Parliament is an open-source news project that organizes news by Confidence, Bias, Newsworthiness, and Context to tell maximally truthful news in a clear and measurable way.
This project is in pre-release, and as such, is a preview. Many features are in development.
Parliament's mission is telling maximally truthful news in a clear and measurable way.
Think about how you understand news. You may consume one or multiple news sources. It may be social media content, cable news, online or print news, your favorite Telegram group, or another news provider. Ultimately, you will infer the claims these sources are making, and when you are sufficiently satisfied, you will either arrive at a conclusion about what you think has happened or consider yourself unsure. Implicitly, you are assigning a unique judgment value to various news outlets, favoring the credibility of some over others. Once you have enough credible confirmations for your own personal threshold, you may consider something to be effectively true.
For people, our perceived credibility in certain outlets is weighed by many things, including our Confidence in the accuracy of the outlet (this X account has been right many times, so they're probably correct now), and by our Bias towards the outlet (this news provider conforms to my beliefs of the world, so I'll trust their interpretation).
We also perceive some news events to be more important than others. This may be due to various factors, such as ramifications that impact our lives, significant geopolitical events, events that confirm or reject our beliefs, local concerns, unique occurrences (man bites dog), or other considerations. For example, on August 9, 1974, when the New York Times published NIXON RESIGNS in capital letters on their front page, they were expressing significant urgency in the event. While the importance of the Story is different to each person, i.e., it may have been more interesting to an American autoworker than to a Nepalese farmer, the New York Times felt it was newsworthy enough for their audience to warrant an all-caps title. This judgment we call Newsworthiness.
News, when told, also includes a varying contextualization around the event that may significantly alter our understanding and perception of an event. In the most simplest case, an image can be so powerful, only to reverse its meaning when it is zoomed out. The fairness and completeness around which a Story is told we call Context.
Each human consumes news differently due to their unique experiences and circumstances. This is partly why the same event can be perceived so differently by different individuals. Parliament's mission is to tell maximally truthful news in a clear and measurable way.
Let's dive further.
Confidence: The likelihood of something being true based on how true it has been in the past.
How do we know that something is true? This is a philosophical question, and there are several product approaches to addressing it. X, for instance, relies on community notes. FactCheck.org focuses on fact-checking. These two approaches work by direct content validation. A strategy that, when implemented correctly, can be extremely accurate.
Parliament takes what you may call the additional level of abstraction to the above. In the view of Parliament, FactCheck.org, @elonmusk, X fact checker, NYT, a random account on IG, etc. are each what we call an Entity. Otherwise put, Parliament does use fact checkers too, it just tries to use all of them. And, based on if they've been right or wrong in the data we've collected, we know how much to trust them going forward. In this mode of thinking, we can assign news Confidence scores, a score between 0.0 and 1.0, where 0 represents maximal certainty that something is false, and 1.0 represents maximal certainty that something is true. Total certainty is unknowable, and both 0.0 and 1.0 can never be achieved.
We will only cover a high level of the confidence algorithm here, as the specifics are written in the code.
Let us say as a human I am introduced to a new user on X/IG/TikTok etc. Now, depending on who introduced me, perhaps the platform's algorithm, perhaps a friend, I may view content differently. Regardless, I will have a measure of doubt as to how much I trust the author (aka Entity), which may change over time.
At Parliament we assume new sources are of even uncertainty; currently we assume new people are 50% likely to be honest.
We then look at the Claims a given Entity has made. If these claims are mutually agreed on by many parties, old and new, we can start to collect a Consensus on a claim, much like a blockchain algorithm, which is a possible future implementation. Once claims form a level of consensus, we can then punish the people who have been wrong, and reward the people who have been right. Of course, once we punish/reward the Entities, other claims may change consensus, creating a finite ripple effect.
Currently we track confidence in an Entity by:
/**
* Calculate the confidence of an entity based on its past statements.
* @param {Entity} entity The entity object.
