Language learning balances clarity and effort

For many years linguists have been debating the underlying reasons for the structural and grammatical similarities between apparently unrelated human languages. While some suggest these typological threads lead back to common linguistic ancestors, others have proposed that they are merely accidental.

A study reported in the journal Proceedings of the National Academy of Sciences (PNAS) this week supports the third idea, widely subscribed to, that the common traits identifiable across languages are a consequence of the universals of human cognition.

The authors of the paper, Maryia Fedzechkina, a doctoral candidate, and T Florian Jaeger, Wilmot Assistant Professor of the Sciences at the University of Rochester, and Elissa Newport, Professor of Neurology and Director of the Center for Brain Plasticity and Recovery at Georgetown University, gave ScienceOmega.com more information about their research and its interesting results.

The claim made by some linguists that languages are shaped by learning biases as well as pressures pertaining to communicative use has long remained controversial and somewhat difficult to test empirically for two principal reasons.

"First, although there are mathematical theories of communication that tell us what properties efficient communication systems should have, these theories were rarely applied to linguistic research," pointed out Professor Jaeger.

"Second, the traditional approach to the study of cross-linguistic generalisations faces serious challenges due to the cost of language documentation and the fact many languages are historically and geographically related – languages are often in contact over long periods of time and they share historical roots, which means that they do not provide independent data points."

A recent investigation published in Nature apparently found no statistically significant typological evidence for language universals, purporting to reject decades of linguistic work that had claimed the contrary. Although the work has come in for criticism from many quarters, it is symptomatic of the difficulties that modern typological work faces according to the authors of this new study.

Professor Jaeger’s lab have focused in previous work on mathematical principles of efficient information transfer in the context of speech, finding many properties that strike an efficient balance between effective and fast communication in conversation. Conversational speech is full of reductions (saying a word faster and with less clear pronunciation) and contraction (eg, President Clinton didn’t rather than President Clinton did not). Jaeger’s work suggests that these reductions are not arbitrary – they tend to occur when the context already provides information about the reduced material.

"Although these reductions are often considered ‘informal’, they are an essential part of language. We don’t usually notice it, but not only our pronunciations, but also our word choices are in large parts determined by what’s already expected in the current context," said Professor Jaeger. "We strike a pretty efficient balance between the chance of being understood and the effort we put into our productions."

"We were interested in how such properties enter languages," Professor Newport explained. "My colleagues and I had previously developed the artificial language learning paradigm and used it in experiments to study learning biases that operate during language acquisition, with both child and adult learners."

This work demonstrated the learner’s tendency to change a language as they learn it to make it more regular. It had also provided evidence that learning behaviour in the lab resulted in patterns that resembled those from languages around the world. Fedzechkina saw the potential in uniting the two strands of research to investigate whether the same preference for efficient information transfer observed affecting language production in adults also operates during language acquisition.

The team invented two miniature artificial languages in order to observe the process of language acquisition. The 40 student participants, none of whom spoke a language other than English, were asked to learn 15 nouns, eight verbs, and the grammatical structure of these languages over the course of four, 45 minute sessions. Short animations, computer images and audio recordings made up the lessons.

Compared with fieldwork and language documentation – methods typically used for the study of language universals – the technique of creating an artificial language for experimental use has several advantages, as Professor Newport outlined:

"Researchers using an artificial language paradigm need to be careful to create languages that are very natural and language-like, and need to use software to create auditory and visual stimuli such as videos or pictures to accompany the auditory language presentation. But the advantage is that we can then control carefully the input and the structure of the language to be learned."

Artificial languages afford researchers the opportunity to investigate a variety of phenomena. Although they are constrained by the complexity of the language since it needs to be acquired in a short time in the laboratory, the only other limit is their personal creativity.

"The ease with which participants learn an artificial language in the lab depends on its complexity," commented Professor Newport. "Some smaller and less complex languages can be learned within one session of as little as two minutes length. Other more complex examples involving a lot of variation, like ours, take several sessions distributed over several days. Most of our participants showed a good understanding of the language after two days, though they continued to learn for the entire four days."

