Talk:Welsh to English

Note: Comments should not include '=' as it confuses the Wiki templating system (as I just found out myself)

Note 2: Suggestions for part-of-speech disambiguation should go here.

OK, I'll try, but I'm not entirely sure of the distinction. some of the stuff at the end of that page, for instance, is covered here. - Donnek

Note 3: Comments should not include the '|' symbol either, at least within double quotes, since it too confuses the wiki.

English to Welsh

Macros

This will contain chunks of rules that we need to split out to make them more maintainable

Notes for areas to be covered

A sort of scratchpad / todo list, based on things that come up when putting phrases into the testing webform.

Conjunctive genitive

gwallt yr eneth - *hair the girl - the hair of the girl - the girl's hair

llaw y bachgen - *hand the boy - the hand of the boy - the boy's hand

Note that the noun phrase in English is definite - contrast "merch y meddyg" (the doctor's daughter) and "merch meddyg" (a doctor's daughter).

For an English phrase 
of the type "def + noun1 + of + def + noun2"
or of the type "def + noun2 + 's + noun1"
convert in Welsh to "noun1 + def + noun2".

Here can noun1 be a simple noun, or can it be a noun phrase? For example "the red cat of the young boy" - Francis Tyers

e.g.

For the pattern det.def + noun1 + of + det.def + noun2:

Output noun1 + det.def + noun2

Yes, as long as you like, eg,

cath goch bachgen bach merch ifanc bert rheolwr y banc mawr du

the red cat of the little boy of the pretty young daughter of the manager of the big black bank

It's only the last NP of the sequence that gets the def.det. Donnek

Ok, so this requires a three level rule.

t1x -> t2x SN_(the cat red) of_(of) SN_(the boy little) of_(of) SN_(the daughter young pretty) of_(of) SN_(the manager) of_(of) SN_(the bank big black)

t2x -> t3x SN_(the cat red) SN_(the boy little) SN_(the daughter young pretty) SN_(the manager) SN_(the bank big black)

t3x -> gen (cat red boy little daughter young pretty manager the bank big black)

What I'll do for now is get the chunks working ('SN' -- noun phrase, and 'of'), for values of 'noun', 'det noun', 'det adj noun', 'det adj adj noun', 'det adj adj adj noun', etc. Then look at taking care of more frequent cases (e.g. the first example). Francis Tyers

For a Welsh phrase of the type "!det + noun1 + def + noun2"
convert in English to "def + noun1 + of + def + noun2"
or to "def + noun2 + 's + noun1".

The second noun is probably historically a genitive, but it has lost all case markers. The equivalent in Irish would be:

ceann an chapaill - *head the of-horse (gen) - the head of the horse - the horse's head

ceann capaill - *head of-horse (gen) - the head of a horse - a horse's head

Welsh to English

A couple of things have come up:

Using a NOUN1 + DET.DEF + NOUN2, we get "the daughter of the doctor", but:

Ond mae fy nhad, mam a thad fy ngŵr wedi talu trethi ac erioed wedi defnyddio'r gwasanaeth iechyd. → But My father is, mother that and my father man after pay taxes and never after use the service health.

If we use NOUN1 + DET + NOUN2, we get "the daughter of the doctor", and:

Ond mae fy nhad, mam a thad fy ngŵr wedi talu trethi ac erioed wedi defnyddio'r gwasanaeth iechyd. → But My father is, mother that and my father of my man after pay taxes and never after use the service health.

Is NOUN1 + DET.POS + NOUN2 reliably going to be DET.POS + NOUN2 + 's + NOUN1 (or DET.DEF + NOUN1 + of + DET.POS + NOUN2)?

Yes. - Donnek

Actually, this is trickier, I guess:

fy nhad, mam a thad fy ngŵr → my father, mother and my husband's father

We'd need to consider "DET NOUN COMMA NOUN CNJCOO NOUN DET NOUN" or at least "DET NOUN COMMA NOUN" → "my father and mother" + "NOUN DET NOUN" → "my husband's father", which would give:

my father and mother and my husband's father

Which although awkward, isn't as horrific as the current translation. I'd prefer to make small chunks where possible, even if (for now) it makes a worse translation, as when we are able to collapse NP CNJCOO NP → NP, in an intermediate stage, it will be more clean.

