Artificial intelligence that can study human language patterns

The model can automatically learn higher-level linguistic patterns that apply to many other languages

A new machine-learning algorithm independently finds linguistic norms that frequently coincide with those developed by human experts.

Due to the remarkable complexity of human languages, linguists have long believed that a machine could not be trained to evaluate speech sounds and word patterns in the same manner as human investigators do.

Yet researchers from MIT, Cornell, and McGill Universities have already made strides in this direction. They have demonstrated an AI system's capacity to teach itself the phonological and grammatical rules of a human language.

This machine-learning model develops rules that illustrate why the forms of those words vary when given words and instances of how those words change to communicate distinct grammatical functions in one language. For instance, it might discover that in Serbo-Croatian, the feminine form of a word requires the addition of the letter "a" at the end.

To get better outcomes, this ML model can also automatically learn higher-level linguistic patterns that apply to many other languages. The model was able to produce the right set of rules to describe those word-form alterations for 60% of the problems.

This method could be used to investigate linguistic hypotheses and discover subtle differences in word meanings between several languages. It is particularly special because the system learns models using little bits of data, like a few dozen words, that are easily understood by people. Additionally, the system makes use of numerous tiny datasets rather than a single large one. This is closer to how researchers propose hypotheses, which is to look at numerous related datasets and develop models to address phenomena across those datasets.

The researchers chose to investigate the relationship between phonology and morphology in their endeavor to create an AI system that could automatically train a model from numerous related datasets.

Since many languages share similar core characteristics and textbook exercises highlight certain linguistic phenomena, data from linguistics textbooks made for an excellent testbed. College students are also highly capable of handling textbook issues, although they frequently draw on past knowledge of phonology from earlier courses when considering new challenges.

The researchers utilized a machine-learning method called Bayesian Program Learning to create a system that could learn grammar or a set of rules for putting words together. With this strategy, the model solves an issue by building a software program.

The software in this instance is the grammar that the model believes to be the most plausible means of explaining the words and their meanings in a linguistics problem. They created the model using Sketch, a well-known software synthesizer created by Solar-Lezama at MIT.

Additionally, they created the model to teach it the characteristics of "excellent" programs. For instance, because the two languages are similar, it might learn some general rules from solving straightforward Russian problems that it would then use to solve a more challenging Polish problem. This makes the model's solution to the Polish problem simpler.

When the system was put to the test using 70 textbook problems, it found a grammar that accurately matched the majority of the word-form modifications in 79 percent of the issues and the whole set of words in 60% of the cases.

The researchers then attempted to pre-program the model with some information that it "should" have learned if it had been enrolled in a linguistics school, and they demonstrated that it could handle all problems more effectively.

The model frequently offered original solutions. In one case, it found the proper response to a Polish language problem that took advantage of a textbook error in addition to the expected response.

The model was also put to the test to see if it could learn some generic phonological rule patterns that could be applied to all problems.

In the future, the researchers intend to use this idea to address unforeseen problems in a variety of different domains. They could also use the method in more circumstances where it is possible to apply advanced knowledge across related datasets. They might create a system, for instance, to deduce differential solutions from datasets on the motion of various objects.