* @param {Statement[]} statements The array of statement objects.
* @return {number} The confidence of the entity.
*/
function calculateEntityConfidence(entity, statements) {
let totalScore = BASE_CONFIDENCE;
let count = 0;
// Loop through statements most recent first
for (let i = 0; i < statements.length; i++) {
const statement = statements[i];
if (statement.confidence == null) {
continue;
}
count++;
// more recent statements are penalized/rewarded more
const decay = Math.pow(DECAY_FACTOR, i);
const decidedPro = statement.confidence > DECIDED_THRESHOLD;
const decidedAgainst = statement.confidence < 1 - DECIDED_THRESHOLD;
if (!decidedPro && !decidedAgainst) {
continue;
}
let isCorrect = false;
let isIncorrect = false;
if (decidedPro) {
isCorrect = entity.pids.some((pid) => statement.pro.includes(pid));
isIncorrect = entity.pids.some((pid) => statement.against.includes(pid));
} else if (decidedAgainst) {
isCorrect = entity.pids.some((pid) => statement.against.includes(pid));
isIncorrect = entity.pids.some((pid) => statement.pro.includes(pid));
}
// this allows us to weight the confidence equally for distance to 0 or 1
const adjustedConfidence = Math.abs(statement.confidence - 0.5) * 2;
if (isCorrect) {
totalScore +=
CORRECT_REWARD * (1 - totalScore) * decay * adjustedConfidence;
} else if (isIncorrect) {
totalScore +=
INCORRECT_PENALTY * totalScore * decay * adjustedConfidence;
}
}
if (count === 0) {
return null;
}
return Math.max(0, Math.min(1, totalScore));
}
Separately, we calculate a Claim Confidence by:
/**
* Calculate the confidence of a statement based on its entities.
* @param {Statement} statement The statement id object.
* @param {Entity[]} entities The array of entity objects.
* @return {number} The nullable confidence of the statement.
*/
function calculateStatementConfidence(statement, entities) {
if (entities.length === 0) {
return null;
}
let weightedSum = 0;
let totalWeight = 0;
for (let i = 0; i < entities.length; i++) {
const entity = entities[i];
const pro =
statement.pro?.some((pid) => entity.pids.includes(pid)) ?? false;
const against =
statement.against?.some((pid) => entity.pids.includes(pid)) ?? false;
if (!pro && !against || !entity.confidence) {
continue;
}
let confidence = entity.confidence;
// Reverse confidence if the entity is "anti" the statement
if (against) {
confidence = 1 - confidence;
}
// Inverse Quadratic Weighting: weight = 1 - (1 - confidence)^2
// This is done to weight high/lows confidence more heavily
const weight = 1 - Math.pow(1 - confidence, 2);
// Add to weighted sum
weightedSum += weight * confidence;
totalWeight += weight;
}
// if no entities have confidence, don't update
if (totalWeight == 0) {
return null;
}
// Calculate new confidence
const newConfidence = weightedSum / totalWeight;
// Ensure confidence is within [0, 1]
return Math.max(0, Math.min(1, newConfidence));
}
Put simply, an Entity's confidence is calculated like a credit score, while a Statement's confidence is an average. We choose to trigger this only when our confidence in a claim changes beyond a defined ConfidenceDelta, so as to not overburden practical implementations. At any given snapshot, that is our best confidence that something is true, and can be applied to an Entity, Claim, or Post. As entities' confidence score changes, so does our confidence in the claims they make.
Bias: An angular grouping of
Entitiesbased on theirPosts,Claims,Opinions, andPhrases, such that the closer the angle the more maximal the agreement.
In principle, understanding political Bias requires a currency for comparing bias in different Posts, Entities, Phrases and Opinions. In practice to solve this problem, people will often say things like right wing or left wing or progressive or conservative. However, someone that is extremely conservative may agree with someone that is extremely liberal, how do we model this?