I think that reads well, actually. I realise there is a tension between doing something that will work for common examples, but is too much of a hack for less common examples, and doing something more comprehensive but which will take longer to finalise. All other things being equal, I would tend towards the former for 0.1, but only you can say what is the best way forward, since you know how Apertium is put together. You've probably struck a decent balance here. - Donnek

Ok, the output is now:

But my father, mother and the father of my man is after pay taxes and never after use the service health.

Which is reasonably ok... The way I have done it is:

In level-1, made three chunks:

• SN(my father , mother)
• CC(and)
• SN(the father of my man)

Then at level-2, I detect:

• SV SN CC SN and transform to: SN CC SN SV.

It doesn't seem to have caused any regressions, but might warrant further testing. - Francis Tyers

"was"

"roedd" ([he/she/it] was) is unknown, but I seem to remember adding entries for "to be" to the dixes in the mists of time. Was I dreaming? (roedd <- yr + oedd)

There are entries for 'bod', but 'roedd' doesn't get processed as all of the 'bod' entries start with 'b' (see this link). I will need to fix this in the analyser. If I understand you correctly, 'roedd' is a contraction of 'yr' (determiner ...) + 'oedd' (verb 'bod', past tense ...)? Francis Tyers

Some serious errors have crept in to those entries. I've sent an amended version to you by email. You're right - roedd -> yr + oedd, but in the amended version I've sent, I've put (e.g.) "roedd" and "oedd" as alternate forms, because "Roedd" is the spoken form, and even in written Welsh you hardly ever see "Yr oedd" nowadays. Donnek

The "bod" paradigm should now be all ok, there remains however to choose the restrictions (e.g. which forms we will generate for each set of tags). - Francis Tyers

the boy was in the garden -> *y bachgen bu yn yr ardd - bu'r bachgen yn yr ardd

Almost correct, except for word-order, and the fact that the preterite is being used instead of the imperfect ("roedd y bachgen yn yr ardd"). The preterite needs to be marked as only being used in written Welsh, and to have a lower likelihood than the imperfect. This is too rough a rule, but would do for the time being.

Marking and word-order

The above brings up a useful point about this. If the standard VSO sequence is changed to SVO (ie unchanged from the English standard), this is a marked pattern, conveying a relative clause. In written Welsh, the verb will be preceded by "a" + soft mutation, but in spoken Welsh the "a" usually disappears.

y bachgen [a] fu yn yr ardd ddydd Llun (the boy who was in the garden on Monday)

yr eneth [a] welodd y ci (the girl who saw the dog)

contrast

gwelodd yr eneth y ci (the girl saw the dog)

Hmmm. Relative clauses are going to be difficult.

For Welsh pattern "noun + a + soft-mutated_verb"
output English pattern "noun + who/which + verb".

The dictionary only has 'a' down as a co-ordinating conjunction "and", does it have other meanings? - Francis Tyers

Yes. "a" - relative "who, which" in a relative clause where the subject is the same as that of the main clause, and "a" - interrogative pre-verbal particle (eg a weles ti hwnna? - did you see that?). Both are followed by soft mutation. Note that interrogative "a" is usually omitted in speech, leaving only the mutation. - Donnek

"yn" as stative

For Welsh pattern "yn + adj"
output English "adj"

There is a problem here in that this pattern can also be an adverb:

siaradodd yn hapus am ei fywyd - he talked happily about his life

For English pattern "adverb_formed_from_adj + ly"
output Welsh "yn + adj"

This second one will be difficult to do, as we don't have adverbs in the English dictionary marked as derivatives from adjectives or not. - Francis Tyers

OK. Unfortunately, since "yn + adj" can be either an adj or an adv in Welsh, I don't even mark them separately in Eurfa - perhaps I should. Would one option be to replicate all the Welsh adj entries in Apertium by preceding them with "yn + space", and adding "-ly" to the English side? This would get the EW direction, but I don't know whether it would cause problems on the WE direction. - Donnek

