What comprehension problems do language assistance systems face?

“Sorry, I didn’t understand him.” Anyone who gets this answer from their own voice assistant will roll their eyes at the latest attempt — and sometimes reach for the screen themselves. To avoid misunderstandings – or lack of understanding – voice assistants go through testing cycles designed to uncover errors and refine the system. Years after the first devices, the technology behind Siri, Alexa, Google, and company has progressed, but training systems despite artificial intelligence is still a challenge for companies and testing providers.

Advertisement

When it comes to development and further development, it is not only different languages ​​and regional differences that must be taken into account, explains Jan Wolter, Product Head and Managing Director of Applause EU, a company that, among other things Apart from this, it also conducts tests. Language support systems before and after market launch. There are many factors that decide whether my device understands me or not. Wolter says that differences in language exist not only within a country, depending on regions, but also across broader factors. “This includes age, gender, social background: people speak very differently.”

Amazon’s Prime Video will get more ads in 2025

language develops

Language habits continue to change even after a generation. Different words are used in different age groups, for example in youth language. Young people sometimes give new meanings to old words and add new words. The constant evolution of language habits means that tools have to cope with a larger data set.

Pronunciation or mispronunciation sometimes drastically changes the sound of a word. “The system has to expect that a word can be mispronounced,” says Wolter, citing brand names as an example. These are sometimes pronounced very differently – it is not for nothing that there are many lists on the Internet that deal with exactly this topic.

Context helps in classification

To solve this problem, the system has to understand the context, says Walter: “The system has to understand the context and recognize what language or dialect it is.” Not all linguistic phenomena can be solved using synonyms. For example, many people in West Germany think of something completely different when they hear the word “pancake” than Berliners do. Experienced users can often work well with systems because they provide them with lots of context. But this cannot be assumed. “We work with a testing community of about 1.5 million people around the world, to whom we give a goal, but not a path,” says Walter. “Whether a system understands a given command does not help in everyday life if users have their own way of achieving the goal.” Furthermore, language is always changing. “You can’t finish something and leave it there for ten years.”

The diversity of spoken languages ​​poses challenges to development. Due to rapid growth and special diversity, sometimes data sets are required that are difficult to obtain. This also applies to the area of ​​inclusion. If the user stutters or has Tourette’s, the system should be able to handle it. “The main problem here is gathering good data that can feed the system,” says Walter. This is extremely difficult because online surveys often reach the same people. “It is even more important for our customers that the topic comes into greater focus.”

Wolter sees a particular need for development in ensuring that language assistants not only understand the language well, but also adapt to the individual – and, if necessary, adopt language habits. “I don’t know whether every language assistant has to speak the youth language, but action is still needed to ensure that they actively recognize that the output should be in simple language.”

(Are)