How NASA’s Forgotten Search Engine for Moon Rocks Boosted AI

Despite the ubiquity of AI assistants like Siri or Alexa, teaching machines how to understand and wield human language remains a huge stumbling block in the field of artificial intelligence.

This field of research—called "natural language processing" (NLP) and a core component of a machine passing the Turing test—got a big leg up in the 80s, when computer scientists began to turn to machine learning to help computers 'understand' human-speak. Now, AI uses even more sophisticated algorithms called neural networks that mimic the structure of the human brain to understand human language when it is encountered in the wild.

In the 1970s, researchers had to program computers with a list of words and their meanings, as well as the rules of English grammar. The computer simply obeyed these hand-written rules. Now, with deep learning, we only have to give the computer examples of human language and let it figure out how it works. But NLP wouldn't have gotten to where it is today without a conceptual boost from NASA, which spent two years in the early 70s developing a natural language processing system to help its geologists classify moonrocks.

NASA's system, called the Lunar Sciences Natural Language Information System (LSNLIS), was conceived as a "stepping stone on the path of discovery that may someday make computers as generally available and conveniently accessible as one's next-door neighbor." It was one of a handful of NLP projects being undertaken at the time (such as MIT's ELIZA, the psychoanalyst chatbot), but despite its hyper-specific implementation, LSNLIS pioneered the natural language search processes that would make things like Google search or AI voice assistants possible decades later.

But in 1972, NASA had no intention of making its question-answering system available to the general public. Rather, it was an experiment in natural language processing that would allow lunar geologists to query a database to compare and evaluate chemical analyses from lunar samples without having to worry much about how to phrase complex search requests to make them understandable by the machine. Rather, the geologists should just be able to ask the computer a question about lunar sample data ("What is the average concentration of aluminum in glass?") and the machine would be able to process their request—sort of like a Google for moon rocks.

"We are not interested in questions which require evaluation, judgment, or conclusions on the part of the system (e.g. 'Does the moon have a hot core?')," the researchers wrote in their final report. "It is the task of the scientist to interpret the data, and we are trying to aid him in this task—not replace him."

Read More: Meet the First Chatbot Sent Into Outer Space

The researchers programmed their system in the LISP, the second oldest programming language, which would query a large database of chemical information about lunar samples collected during the Apollo missions. Unlike today's NLP systems, which are able to analyze natural speech without being programmed to recognize certain patterns, NASA's system was pre-programmed with a corpus of possible queries and some 3500 vocabulary words provided by geologists who were familiar with the lunar sample data.

According to the final report on the LSNLIS, the system consisted of three main components. First, a question was fed to a parser that analyzes the syntax (word order) of a user's query. After analyzing the structure of the query, this is then fed to a semantic analysis component. Together these two components of the system work to translate the user's question into a machine-readable request, which is then executed by the system to produce an answer to the query based on what information is available in the database.

After a year of development, the system was tested at second Lunar Science Conference in 1971. Geologists were allowed to run searches on the database by typing questions in natural English and of the 111 questions asked, the machine was able to interpret 78 percent of them correctly and return the relevant data. Not bad for a first test run, but its success might have been partly due to geologists knowing how to ask the right questions—the system had a much more difficult time answering questions supplied by users who were not geologists.

In any case, NASA's lunar analysis system marked a major step forward in natural language search processing. Its unprecedented ability to parse and translate natural language syntax and semantics into machine-readable code laid the foundation for much more sophisticated search engines like Google, and ultimately the AI that will one day pass the Turing test thanks to its uncanny ability to understand and imitate natural speech.

Subscribe to pluspluspodcast , Motherboard's new show about the people and machines that are building our future.