The above rule has been applied (way!), but does not catch mutated adjectives ("yn" causes soft mutation):

*tyfodd fo yn mawr -> he grew big

tyfodd fo yn fawr -> *he grew in *fawr

This was a dictionary error, 'fawr' did not have the initial-m paradigm. Now added. - Francis Tyers

OK - there are a couple of others I've come across: mwy (fwy), bach (fach), gwyn (wyn). there may be a few more. - Donnek

Taken care of the first two, 'gwyn' doesn't seem to appear in the dictionary (only as 'complaint'). does it inflect at all? - Francis Tyers

LOL! There are some obvious words not in Eurfa, tut tut to me! gwyn (white), *gwen (in practice "wen", fem), gwynion (occasionally, plural), gwynnach (whiter), gwynnaf (whitest). There may be fem comp and super forms too, but we can ignore those. By the way, "da" also has this problem too. - Donnek

 :) -- Ok, I've added gwyn/gwnnach/gwynnaf for now, adding the genders would probably mess up some rules and these are probably fairly low frequency and can be taken care of later. - Francis Tyers

We could also extend this to nouns:

roedd hi'n waith anodd -> *was in ~a #difficult<adj><sint> work - (it) was hard work

(though "work" gets lost in the second proc run).

For Welsh pattern "yn + non-place noun-phrase"
output English "noun-phrase"

This is a bit complex. There are two "yn"s in Welsh: "yn" showing state or condition, or extension in time (yn hapus - happy; yn mynd - going), and "yn" the preposition showing location in a specific place (yn y tŷ - in the house (contrast: mewn tŷ - in a house); yn Nolgellau - in Dolgellau). (They are probably related historically.) The stative "yn" soft-mutates nouns and adjectives, but not verbs; the location "yn" nasal-mutates (and changes to "ym" to match an initial "m" in the following noun, eg ym Mangor - in Bangor).

So - as it stands, the above will clash with 1.3.10 (change prep+noun to prep+det.indef+noun), even though "yn" the preposition will never occur before a non-specific noun (it must have specificity), and even though the above is not actually "yn" the preposition (it's "yn" the stative). We can't use the stative soft-mutation to decide, because (a) that doesn't apply to some consonant initials, and (b) other prepositions cause mutation too, and it would be overkill to check for each one. So the easiest thing is to adjust 1.3.10 to exclude "yn" as one of the prepositions that will be caught. Is it easy? I don't know :-)

I've added the yn "stative" to the analyser as well as the yn "preposition", but until we retrain the tagger it will not pick the former. If you could think of any rules that will choose the right one in a given context it would help (for ideas on the kinds of restrictions to these rules, see here and here). - Francis Tyers

The simplest would be:

Welsh word "yn" is a preposition

when it is followed by "det.def" or by a capitalised word

otherwise it is a stative

That may not be perfect, but it is good enough. I'll bear in mind the tagger pages, but it may take a while to get to that stage. - Donnek

Preferential choice between noun and verbform

atebodd hi'r cwestiwn -> *answered shethe #hold an inquiry - she answered the question

proc selects 'cwestiwn' (question) - correct - and 1p pl imperative of 'cwestio' (an infrequent verb for 'hold an inquiry'). The 1p pl present would also have been a possibility, and indeed a more likely one. tagger selects the second of these.

Not sure how widespread this would be, but the tagger should give precedence to the noun choice whenever the verb form is preceded by 'y':

For Welsh pattern "{y,yr,'r} + word_tagged_as_either_noun_or_verb"
output "{y,yr,'r} + noun"

This is not perfect, because "y | yr" can also be an indirect relative clause pronoun before a verb, but it would catch most things until we can resolve the latter point.

gwelodd y dyn y llyfr -> *the man saw the books - the man saw the book

This is similar, but is tricksy because it is superficially correct apart from the plural. But in fact, tagger is reading "llyfr" as pres 3p sing of "llyfru" (to book). Apart from being infrequent, and therefore much less likely to appear ("bwcio" would be the usual word), Eurfa has "llyfra" as the pres 3p sing, so there may be a paradigm problem too. The above rule would throw out the verb in the meantime.

