How It Works

At Left, Right, and Centre (LRC), we leverage Large Language Models (LLMs) to innovate political discourse, objectifying the subjective elements of politics to provide a framework for critical thinking.
Our highly automated platform curates a spectrum of political viewpoints and uses LLMs to minimize bias and deepen the reader’s understanding of diverse perspectives.
Our strategy is designed to ensure that the AI remains unbiased and maintains impartiality in processing and presenting information.

Navigating the Platform

Every post on our platform addressing an issue is organized into five main sections to facilitate a comprehensive and balanced discussion:

1. Introduction: This section offers a brief overview of the topic.
2. Jargon: Introduces the necessary terms for understanding the core concepts.
3. Viewpoints 💭: This section presents a neutral ground that promotes understanding and seeks to reduce polarization by representing all viewpoints equitably in the ‘Ideological Differences’ segment. The ‘Left’ viewpoint and ‘Right’ viewpoint segments articulate the respective ideological positions.
4. Prominent Voices 📣: Here, opinions by influential figures and institutions on the topic are highlighted. Each of the prominent voices is cited by providing a direct link to the news article it originates from.
5. Sources 📚: In this section, we list some of the articles that are relevant to the topic in IEEE style with citations.

The Process in Depth

Overview

When preparing to publish a topic on our website, we follow a meticulous process that emphasizes our mission while recognizing our constraints. The steps involved are:

1. Content Fetching: We fetch data from a variety of sources to ensure impartiality, and we pull in hundreds of articles.
2. Article Clustering: We group similar articles to streamline our analysis and reduce redundancy.
3. Classification: For each article, we analyze it for key metrics and score it on how left- or right-leaning it is.
4. Summarization: We condense the essential arguments from each article to highlight the key points based on inputs from the classification phase.
5. Feature Extraction: Important terms and influential voices are highlighted to enrich the reader’s understanding.
6. Balanced View Generation: We consolidate the information to present a balanced perspective that acknowledges different viewpoints.
7. Review: The final content undergoes a rigorous review to ensure it meets our standards of fairness and completeness.

Once this is run for each topic (by which the review is a success), the post is published on the website. Let’s go through each of these steps in greater detail.

1. Content fetching

To ensure that we get the full picture of the topic at hand, we source our information both from curated as well as dynamic lists. The curated list has some of the more well-known news sources, so we ensure that we get our news articles from a broad spectrum.

We also find other sources based on searching the content dynamically by searching the topic keywords and include any articles from sources that aren’t available on our curated list. This helps ensure that we include relevant opinions and perspectives that originate from outside sources, rather than simply stick to the known sources.

This results in a sizeable number (about 250-350) of articles in our collection, with which we can proceed to the next step.

2. Article clustering

In our analysis, we noted that news websites relentlessly copy from each other. This results both in unintentional echo chambers, as well as lots of duplication of data. To avoid this, we use sentence-transformers to group similar articles.

We conduct hyperparameter tuning on clustering algorithms by iteratively testing many parameter configurations to optimize cluster validation. The aim is to balance two key metrics:

1. Intra-cluster Homogeneity: Enhancing similarity within clusters, evaluated using methods like the Silhouette Score.
2. Inter-cluster Heterogeneity: Maximizing dissimilarity between clusters.

This model selection process determines the most effective clustering configuration. Once we have article clusters, we pick a representative article from each cluster and can proceed to the next stage

3. Classification

Starting from this step, we begin to extensively use Large Language Models (LLMs) to organize our data. We have designed a prompt to classify an article on a spectrum, between extreme left bias on one end, and extreme right bias on the other.

These are the major metrics we use for classification:

1. Religious and Communal Tensions
2. Caste-Based Politics
3. Economic Inequality
4. National Security
5. Environment and Development
6. Gender Equality and Women’s Rights
7. Foreign Policy
8. Freedom of Speech and Expression
9. Urbanization and Land Use Policies
10. Alignment with political parties
11. Judiciary
12. The basic structure of the constitution
13. Independence of Central Institutions
14. History and Education
15. Uniform Civil Code

Based on these metrics, an article is scored and classified, so we not only know whether it is left-leaning or right-leaning, but we also know how extreme the lean is. As you will see, this will serve us in the upcoming steps.

4. Summarization

With the articles bifurcated into left-leaning and right-leaning, we feed the separated articles into a summarization prompt to get:

• A brief introduction of the topic under discussion
• A summary of all the left-leaning articles
• A summary of all the right-leaning articles

The aggregation and summarization of these articles will allow us to further extract useful information highlights that would interest the reader.

5. Feature extraction

Post classification, we begin to extract features from the summarization of the articles. Since we are working on top of the previously generated summary, this does not exceed the context window of LLMs.

We pick up the political jargon used in the articles so that a reader will grasp the talking points of both sides, even if they are not experts on the subject. We also further scan the summaries for any prominent voices that support those viewpoints. This will be delineated and shown accordingly.

The feature extraction works directly on top of summaries, and new features to extract from the articles can be configured by updating both the summarization as well as the feature extraction prompts.

6. Balanced view generation

We take the summaries, prominent voices, and classification data and feed it to our LLM to produce the content displayed on our platform. The analysis is presented in the form of bullet points, involving a thorough breakdown of the topic, the left view, the right view, as well as the ideological differences between these views.

Thus formed, the results of our analysis finally exit the LLM pipeline for the next step

7. Review

Once the topic has been thoroughly analyzed by the above steps, we have a draft version of the article that we circulate internally among people on both sides of the political spectrum. They will scrutinize our article, and let us know if their viewpoints have any gaping holes in terms of the arguments that have been made or are otherwise misrepresented.

If this is the case, we take their feedback to either fine-tune the topic search, re-run the LLM pipeline (after the article clustering stage) with different representative articles, or inject some of the feedback given to us in article form so that it may better inform the summary.

In conclusion, we have set up Left, Right, and Centre to work with a pipeline that fetches articles for a particular topic, aggregates them by clustering, classifying, and summarizing them, collates the information we have by extracting the features needed to present the article correctly, and interpolates from the bifurcated set to generate a balanced view.