Researchers teach computer to be fluent in Finnish dialects

Computers usually recognize Finnish lone arsenic the normative modular known arsenic kirjakieli. Finnish dialects, however, make a batch of occupation erstwhile interacting with computers, since it is intolerable to talk a connection without speaking successful a dialect of immoderate sort. A probe radical has built artificial quality (AI) models that tin automatically detect, normalize and make Finnish dialects. The results were published in The 2021 Conference connected Empirical Methods successful Natural Language Processing.

Collecting information for making an AI recognize dialectal Finnish and Swedish has been connected the quality recently. The methods devised by the probe radical of Mika Hämäläinen, Niko Partanen, Khalid Alnajjar and Jack Rueter from the University of Helsinki instrumentality this further and alteration an AI to beryllium fluent successful the Finnish dialects.

Within the paradigm of computational creativity, they person developed a method for converting modular Finnish into 1 of the 23 Finnish subdialects. Computers should not lone beryllium capable to recognize dialectal Finnish, but they should besides beryllium capable to explicit themselves successful a dialect.

"With our method, an intelligent strategy specified arsenic a robot tin accidental akku connected lopussa (battery is low), for illustration successful Etelä-Karjala dialect akku o lopussa, Etelä-Satakunta dialect akku ol lopus oregon Länsi-Uusimaa dialect akku o lopus," Hämäläinen says.

For example, the commonly utilized algorithm of Google Translate fails to construe a dialectal Finnish condemnation Oisko sulla jotai esimerkkei siit (Do you hap to person immoderate examples of that) producing a wholly incorrect "English" translation Oisko sulla thing similar that conscionable due to the fact that Google Translate has been built to enactment exclusively connected modular Finnish. This aforesaid improvement tin beryllium observed with immoderate AI tools that enactment Finnish similar Apple Siri oregon dictation successful macOS.

Dialects are detected from some spoken audio and text

The probe shows that detecting dialects is simply a hard task erstwhile relying connected plain text. Dialect recognition is easier erstwhile the exemplary has entree to audio arsenic good due to the fact that galore dialects are marked with distinctive phonetic properties. Thus the latest probe published by the researchers deals with detecting dialects from some spoken audio and text.

"The process of normalizing dialects to modular substance has galore benefits. It allows analyzing dialectal materials utilizing tools for the Standard Finnish, and we tin besides usage the normalized mentation arsenic a hunt point erstwhile we privation to find thing from the dialectal materials", says Khalid Alnajjar.

The researchers punctual that the occupation of knowing dialects is analyzable and nary exemplary tin recognize earthy connection similar humans do. But the created models unfastened galore much absorbing directions for research, specified arsenic the grade to which a dialect deviates from the norm and what are the syntactic differences betwixt antithetic connection varieties.

"With this we tin amended the existent authorities of Finnish natural language processing solutions and physique AI models tailored for individuals. For example, we person already reached awesome results successful speech recognition of 1 person's speech, adjacent successful endangered languages", Niko Partanen says

The probe radical has besides developed a akin normalization methodology for the dialects of Swedish spoken successful Finland (Hämäläinen et al., 2020b) and humanities Finnish (Hämäläinen et al., 2021b).

The dialect generator tin beryllium tested online and the dialect normalizer and generator codification person been released openly connected Github. The dialect recognition codification tin beryllium recovered connected Github arsenic well.

---

---