It is currently using the aberth/u__vblex paradigm (see output here). Is this incorrect? - Francis Tyers

The problem is that "aberthu", apart from the 'regular' "abertha" also has a written "aberth". So yes, it probably is incorrect. The problem is that a lot of less common verbs are very rarely inflected. It might have been better to use something like "gwenu" or "siomi". In the meantime, perhaps just changing "aberth" to "abertha" in the pres 3p sing will do. - Donnek

Number agreement of verb

I added 'rabbits' to the dictionary, but the problem of unknown words and phrase movement is one we're experiencing in Basque too... - Francis Tyers

OK - so it's basically an issue that you can't do much about until the word is logged. Hmm. I suppose that makes sense, since Apertium can't figure out what to do with something until it knows what it should do with it ... In a practical sense, this is going to be problematic if we demo Apertium using unseen text. Is there any way of doing some blind choosing, eg

if this word is

preceded by [y,yr,'r]

we will assume it's a noun

preceded by yn

we will assume it's a verb

unless a verb has been identified in the current phrase

in which case we'll assume it's an adjective

This might break Apertium - I don't know. In theory, though, we might be able to get relative probabilities for a particular sequences from a corpus. - Donnek

I'd be reluctant to add one as we'd not be able to get the translation, on the other hand, it wouldn't cause messing up of word order. It's an open problem, and we're thinking about it :) - Francis Tyers

Prepositional noun phrase should not be a subject

cerddodd fo i'r dref -> he walked in the town

Fine, except that the preposition "i" should really be glossed as "to" ("yn y dref" would be "in the town")

Contrast:

cerddodd i'r dref -> *the town walked in - [he/she] walked to the town

Welsh pattern "prep + det.def + noun" is never a subject phrase

and therefore the "det.def + noun" section shouldn't be shifted. (I can't think of any exceptions to this, but there may be one.)

There was a rule to do this, I've commented it out, I think there was a reason for it, but I can't recall now. I've run the regression tests below and it doesn't seem to have broken anything. Regarding the preposition, should I change "i" to be "to" instead of "in" ? - Francis Tyers

Re "i", yes, change it to "to". - Donnek

The problem here was the dictionary only had i'r → yn+yr... i've added i'r → i+yr and now it is picking the right one, although I don't know what will happen for other contexts... - Francis Tyers

Not sure where that would have come from. The only vaguely relevant thing I can think of is "i mewn i" (into). - Donnek

allan i'r cyfarfod -> *the meeting #exit<vblex><pres><p3> in - out to the meeting

This is similar - "in" should be "to", and should be kept with "the meeting".

However, there is another issue here, which is in effect the same as "Preferential choice between noun and verbform" above. In this case, the verb "allanu" (to exit) is being chosen instead of the much more likely "allan" (out).

roedd o ar dy lyfr -> *was of on your books - it was on your book

1.3.9 would deal with "of", and 1.3.6 would deal with "books". Subject shift would then produce a reasonable translation.

However:

roedd ar dy lyfr -> *your #be<vbser><past><p3> on books - (it) was on your book

Omitting the subject pronoun can happen quite frequently in speech if the subject has already been mentioned. The <sg> tag gets lost at interchunk, which means the verb can't be conjugated (this came up somewhere else, but I think it's been taken off the page - maybe it would be better just to mark the issue heading as "addressed" rather than delete it). But there is an additional issue, in that the possessive pronoun is getting treated as the subject and moved separately. So maybe we need a broader rule to say that "prep + det.def/pr.poss/whatever + noun" is an indivisible chunk, and must be dealt with as a block. No part of it would be moved in this case anyway.

Regarding page cleanup, ok. perhaps having a separate section, and then moving sections down would be a good idea. - Francis Tyers

It would also be nice in the longer term to fill in the pronoun if it is omitted.

For Welsh pattern "verb + non-subject noun phrase"
output English "verb + pronoun agreeing in number and person + non-subject noun phrase"

The NSNP could be a prepositional phrase (marked by an initial preposition), or an object phrase (marked with initial soft mutation).

