A warkin product wi research an development aye gaun on
My Words exists acause o a haarder problem nor owersettin a sentence: haudin a conversation thegither ower a leid boundary fan fit's at stake is somebody's hoosin, safety or health, an deein it in a wye that tells the truth aboot its ain uncertainty. That problem isna sortit yet, sae the research didna stop fan the product shippit.
Far it cam fae
The wark startit as Project Keystone, research by INT6 Ltd intae real-time bilingual conversation mediation usin language models. INT6 biggit the proof-of-concept application, Turnstone, that wis syne wirkit up intae a warkin system an testit in real settins.
The application his sinsyne been shiftit intae C2 Discovery Labs CIC sae that it's hauden for community an societal benefit. My Words is that same testit system, rebrandit an cairried forrit, an the development gyangs on unner it.
Academic validation
Independent validation is bein arranged wi a UK university partner. The ettle is tae wark wi postgraduate students an academic research labs on twa things at aince: makkin the product better, an checkin that it dis fit this site says it dis.
That wark isna feenished an naething on this site hings on it. Fan there are results, includin ony that gyang agin us, they'll be published here wi the method attached. Till then, tak the claims on this site as oors raither nor as independently checkit.
Fit the project is really seekin tae find oot
These are the questions the wark is organised roon, an the eens we wid pit in front o a research partner the morn. Nane o them is settled.
Can a machine haud context the wye an interpreter dis?
A human interpreter cairries the room, the history an the register wioot bein speirt. Foo muckle o that can be made explicit as structured context, an fit's tint in the makkin o it explicit?
Can a system usefully ken fan it's wrang?
Fidelity scorin is ae model judgin anither. Far dis that correlate wi real error, an far dis it gyang wrang in the same direction as the verra thing it's checkin?
Dis a question for clarity help, or jist interrupt?
Speirin beats guessin, in principle. Wi a distressed body in front o ye, foo mony times can ye interrupt afore ye've connached the verra thing ye were protectin?
Foo shid uncertainty be shawn tae a thrang professional?
A marker that's ignored is nae better nor nae marker ava. A marker that stops the conversation ilka turn is waur. This is an interface research question as muckle as a modellin een.
Fit chynges fan the body hauds the device themsel?
Readin in yer ain haans, in yer ain leid, is a different thing fae readin across a desk. Fither that chynges fit fowk will tell ye is a question for measurement, an we hinna measured it yet.
Far dis the boundary really sit?
We assert that this is for early conversations an nae for statutory eens. That boundary shid be testit agin fit really happens in services, nae defendit jist acause we set it doon in writin.
Foo the wark is evaluated
The system produces its ain evidence as a by-product o rinnin: ilka turn cairries a source, an owersettin, a fidelity score, ony revision, an ony question speirt for clarity. That maks a hantle o things measurable wioot instrumentin onything extra.
- Foo aften an owersettin is revised afore delivery, an foo faur the score shifts.
- Foo aften the mediator speirs insteid o answerin, an fit the conversation dis neist.
- Foo the scores spreid oot by leid, sae that uneven coverage shaws up insteid o disappearin intae an average.
- Far warkers owerrode the tool, said the same thing ower again, or gied up on the exchange aathegither.
Fit the instrumentation canna tell us is fither the conversation wis ony gweed. That needs the service an, far it can be deen richt, the body on the ither side o it.
Foo we haud oorsels tae it
Test afore live. New behaviour is pit throwe its paces agin recordit an synthetic conversations afore it iver meets onybody's real een.
Measure fit wis measured. Nae headline accuracy figure staunin in for a distribution we can actually shaw.
Publish the method. Fit we learn aboot deein this safely is written up for ither fowk tae reuse, includin the bits that didna wark.
Negative results coont. A feature that made conversations waur is a findin, an gets reportit as een.
Fit's biggit an rinnin
This is a warkin application, nae a demonstration. In braid terms it his the noo:
Live turn-by-turn mediation
Streamin owersettin wi the conversation history, structured facts an domain context in scope for ilka turn.
Fidelity scorin an self-revision
Scores for ilka turn, revision afore delivery fan the score is unner the threshold, an clarification back tae the spikker far the source is ambiguous.
Twa conversation surfaces
A shared-screen mode an a paired mode far the saicont body joins fae their ain device ower a live connection.
A localised interface
Interface strings owerset intae the seedit leids an cached, sae the surface itsel isna English-only.
Accoonts, groups an audit
Groupin an permissions at organisation level, passkey an password authentication wi twa-factor codes, an an audit log ower conversations an administration.
Conversations that can be replayed
Recordit exchanges can be played ower again agin a new build, sae a chynge is testit on real dialogue an nae on somethin made up for the demo.
An ordinar web stack
A PHP an MariaDB application served ower HTTPS. Deliberately plain, sae it can be reasoned aboot, auditit, an hostit some ither place nor far it sits the day if that turns oot tae be the richt answer.
Wark on this wi us
We're leukin for postgraduate students an research labs wi a question in this space, for services willin tae hae their use studied, an for funders backin evidence-led wark on leid as a barrier tae gettin help.