SymmetricalVoice @Symmetricalvoic

Linguistics and language conservation. Special attention to the Pacific. run by @peterschuelke Joined October 2019

Tweets

2K
Followers

864
Following

1K
Likes

4K

Gašper Beguš @begusgasper

2 years ago

Can we model syntax from speech? Most models of syntax are text-based. Here we propose that basic syntax can be modeled from raw speech. GANs trained on individual words start to concatenate them into multiple-word outputs. Sometimes the model even concatenates three words:

2 28 175 38K 73

Download Image

Adobe Garamond Balenciaga @bkmacaulay

2 years ago

theta roles are just a 'gotcha' for syntacticians, like it's always "what do you mean 'subject'? this verb only assigns a SCORPIO"

0 1 10 926 1

Gašper Beguš @begusgasper

2 years ago

Interpretability is one of the main frontiers in AI Doing interpretability work on speech is easier than in vision. Among others, sound is 1D (huge advantage) Here's a presentation of our IEEE/ACM TASLP paper that introduces a technique to interpret layers in audio CNNs

1 7 30 3K 7

Download Video

Endangered Languages Project @_ELProject

2 years ago

We're hiring! Applications for ELP's summer 2023 internships are now open - if you are passionate about supporting #languagerevitalization, we'd love to hear from you!⁠ ⁠ Learn more about the positions and apply by May 10 at: bit.ly/ELP-interns2023

4 62 89 21K 28

Download Image

Gašper Beguš @begusgasper

2 years ago

GPT4 unlike ChatGPT does not hallucinate papers and actually gives useful recommendations. Prompt: “Give me references to papers that measure vowel duration before stops with various laryngeal features”

0 1 16 2K 2

Download Image

Field Matters @field_matters

2 years ago

The link is in document! Join us at "Climbing the WALS: Typological Case Studies in LLMs Probing Using Syntactic Probing Framework"online talk now!

Field Matters @field_matters

2 years ago

The link is in document! Join us at "Climbing the WALS: Typological Case Studies in LLMs Probing Using Syntactic Probing Framework"online talk now!

0 0 0 256 1

0 1 2 186 0

@langoinstitute.bsky.social @langoinstitute

2 years ago

Join us as we learn #Ghanongga #Ganoqa youtu.be/bK5NqUX8Bfk

0 2 4 249 1

Field Matters @field_matters

2 years ago

1/7 Discussing low-resource languages we highlighted that the amount of the language data doesn’t correlate with the number of their speakers. The number of speakers isn’t the main factor that affects language vitality, the other factors also play a part. Which ones? The 🧵

Field Matters @field_matters

3 years ago

1 0 2 850 0

1 2 6 661 2

Gašper Beguš @begusgasper

2 years ago

Uncovering the unknowns in sperm whale communication🐳 using deep learning We introduce a novel interpretability technique (CDEV) that combines generative deep models and causal inference methods. Our model discovers meaningful properties, both novel and previously hypothesized

15 117 497 167K 168

Download Image

IZ Europäische Sprachen @izeus_berlin

3 years ago

#OTD 146 years ago, Wilhelm Doegen (1877-1967) was born 🥳 Expert on phonetics, phonology, and the inventor of a speech recording device known as Doegen-Lautapparat (1909). During WW1, he recorded speech samples and songs of PoWs in diverse languages. #LinguisticBirthdays #Histlx

1 8 18 4K 0

Download Image

Greg Dickson @GregDicksonNT

3 years ago

Animals in Kriol where the etymology is the English plural: hosis = horse (from 'horses') ens = ant (from 'ants') I reckon there are a couple more examples that will come to me...

0 2 7 230 0

Gašper Beguš @begusgasper

3 years ago

"John is too stubborn to talk to" is the new "colorless green ideas sleep furiously" Prompt: Draw a minimalist syntactic analysis of the sentence "John is too stubborn to talk to" GPT4:

3 6 31 5K 5

Download Image

Gašper Beguš @begusgasper

3 years ago

GPT4 Prompt: draw an OT Tableau for the process /anta/ > [anda]

Gašper Beguš @begusgasper

3 years ago

GPT4 Prompt: draw an OT Tableau for the process /anta/ > [anda] https://t.co/joYAqIfmVl

15 63 603 181K 143

Download Image

1 1 9 3K 1

Download Image

Ryan @Ryncor2

3 years ago

Not a huge Chomsky fan myself, but can we please stop with the LLMs disprove Chomsky thing? Humans are not computers. We don't know that computers learn grammar in the same way, we don't know that humans can learn like computers, and LLMs require significantly more data.

4 1 19 3K 0

Gašper Beguš @begusgasper

3 years ago

GPT4 draws a syntactic tree for the sentence: "I saw an elephant with glasses" 🐘😎 Interesting timing given a recent opinion piece in @nytimes

15 63 603 181K 143

Download Image

Ansley Potter @ansdpotter

3 years ago

Here's a look at our poster from #HSP2023! Adults who spend many hours reading may read object relative clauses faster, BUT this effect is modulated by the syntactic complexity of the text (websites) they frequently read

2 2 22 2K 1

Download Image

Gašper Beguš @begusgasper

3 years ago

One advantage of speech-based models trained on smaller datasets is that we can observe how the network learns as it is being trained (unlike in text-based GPT) and compare “machine acquisition” to language acquisition stages They can be are very similar youtu.be/oClrwnh0DLc