Consider a Parliamentary (roll credits) system of government, in which there are different parties that appeal to different voter segments, which in turn correlates a party to bias groups. Let's say in our Parliament the seating is circular, and similar parties seat near one another. We can organize our Parliament by asking the following question:
If I were to seat every member of Parliament at a round table, such that I wanted to maximize agreement and minimize the disagreement between a member and the two people sitting next to the member, how would I seat the people? The reason this is chosen is more philosophical than mathematical, as it assumes there is a center, right, left, and anti-center, what we label the extreme. This is a mode of grouping that we find to be understandable by many people and is a manifestation of the horseshoe theory. If we accept this assumption, we mathematically represent bias as an angle between 0.0 and 360.0 degrees, and calculate angle updates by simple angular arithmetic. The full algorithm written in the code.
/**
* Takes in a list of iterables (entities or statements) and returns avg.
* Accounts for whether the post is pro or against.
* @param {Object[]} iterable A list of entities or statements.
* @param {Object} target The target object, either an entity or statement.
* @param {boolean} isEntityTarget Whether the target is an entity
* @return {number} The bias (angle) of the entity or statement.
* */
function calculateAverageBias(iterable, target, isEntityTarget) {
let x = 0;
let y = 0;
let count = 0;
// Loop through iterable to calculate the average angle
for (let i = 0; i < iterable.length; i++) {
const biasObj = iterable[i];
if (biasObj.bias == null) {
continue;
}
let pro = false;
let against = false;
if (isEntityTarget) {
// Target is entity: Check if posts are in the pro/against lists
pro = biasObj.pro?.some((pid) => target.pids.includes(pid)) ?? false;
against = biasObj.against?.
some((pid) => target.pids.includes(pid)) ?? false;
} else {
// Target is statement: does statement's lists contain the entity's posts
pro = target.pro?.some((pid) => biasObj.pids.includes(pid)) ?? false;
against = target.against?.
some((pid) => biasObj.pids.includes(pid)) ?? false;
}
let bias = biasObj.bias;
// Reverse the bias if it's in the against list
if (against) {
bias = (bias + 180) % 360;
} else if (pro) {
// do nothing
}
const angleInRadians = (bias * Math.PI) / 180;
// Convert the angle to a vector and accumulate
x += Math.cos(angleInRadians);
y += Math.sin(angleInRadians);
count++;
}
// If no valid biases, return null
if (count === 0) {
return null;
}
// Calculate the average vector direction
const averageX = x / count;
const averageY = y / count;
// Convert the average vector back to an angle
const averageBias = Math.atan2(averageY, averageX) * (180 / Math.PI);
// Ensure the angle is in the range [0, 360)
return (averageBias + 360) % 360;
}
We represent the Center with the color Green. This is done to be analogous with American political colors. Left = Blue, Right = Red, hence RGB -> BGR and Green is center. Center content is defined as content that right and left would each agree on or each disagree on, and content that disagrees with extreme. In practice, these are people we seat 1/2 way between left and right at the top of our circle.
Right is our anchor. We need to choose the first people to seat. We define right pseudo-arbitrarily as a traditional conservative mentality;
- more likely than not to favor small government policy
- more likely than not to favor lower taxes and laissez faire economics
- more likely than not to favor family values
- more likely than not to express nationalism and patriotism
- more likely than not to favor stricter law and order measures
Left is defined as those most opposite of our anchor, right. content that right would most disagree with.
Extreme, represented as pink (255,0,255) where green is (0,255,0), is defined as content that right and left would each agree on or each disagree on, content that disagrees with center, and content that expresses Opinions, Phrases, and Claims which are likely to be disagreed with by consensus.
Hence for a given claim we can now say how confident we are that it is true, and how biased are the people who are making it. Likewise for the entities we can record our confidence in their reporting, and how biased they tend to be. This provides significant value for us to understand and contextualize news. But how do we know what news to show people? To answer this question, we must consider another concept, Newsworthiness.
Newsworthiness: The cross-bias consensus for how newsworthy a given
Storyis.
