tlhIngan-Hol Archive: Mon Aug 10 10:01:09 2015

Back to archive top level

To this year's listing

[Date Prev][Date Next][Thread Prev][Thread Next]

Re: [Tlhingan-hol] Bing progress

Robyn Stewart ( [KLI Member] [Hol po'wI']

I've added hundreds of sentences and their negations. I think they are not
quite honest with me when the say that it is capable of learning about
infixes, or that they have fixed the problem with it ignoring whitespace.
It still stumbles at the distinction between leghbe' and legh be'. I think
its underlying templates work best on SVO languages, so while it generally
uses OVS, it can be thrown off.

Here are some sentences:
DaHjaj legh be'. wa'Hu' leghbe'. puq Qorgh SoS. tugh yaS naDHa' Sa'. ghorgh
be'Hom bIj vav? qaStaHvIS po Qu' Qatlh wIta'. jIHeghpa' 'eSpanya' vISuch
vIneH. nIQ 'ey bovutmeH Suy' pI' qemnIS wIjwI'ra'. moHaq rurDI' mu' pojHa'
Bing. mughwI' laHmey vI'aghmeH mu'tlheghmey Sar vIqon 'e' vInID.

The public version translates as:
The blind today. Yesterday I was blind. Care of the child mother. the primal
officer soon General. When a boy that she sends you to your father? It is
difficult in the morning we Do. before I die, I want to visit Spain.
Breakfast is tasty, I must bring to bovut, my fat Shooy beast. bing pojha'
words, when like a prefix. translator's skills to vi'ag amid the Varied

The currently deployed in-development version produces:
The woman sees today. the woman sees the yesterday. the mother takes care of
the child. the officer soon, the General is primal. When the father punishes
the girl? during the mission, we are difficult in the morning. before I die
I want to visit Spain. bovut is delicious breakfast brought to my fat pig
nibbles, you need to. When the words are like the prefix bing misanalyses.
translator to demonstate skills I tried to write a variety of sentences.

Likewise translating this to Klingon:
I don't know your name. You are afraid to attack the government. Before the
meeting, let's eat lunch. Is there a doctor on the ship? While the woman is
deciding, we can kill the blue bird. You can't manage a Klingon crew.  We
don't wash our car during the dry months.  I didn't know my mother very
well. She was very strong and intelligent. They can't speak French. The
children don't want the slimy fruit. He wears expensive trousers and cheap
shoes. The red book is about an advanced civilization. The admiral is angry.
The sergeant is not qualified. An airplane doesn't resemble a cloud.

ponglIj'e' vISovbe'. bInarghtaHvIS qum HIvmeH. pa' qep, Ha' megh vISop.
tu'lu' Qel DujDaq? wa'vatlh wuq be', cha'par SuD laH HoH maH. tlhIngan beq
pagh vu' SoH. puH Duj ghewmey DaSay'moHtaHvIS maH during jar QaD. SoSwI'
vIchel vISov maj. ghaH HoS 'ej val. french pagh jatlh chaH. naH char ghewmey
neH puq. yavDaq qamDu'Daj lIS wagh qutlh waq 'ej tuQ ghaH. paq petaQ
vIHtaHbogh 'Itlh tayqeq. 'aj QeH. bu' 'oHbe' 'um. chen doesn't rur muD Duj.

In development:
ponglIj vISovbe'. bIHeghvIpchugh qum luHIv. qaSpa' qep, Ha' megh Sop. SaH
Qel DujDaq? wa'vatlh wuq be', bo'Degh SuD HoHlaH maH. tlhIngan beq pagh vu'
SoH. puH Duj ghewmey Say'moH maH qaStaHvIS jar QaD. SoSwI' vIchel vISov maj.
ghaH HoS, 'ej val. vIraS jatlhlaHbe' chaH. char naH vIneHbe'mo' 'e' puq.
wagh qemtaHbogh qutlh waqmey je 'e' tuQ. Doq paq bop 'Itlh tayqeq. QeH 'aj.
bu' 'um wej. chen jatlhbe' rur muD.

Because Bing uses a statistical approach,  throwing a large number of
sentences at it to fix one problem can break something that was previously
working. The in development version no longer uses Hamlet, which eliminates
a lot of weirdness. I'm not sure when the last update of the public version
was; it still doesn't include the nerd vocabulary, which I incorporated
about a month before the Spoken Nerd DVD came out, but they crazier errors
aren't as common now.

And wow, right there you can see an error in the input: I must have
misspelled demonstrate at some point and it picked up on it.  I'll have to
find that.  Likewise I can try to track down the "primal" thing.  Going from
Klingon to English, it creates a rough translation, and then tries to make
that match patterns that it knows about in English, which explains why it
sometimes veers off so badly.  The engine has learned to omit the word "is"
when translating from English to Klingon, so that probably helps to explain
its like/is like confusion and also--I just thought of this, thanks--its
inability to learn the difference between paq Doq and Doq paq.  I threw
thousands of machine generated sentences using almost all the adjectival
verbs in the corpus with a wide variety of nouns and the before and after
tests of its ability to get the adjectives in the right place showed as many
shifts from correct to incorrect as the reverse.

The Hub *is* complicated, and they keep updating it, so I can't even give
one road map for occasional users to use.  I don't think they realize what a
barrier it is for people to use.  It's just as useful to me to get feedback
from the public version, because I can try to make sure the next one won't
do that.

- Qov

-----Original Message-----
From: Lieven [] 
Sent: August 9, 2015 23:08
To: Robyn Stewart
Subject: Re: [Tlhingan-hol] Bing progress

Am 07.08.2015 um 20:57 schrieb Robyn Stewart:
> Haha, great.

> Oh Bing, how many negated sentences will you have to parse before you get

I suppose you have added tons of simple negated phrases, right?

 > It can't seem to get the hang of parHa' vs. rur,  > and just when I think
it has OVS it messes it up again.

Yes, I guess the problem is withing the english, where the word "like" 
has multiple meanings. How do they handle that in general?
On the other hand, evene a computer should noticed that when {rur} is used,
an english phrase usually has a verb attached, "is like". Hm.

> Are these from the public version or the in-progress that you have to log
in to use?

These are from the public version, I forgot how to log in to the translation
hub. I found it too complicated at the beginning, so I lost interest in
working there. Sorry.

Lieven L. Litaer
aka Quvar valer 'utlh
Grammarian of the KLI

Tlhingan-hol mailing list

Back to archive top level