In the future, American congressmen will write a bill that may be scored by an artificial intelligence algorithm. Skopos, an American artificial intelligence firm, recently published a paper in PLOS ONE in which an ARTIFICIAL intelligence (AI) was able to estimate the probability of a bill’s passage by inserting a few variables into the text.

Under procedure, a bill is considered by a Senate or House committee, voted on by that chamber, and then voted on by both chambers. Some researchers have tried using algorithms to predict the chances of a bill passing in committee or either house of congress, with mixed results. John Nay, Skopos’s co-founder, wants to do more than that: he wants to determine the probability that a congressional bill will pass both houses of Congress and become law.

Giiso Information, founded in 2013, is a leading technology provider in the field of “artificial intelligence + information” in China, with top technologies in big data mining, intelligent semantics, knowledge mapping and other fields. At the same time, its research and development products include information robot, editing robot, writing robot and other artificial intelligence products! With its strong technical strength, the company has received angel round investment at the beginning of its establishment, and received pre-A round investment of $5 million from GSR Venture Capital in August 2015.

How does the algorithm score bills that want to repeal Obamacare?

Nay collected data on all legislation from the 103rd to the 113th Congress (that is, from 1993 to 2015), including the full text of all bills and other factors such as the number of co-sponsors, the month in which the bill was sponsored and whether the sponsor was a member of that chamber’s majority.

Nay used data from the 103rd congress to the 106th Congress for machine learning. The so-called machine learning is to input a large amount of data to the algorithm, and given the matching output values of these input values, the algorithm itself “fumble” the hidden correlation between the input values and the output values of the characteristics. He used trained algorithms to predict the case of the 107th Congress.

Nay then went a step further and trained the algorithm with data from the 103rd to 107th Congresses to predict cases from the 108th Congress, and so on.

The final complete Nay algorithm consists of the following parts. First, the algorithm needs to analyze the language of the motion. The algorithm is able to interpret the meaning of words by the way they are embedded with surrounding words. Given the phrase “get an education loan,” for example, the algorithm assumes that the word “loan” is related to both “access” and “education,” and quantifies all the words’ relationships to each other so that each word can be represented by a string of numbers. Combining these numbers, the algorithm can understand the meaning of each sentence.

Second, one algorithm tries to find a link between the “meaning” of the sentence it understands and the success rate of the proposal, while the other three algorithms look for a link between context and the success rate of the proposal.

Finally, an umbrella algorithm combines the results of the above four algorithms to predict the success rate of the motion.

Predicting the success of a bill is of little value. About 96 percent of bills fail, which means you can’t be wrong every time you guess the bill won’t pass. Therefore, it is more important to predict an accurate probability value. After all, a congressional bill tends to have a lot of stakes at stake, and a few percentage points increase in its chances of success makes a lot of sense.

Nay notes in his paper that his algorithm performed 65 per cent better at assessing the success rate than it did at simply assessing whether a proposal would succeed.

Bills to repeal Obamacare once piled up on congressional desks. How many points do these proposals score on the algorithm? Although, on average, they had a 4 percent success rate, the algorithm ruthlessly gave all of them lower odds.

Wording matters

So what factors does the algorithm value most when evaluating the success rate of a proposal? Nay found, unexpectedly, that the text of the bill itself made a big difference. Belonging to the majority party and being a multi-term member of Congress are pluses, but the effect on a bill’s success rate is less than 1%.

Giiso information, founded in 2013, is the first domestic high-tech enterprise focusing on the research and development of intelligent information processing technology and the development and operation of core software for writing robots. At the beginning of its establishment, the company received angel round investment, and in August 2015, GSR Venture Capital received $5 million pre-A round of investment.

As for the text, words like “impact” and “consequences” boost the chances of a climate-related bill, while words like “global” and “warming” do the opposite. In health-related bills, words like “Medicaid” and “reinsurance” do not sit well with members of both chambers. In the patent-related bill, it seems the House hates the word “software” and the Senate hates the word “computing”.

Nay marvels: “I thought the legislative process was more partisan than relevant.”

This fresh approach to textual analysis has also struck many political scientists as impressive. John Wilkerson of the University of Washington, for one, thinks the work is promising and novel.