\.Faithfulness

52\.Coverage

63\.Distinctiveness

74\.Ignoreformattingdifferences

8

9Output:

10Rating:\[\[X\]\]

11

12Originalconversation:

13\{conversation\}

14

15AIResponse:

16\{response\}

## Appendix GBias Neutralisation Prompts

### G\.1Encoder\-Decoder Neutralisation Prompt \(Training\)

Used implicitly through fine\-tuning on WNCpryzant2019\. No explicit inference prompt required\.

### G\.2LLM Safety\-Fallback Neutralisation Prompt

1Thefollowingsentencecontainsbiasedlanguage:

2

3\{sentence\}

4

5Rewriteitinaneutralandobjectivetone\.

## Appendix B: Evaluation Results

Table 3:Semantic diversity \(mean cosine distance; higher is better\) for criteria\-based prompting onSocial\-Chem\-101andCMV\.Table 4:Bias neutralization performance on combined WNC \+ custom test set\.
## Appendix C: Application Screenshots

![Refer to caption](https://arxiv.org/html/2607.18983v1/Figs/AutoJournScreen1.png)Figure 6:Screenshot of the AutoJourn App from the website\. It shows the search results after a query search\.![Refer to caption](https://arxiv.org/html/2607.18983v1/Figs/multiPSumm.png)Figure 7:Demonstration of the Multiple Perspectives being displayed on the right\-side panel of the AutoJourn Application and the Summarization\.![Refer to caption](https://arxiv.org/html/2607.18983v1/Figs/SummaryAgreeDisagree.png)Figure 8:Summarisation module demonstration: shows the agree\-disagree stance specific summaries based on a particular topic\.![Refer to caption](https://arxiv.org/html/2607.18983v1/Figs/BiasClassified.png)Figure 9:User interface demonstration showing bias\-type classification in action within the automated journalist application\.![Refer to caption](https://arxiv.org/html/2607.18983v1/Figs/ReplaceBias.png)Figure 10:Example of bias neutralization in our system: the biased sentence is rewritten into a neutral form while preserving factual accuracy\.