"-ing" as "yn + verb"

For English pattern "subject + verb<vbser> + verb + ing"
output for Welsh "verb<vbser> + subject + yn + verb"

Inflected verbs not being parsed

aeth -> *aeth - (he/she/it) went

However, "aeth" is listed in cy.dix.xml (line 27491) as past 3p sing in the mynd_vblex paradigm, which is what "mynd" (to go) gets conjugated against (line 54444).

Ah - a bug in the segmentation.

*myndaeth fo -> he went

he went -> *myndaeth fe

The infinitive is getting added to the irregular forms, instead of being replaced by them.

Yep, this is a problem in the paradigm for 'mynd', I'll need to rewrite it, fortunately it is only used once... New paradigm output here - Francis Tyers

Fine, but the imperative forms also need "mynd" excised. - Donnek

Done. - Francis Tyers

Insert det.indef in prepositional NP

daeth Taid â lamp -> Grandfather came with lamp (preferably "with a lamp")

dychwelodd y rheolwr gyda gŵr tew -> the manager returned with fat man (preferably "with a fat man")

For Welsh pattern "prep + noun"
output English pattern "prep + det.indef + noun"

this should probably be working now, - Francis Tyers

I am tempted to retire this in favour of a broader rule:

For Welsh pattern "non-specific noun.sg"
output English "a + non-specific noun.sg"

"Non-specific" here means a noun that is not qualified by det.def, pr.poss, etc.

daeth car ar hyd y ffordd -> *car came along the road - a car ...

mae'r athro yn licio chwarae gêm o golff -> *the teacher is liking play game he golf - the teacher likes to play a game of golf

1.3.16 would deal with "play". For "gêm o golff" we would need to prevent "*a game of a golf" (which the existing rule would in fact have produced for "o golff"). Perhaps:

For Welsh pattern "non-specific noun + o + non-specific noun"
output English pattern "a + noun + of + noun"

This would need to fire before the revised rule above, or we need some other way of sorting out the possible doubling of "a" (a a game of golf) - LOL - let it happen and then have a rule:

For English pattern "det.indef + det.indef"
output "det.indef"

gwelodd y bachgen gath yn yr ardd -> the boy saw cat in the garden (preferably "a cat")

Preferential choice between verbforms

bydd y lamp yn rhoi golau -> *are the lamp giving light - the lamp will be giving light (and presumably we could massage this into "the lamp will give light" later, since that would be the more natural English equivalent)

A couple of things here. The most important is that tagger chooses the less frequent imperative out of the imperative/future choice for the verb. Presumably this then means that the subject shift can't take place. But even with the imperative choice, the imperative 2p sing info gets lost between interchunk and postchunk, and replaced with a generic? present which gets output as "are". Odd.

(I'm assuming that "bydd" would get output as "will be", since that would be the correct English tense.)

I fixed this (crudely) by commenting out the imperative for "be" 2pSg (You are!) When I train the tagger next I'll see if i can take care of it there. - Francis Tyers

Another example:

roedd y bechgyn yn gallu croesi dan y ffordd -> *the boys were #be<vbmod><ger># able to #vblex><vblex><pres> under the road - the boys were able to cross under the road

Here, proc decides to ignore "croesi" (cy-en.dix line 6203) as an infinitive in favour of conjugating it as present 2p sing. The infinitive option doesn't even show! Can we add some such rule as:

When you have homographous Welsh verb options <inf> and conjugated_verb
choose <inf> unless verb is followed by pr

This isn't very good, because you could have the conjugated from without a pronoun, but it might deal with this to some extent.

The example above also has some funkiness going on with "gallu" - "were be able" needs to be transformed into "were able". However, I don't know enough about how Apertium treats modal verbs to make a suggestion.