There is almost an unlimited amount of news, and choosing what news to publish is a core part of any platform. For instance, a heavily biased platform may choose to publish news that corroborates their narrative more than news that supports a different narrative. Even within this context, regionalization of the platform, reader interest, and other aspects control what gets published. Under the hood, something is filtering large amounts of information and deciding what is newsworthy. In contrast, social media feeds may show different news to different people, as what is important and interesting to one person may be different than to another. In aggregate some Stories and Posts will be shown more than others, but this may reflect the bias or even abuse of the platform or the algorithm. If we wish to label all Stories, we must determine a way they can be labeled without bias. To solve this question, we calculate Newsworthiness.
A heuristic would be to consider something newsworthy if it is discussed with high frequency. Social media feeds and news outlets rely on this heuristic significantly. As discussed, this is not sufficient, and it favors bias. However, if we already know bias as we have calculated above, we can consider something to be Newsworthy if it not only has a high frequency in our given publication period, but it also has a high frequency amongst some or all bias groups. This exact algorithm is still being calculated but it yields a number between 0.0 and 1.0 where 0.0 represents the least possible newsworthy event and 1.0 represents the most possible newsworthy event. For regionalization we don't change the scale at all, it is a global concept from which we can still filter by region for localized news.
Context: The cross-bias consensus for what information should be included in a
Story.
News providers will invariably include different information when telling a Story. Based on the information included, a viewer may perceive content very differently. If we were to find cross-bias consensus on what Context to include in a given Story, we could maximally ensure fairness in reporting.
We calculate context by ensuring information reporting meets a Center degree of Bias. That is, if we have significant amount of centrist sources, we will aggregate their contextualization of a Story and use that in our reporting. For each Claim, Opinion, or Phrase that is beyond centrist, we should be sure to include labeling of the context, as well as context from the angular opposite. In this way, we are always reporting news with fair Context that matches our Centrist Bias.
A Story is an event that happened.
- It is created from a collection of Posts that are 'talking about the same thing.' (more on this below)
- A Story has a title, a description, a headline, a subheadline, which are textual descriptions of the Story.
- A Story has a "happenedAt" timestamp, which represents when the event happened in the real world.
- A Story has a "newsworthiness" value, which is a number between 0.0 and 1.0, where 1.0 is the most possibly newsworthy event.
- A Story has a lat and long, which are the best estimates of the location of the Story.
- A Story may have photos, which are images that are associated with the Story.
- A Story may have Claims, which are statements that are either supported or refuted by the Posts.
A Post is a social media post (text, image, video), news article, or any other news source posting.
- A Post can have a title, a description, and a body, the latter two of which are optional.
- A Post may have images, videos or other media, or even link to other Posts.
- A Post has an author, which we call an Entity.
- A Post has a "sourceCreatedAt" time which represents when the Post was originally created on the social/news platform.
- A Post can belong to a Story and have Claims.
An Entity is an author, sometimes one person and sometimes a whole news outlet, depending on the level of granularity that can be obtained.
- An Entity has a Bias and Confidence score tracking the Entity's history.
- An Entity is associated with a list of Posts, Claims, Opinions, and Phrases.
A Statement is either a Claim, Opinion, or Phrase.
A Claim is a statement about an event that is verifiable.
- A Claim is a statement that has "pro" and "against" list of Posts either supporting or refuting the Claim.
- A Claim has a "value" field, which is the statement itself.
- A Claim has a "pro" field, which is a list of Posts that support the Claim.
- A Claim has an "against" field, which is a list of Posts that refute the Claim.
- A Claim has a "context" field, which provides more information about the Claim, and helps for search.
- A Claim is fundamentally verifiable; a non-verifiable "claim" is called an
Opinion. - A Claim has an associated
Confidence. - A Claim has associated
Entities.
An opinion is a statement about an event that is not verifiable.
- An Opinion is a statement that has "pro" and "against" list of Posts either supporting or rejecting the Opinion.