The case markers included to modify nouns depending on whether they are the subject or object of a sentence are common to Russian, Japanese and other languages, but are not generally prominent in modern English. While the case markers in Old English allowed for relatively free word order, gradual changes in pronunciation began to make meaning ambiguous. This resulted in word order becoming more important and hence more rigid. This is an historical example of the adaptation of language by users, apparently to ensure clear and concise communication.

"We found that the bias towards efficient communication systems plays a role during language acquisition and forces the learners to restructure the input language they’re learning, making it more communicatively efficient," Fedzechkina asserted. "Our findings suggest that communicative pressures can shape language structures already as learners acquire them."

In both experiments, the language-learning participants adapted the rules when faced with a sentence construction which was potentially ambiguous or confusing. When the meaning could have been obscured by not using a case marker the learners were more likely to put these markers to use, distinguishing object from subject.

In the real world, languages undergo change all the time through language contact, random errors, and lexical innovation to name just a few.

"Not all changes get picked up by the next generation and persist in the language," stated Fedzechkina. "Some changes actually disappear fairly quickly or never make it from small subgroups of the language community to broad acceptance by most speakers."

She pointed out that the biases learners possess may act as a filter as the language is passed from one generation to the next.

"Learners preferentially pick up those changes that are beneficial for efficient information transfer and make them more pronounced in the language," Professor Jaeger said. "In a sense, language learners make a language better suited for communication. Our work suggests that maybe not all change that is often labelled as ‘corruption’ or ‘degradation’ is bad. Instead, many reductions, such as ‘math’ for ‘mathematics’, ‘the web’ for ‘the world wide web’, ‘don’t’ for ‘do not’ and so on, often make language more efficient."

The results provide support for the view that language acquisition is affected by language function.

"Some researchers in linguistics don’t consider communication an important aspect in explaining why languages are structured the way they are, while others have challenged that," Professor Jaeger attested. "We suggest that communication is important enough to affect the development of language. Specifically, learners seem to prefer language codes (grammars) that strike an efficient balance between clarity and effort.

"This preference might in turn reduce to an even more general ‘bias’ that is inherent to our brain, a preference to – on average – reduce the expected effort required to achieve our goals."

Although experiments of this type are comparatively easy to conduct, the recruitment process, scheduling and administering of participants can be a challenge for studies which take multiple days to learn. In Professor Newport’s lab, these studies are run with young child learners, who often show even stronger tendencies to change the language as they learn it. In Professor Jaeger’s lab, collaborations between Fedzechkina and Dr Harry Tily, a speech scientist at Nuance Technologies, are seeing the extension of the artificial language paradigm to a web setting via crowd-sourcing.

"This comes with its own unique set of challenges but holds the promise of obtaining large amounts of data from learners with many different language backgrounds," Professor Jaeger enthused.

I asked the team about the next stages of their research.

"As a next step, we are interested in understanding the developmental timeline of the bias toward efficient linguistic systems we observed in these experiments," Fedzechkina told me. "How early does it develop? Does one need substantial experience with language to show this preference for efficient linguistic systems, or does it develop very early in life before sufficient language experience is acquired?

"This will tell us whether the preferences we observe in our experiments are innately given or whether they are themselves learned. So far we have only tested adult learners so our findings do not speak to these questions. Our next step is to explore the preferences of young children in similar tasks."

In a broader context, the researchers say their findings contribute to recent studies showing that the artificial language learning paradigm is suitable to test linguistic theories of language universals and change, which could benefit future studies in many ways. Professor Newport and collaborator Jenny Culbertson at George Mason University have investigated word order universals that become more consistent through language learning.

"As yet we have only tested one of many theories of language evolution – the preference for efficient information transfer. There are many more theories that have been proposed to explain linguistic phenomena that the artificial language learning paradigm will allow us to explore," concluded Professor Newport.