This is now giving:

the boys were being able to cross under the road

Which is an improvement. On the other hand, "^be<vbser><past><p3><pl>$ ^be able to<vbmod><ger>$" is probably redundant :) -- Not sure how to deal with this. - Francis Tyers

What about a purely surface hack, eg:

For English pattern "was/were + being + able"

output "was/were + able"

Or a mapping to convert "be able" to "can/could"? - Donnek

Comparative adjectives with "less/more"

tyfodd y twnnel yn llai llachar -> *the tunnel grew small bright - the tunnel grew less bright

tyfodd y twnnel yn fwy tywyll -> *the tunnel grew big dark - the tunnel grew more dark

For Welsh pattern "fwy/llai + adj"
output English "more/less + adj"

For English pattern "more/less + adj"
output Welsh "fwy/llai + adj"

I'll see if i can copy in a rule from Spanish--English for this :) - Francis Tyers

Synthetic comparative adjectives

Many of these seem to have faulty dictionary entries:

tyfodd y twnnel yn fwy -> the tunnel grew big (should be "bigger")

tyfodd y twnnel yn llai -> the tunnel grew small (should be "smaller")

tyfodd y twnnel yn hirach -> the tunnel grew long (should be "longer")

tyfodd y twnnel yn uwch -> the tunnel grew high (should be "higher")

Dictionary error in the bidix, now fixed. - Francis Tyers

Cool! - Donnek

Note, also a rule needs to be written for:

the expensive house → y tŷ drud

the more expensive house → y tŷ drutach

the most expensive house → y tŷ drutaf

Verb + preposition

Re "coolness factor" below (woop woop!), we need to cater for verbs such as "ymchwilio" which are followed by a preposition that is different from English, or where there is no preposition in English.

For example:

ymchwilio i - research into, investigate

siarad am - talk about

dweud wrth - say to, tell

gofyn am - ask for

Is there any way to get the verb+prep phrase parsed as a phrase, rather than separately? Perhaps an entry in one of the dictionaries? This would only need to be done for those phrases where the preposition differs in English and Welsh.

Not, for instance for:

neidio dros - jump over

cerdded i - walk to

delio gyda - deal with

where there is a regular correlation between the meanings of the Welsh and English prepositions.

Yes, these are multiword constructions, like for example "He became accustomed to the taste." → "cynefinodd Fe i y blas." (try it in the testing interface). Is there a way of getting a list of these? (actually there are many I currently need to fix in the bidix/English dict, but if you have a list I can look at them. At the moment we only seem to have multiword verbs on the English side. - Francis Tyers

I will try to compile a list of the most common, and send it to you tomorrow. - Donnek

Subordinate ("reported speech") clauses with "bod" + noun

Also referring to the cool sentence, we have two sentences as follows:

(1) roedd y Comisiwn yn ymchwilio i'r honiadau - the Commission was investigating the allegations

(2) mae yr AS wedi methu datgan £103,000 o roddion - the MP has failed to declare £103,000 of gifts

Subordinate clauses, like the relative clauses, will be difficult. But a first stab at this might be as follows:

For Welsh pattern "[b/f]od + [det.def] + noun + [qualifiers] + wedi + verb"
output English "that + [det.def] + noun + [qualifiers] + has/have (number agreeing with noun) + verb_past_participle"

clywodd y dyn bod y trên wedi cyrraedd yn hwyr -> *the man heard be the train after arrive #late

This would be:

VBSER(INF) + DEFINITE_NP + wedi + VERB

THAT + DEFINITE_NP + HAS + VERBPP

Where DEFINITE_NP is any noun phrase preceded by the definite article? - Francis Tyers

Actually, thinking about this again, it doesn't have to be definite - it just so happened that in those sentences it was. You could have something like:

clywodd ysbïwr bod y trên .... -> a spy heard that the train ....

So the NP could be "[det.def, rhyw (some), pr.poss] + [adj - eg hen] + noun + [qualifiers - adjectives, demonstratives, etc]", or it could just be "pr.subj" (clywodd fo bod y trên ...). The same applies to the "am" construction below. Another point is the the VBSER can be soft-mutated - "fod" instead of "bod". - Donnek

This rule is broadly working for now. At least it is inserting the 'that', a form of 'have' and changing the verb to a pp. It is not however robust, and seems to me a bit hacky. Could you give some more examples so I can fine tune it? - Francis Tyers

Hacky? Surely not .... I did say that relative and subordinate clauses will be difficult, so we may have to refactor as we go along. An alternative to the above (which would also cover the adjective example below) would be:

For Welsh pattern "[b/f]od + NP + complement"

output English "that + NP + is + complement"

This would give:

clywodd y dyn bod y trên wedi cyrraedd yn hwyr -> *the man heard that the train is after arrive #late

You would then have further rules to transform "is + after + verb" to "has + verbpp", and "is + for + verb" to "will + verb". (Irish and Gaidhlig have a similar construction, by the way, using "ar, air" instead of "wedi", so whatever rule bundle you use here would be transferable to that branch of Celtic too.)

