If you missed the in-person event, we have a virtual booth that contains more detailed information and additional papers/articles.
The Long Night of the Sciences has become an established form of public relations activity in Germany. One night of the year, large scientific institutions hold lectures and demonstrations for the general public in order to present themselves and a general overview of their research topics. The events cover all aspects of science, from natural sciences to social sciences. The public can usually visit the insitutions around 5pm and 1am on the ‘Long Night of Sciences’ for free.
The Lab in Action
The SLaM Lab at NDW - The role of Linguistics in AI for Social Good
How can a computer detect Parkinson’s Disease from a voice recording? Or determine if a person has had too much to drink? What do we need to build voice assistants that can understand diverse voice? With the growing role of language technologies in our lives, we need to build tools that are both intelligent and equitable, in order to meet diverse needs of linguistically-diverse populations. Guests will learn about the intersection between linguistics and AI through hands-on demonstartions. What linguistic principles underlie these challenges, and how can these be computationally implemented? Such systems range from familiar virtual assistants, to non-invasive screening for neuromuscular disorders, to particular populations. How can we improve these systems for the public good?
Virtual Health Assistants: Usage and Problems
Virtual health assistants are virtual people we can have conversations with. They can help by asking guiding questions or giving instructions, which makes them particularly useful for people who feel shame or embarrassment asking other people for help. Among other things, this is useful for preventive talks or screenings, for breast cancer, testicular cancer or for urine samples.
AIs of this kind have to be trained with language data, so that they can recognize the ‘patterns’ of a language. In a best-case scenario, the data is carefully selected, although often, language data from majority languages is used. This way the AI only recognizes a standard variety of a language and it cannot understand certain dialects and minority languages. These people thus do not have the same opportunity and access to virtual health assistants, and other applications using automated speech recognition. To combat the discrimination of minority languages, we ought to collect more language data on these languages and make them more accessible to researches working on AI and related tech.
Alcohol and Pitch in Tone vs. Non-Tone Languages
In tone languages it is essential to have control over your pitch in order to distinguish sounds and with that, different words. In a language like Mandarin for example, it is crucial to be able to produce and perceive up to four contrasting tones that can distinguish the meaning of words. Considering that alcohol affects pitch control, it is interesting to see if speakers of tone languages maintain better control than non-tone languages and how this difference is translated to a second-language contrast. Chang and Tang (2022) investigate these questions with three groups of bilingual speakers: Mandarin L1 and English L2, Korean L1 and English L2 and German L1 and English L2. An experiment with a drunk and a sober condition showed that the native Mandarin group had the best control over their pitch in their L1 and L2, whilst intoxicated. The native German speakers, so not a tone language, had the highest degree of pitch variation under the influence, while the native Korean speakers experienced less impact, a possible result of the fact that Korean speakers benefit from their first language’s contrast, i.e. phoneme categories since they can apply this sensitivity to their second language even when speech is impaired.
Linguistic Biomarkers for Parkinson Disease
Biomarkers are abnormalities concerning genes, hormones, DNA and cells, and can indicate the onset or predisposition to a certain disease. In the case of Parkinsons disease, the basal ganglia, a specific region of the brain, which is responsible for motor control, is affected. As a result, speech production and articulatory processes can be one of the first aspects to be compromised. With a loss of motor control and articulatory precision, speech can be slurred and slowed down, consonants become less clear and the voice might be breathy and less loud overall.
One parameter for measuring these sypmtoms is the degree of palate-tongue contact in sounds. This change of quality in consonants, in this case due to the decrease of contact and tongue mobility, is called lenition. Electropalatography is a useful tool for measuring lenition, since the method is non-invasive and can capture minute deviations and early stages of PD. Statistical models and computer programs on the other hand, can help to quantify and analyse the obtained datasets fast and efficiently. All in all, this approach to linguistic Parkinson disease biomarkers can offer valuable insights that help to recognize speech abnormalities early on.
Want more information? Click here to get to the virtual booth.