How do artificial-intelligence systems process language

In several recent papers, researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and the Qatar Computing Research Institute have used a recently developed interpretive technique, which had been applied in other areas, to analyse neural networks trained to do machine translation and speech recognition.

They find empirical support for some common intuitions about how the networks probably work. For example, the systems seem to concentrate on lower-level tasks, such as sound recognition or part-of-speech recognition, before moving on to higher-level tasks, such as transcription or semantic interpretation.

But the researchers also find a surprising omission in the type of data the translation network considers, and they show that correcting that omission improves the network’s performance. The improvement is modest, but it points toward the possibility that analysis of neural networks could help improve the accuracy of artificial intelligence systems.

“In machine translation, historically, there was sort of a pyramid with different layers,” says Jim Glass, a CSAIL senior research scientist who worked on the project with Yonatan Belinkov, an MIT graduate student in electrical engineering and computer science.

“At the lowest level there was the word, the surface forms, and the top of the pyramid was some kind of interlingual representation, and you’d have different layers where you were doing syntax, semantics".

"This was a very abstract notion, but the idea was the higher up you went in the pyramid, the easier it would be to translate to a new language, and then you’d go down again. So part of what Yonatan is doing is trying to figure out what aspects of this notion are being encoded in the network.”

The work on machine translation was presented recently in two papers at the International Joint Conference on Natural Language Processing. On one, Belinkov is first author, and Glass is senior author, and on the other, Belinkov is a co-author.

On both, they’re joined by researchers from the Qatar Computing Research Institute (QCRI), including Lluís Màrquez, Hassan Sajjad, Nadir Durrani, Fahim Dalvi, and Stephan Vogel. Belinkov and Glass are sole authors on the paper analysing speech recognition systems, which Belinkov presented at the Neural Information Processing Symposium last week.

Neural nets are so named because they roughly approximate the structure of the human brain. Typically, they’re arranged into layers, and each layer consists of many simple processing units — nodes — each of which is connected to several nodes in the layers above and below.

Data are fed into the lowest layer, whose nodes process it and pass it to the next layer. The connections between layers have different “weights,” which determine how much the output of any one node figures into the calculation performed by the next.

During training, the weights between nodes are constantly readjusted. After the network is trained, its creators can determine the weights of all the connections, but with thousands or even millions of nodes, and even more connections between them, deducing what algorithm those weights encode is nigh impossible.

The MIT and QCRI researchers’ technique consists of taking a trained network and using the output of each of its layers, in response to individual training examples, to train another neural network to perform a particular task. This enables them to determine what task each layer is optimised for.

In the case of the speech recognition network, Belinkov and Glass used individual layers’ outputs to train a system to identify “phones,” distinct phonetic units particular to a spoken language. The “t” sounds in the words “tea,” “tree,” and “but,” for instance, might be classified as separate phones, but a speech recognition system has to transcribe all of them using the letter “t.”

And indeed, Belinkov and Glass found that lower levels of the network were better at recognising phones than higher levels, where, presumably, the distinction is less important.

Similarly, in an earlier paper, presented last summer at the Annual Meeting of the Association for Computational Linguistics, Glass, Belinkov, and their QCRI colleagues showed that the lower levels of a machine-translation network were particularly good at recognising parts of speech and morphology — features such as tense, number, and conjugation.

Discover more here.