- An Opinion has a "value" field, which is the statement itself.
- An Opinion has a "pro" field, which is a list of Posts that support the Claim.
- An Opinion has an "against" field, which is a list of Posts that refute the Claim.
- An Opinion has a "context" field, which provides more information about the Claim, and helps for search.
- An Opinion is fundamentally non-verifiable; a verifiable "opinion" is called a
Claim. - An Opinion has an associated
Bias. - An Opinion has associated
Entities.
A Phrase is a word or group of words that are used to express Posts, Claims, and Opinions.
- A Phrase utilizes repeated and specific language.
- A Phrase has an associated
Bias. - A Phrase has associated
Entities.
Parliament is open source. For questions related to this GitHub repo, please join our discord or raise an issue in this repo.
TheParliament.app (to be released shortly) is a commercial hosted and deployed implementation of Parliament. For investment information and other queries, please email contact@theparliament.app.
- Discord: Parliament currently supports an official Discord server.
- X, Reddit, IG, Tiktok and Others: No current accounts, but these will be created in the near future.
You believe in our mission.
There are serious quality issues with current news. It is often misleading, decontextualized, or even fake. Parliament is an apolitical solution to this problem. If, and only if, this mission is for you, only then do we recommend you contribute.
You should not participate if you are a political activist looking to push a political agenda. Parliament is apolitical, and will always be.
While Parliament is open-source and non-commercial, and the repo can be used and forked according to our use policy (coming soon), TheParliament.app (also coming soon) is the official commercial implementation of Parliament and the only implementation that can use Parliament branding. TheParliament.app is reserving a meaningful amount of equity for top contributors.
Contributors with accepted commits get swag at different milestones.
Learn Flutter, backend, and AI/ML.
Join our Discord to start asking questions and becoming part of the community.
- Clone Parliament to your machine.
- Install our frontend framework Flutter.
- (Optional for iOS) Install XCode on your machine. If you are not on a Mac, you can develop on Chrome and skip this step. Note that at this time Chrome UI will look strange, as we have concerted our efforts on iOS.
- Setup a Firebase project, which comes with free credits.
- Add an iOS and/or web project. Copy your variables and create a
.envfile under the project root folder - (Strongly recommended) We recommend using Flutterfire CLI to assist your setup. Be sure to run through this end to end to connect your Flutter project to Firebase.
- You will need to ensure Firestore, Firebase Auth, Google Sign in, Firebase Storage are enabled. Google each of these if there is any issue. Google sign in will not be handled by Flutterfire CLI, but enabling it is trivial.
- You will need NodeJS on your machine. We are on version 20, but 22 should also work.
- Setup an an OpenAI account, which comes with free credits.
- (Optional for advanced X scraping) Setup an X account that is Premium, (paid, can skip for simpler development).
- (Optional for advanced news API) Setup an account with Newsapi.ai. This comes with free credits.
- (Optional for maps support) Enable Static Maps API in your new Google Cloud Project, which got created when you created a Firebase project. From there, you will need to setup your Maps API. Setup
dark_mode Static – Rasterandlight_mode Static – Rastermaps. For each map, add a style that matches the colors in Parliament's palette, by making the water equal the background color, and the surface equal theprimaryColorfrom the theme. All borders are set to the onPrimaryColor.
Try running the backend and frontend in different windows.
Firebase may prompt you to add env vars, but if it doesn't, or if you run into billing or other account setup issues, Googling them should be straightforward to solve. Feel free to add helpful info to this readme.
How do I know if I am set up? Once you run the project correctly, the frontend will appear in either iOS or web. You must be able to sign in with Google. In the top right, click the 'Plus' button, and copy this link https://x.com/Eagles/status/1835413000569704460 and click the paste area. The page should spin and load for a few moments. Afterwards press "generate". From here, navigate back to the home screen, pull to refresh, and if you see a Story you are set up!
For any questions connect in our Discord.