There is also another similar construction using "ar" in place of "wedi" and "am" - this one means "about to":

clywodd y dyn bod y trên ar gyrraedd yn hwyr -> *the man heard be the train on arrive late - the man heard that the train was about to arrive late

So an additional rule "is + on + verb -> was + about to + arrive".

Oh, there's another one too, with "newydd", meaning "just now":

clywodd y dyn bod y trên newydd gyrraedd yn hwyr -> *the man heard be the new train arrive late - the man heard that the train had just arrived late

This one has been caught by the adjective rule, but "newydd" belongs to the VP, not the NP in this case, so we'd need some prioritisation.

Other examples:

roedd y bachgen yn dweud bod y tŷ wedi mynd ar werth -> the boy was saying that the house has gone on value

(fine, except that "ar werth" means "on sale" - see 1.3.19 below.

dywedodd hi bod y trên yn hwyr -> *she said be the train late - she said that the train is late.

You could deal with this one by adding a similar rule:

VBSER(INF) + NP + ADJ

THAT + NP + VBSER + ADJ

dwi'n meddwl bod y glaw wedi stopio -> *dwithinking that the rain has stopped - I think that ....

(We need to get the present tense of "bod" sorted out too)

dywedodd yr eneth bod y siop am agor ar amser -> the girl said that the shop will open on #time

(I'm noticing the lack of adverbs, eg "tomorrow", "today", "afterwards", etc. I suppose the remaining bits of Eurfa need importing at some point.)

Donnek

The above rule would give "the man heard that the train has arrived late" - not perfect, since in English we would use pluperfect rather than perfect here, but a lot better.

We can extend this to another construction:

For Welsh pattern "[b/f]od + [det.def] + noun + [qualifiers] + am + verb"
output English "that + [det.def] + noun + [qualifiers] + will + verb"

clywodd y dyn bod y trên am gyrraedd yn hwyr -> *the man heard be the train for arrive #late

The above rule would give "the man heard that the train will arrive late" - not perfect, since in English we would use conditional rather than future here.

This now seems to be working. - Francis Tyers

These could be improved if it were possible to refer back to the verb of the main clause. Thus where it is past, the subordinate would use pluperfect or conditional; where it is non-past, the subordinate would use perfect or future.

We can probably set this as a variable, but what would be the triggers to set/unset the variable? - Francis Tyers

End of the clause? Full stop or comma, perhaps? - Donnek

There are other varieties of subordinate clause that I give other suggestions about.

Incidentally, in the above the det.def should be taken to include other prenominal qualifiers like possessives.

Verbal nouns / Infinitives

roedd y dyn yn gwerthu pethau rhad -> the man was selling cheap things

roedd fo yn palu -> he was digging

Both of these are fine.

Making the verbal noun/infinitive the subject doesn't work quite so well:

roedd gwerthu pethau rhad yn hawdd -> *was sell cheap things easy - selling cheap things was easy

roedd palu yn waith caled -> *was dig in a hard work - digging was hard work

The latter would be benefit from the extension of the "yn as stative" rule to nouns, as suggested in 1.3.4 above. But we also need to define the VN as a subject, so that it can be shifted. This is not easy, because the rule may cause problems with other constructions later. But we can take a stab at it.

First, we can use the infinitival form in English - "to sell cheap things was easy" and "to dig was hard work" are equivalent to the above sentences.

For Welsh pattern "verb<vblex><inf>"
output English "to + verb"

For English pattern "to + verb"
output Welsh "verb<vblex><inf>"

This allows the same rule to be used in sentences like:

ceisiodd y dyn agor y bocs -> *the man sought open the box

which should produce "the man tried to open the box". (Can we delete the "seek" entry for "ceisio" until we have refined choices between different entries? The "try" entry is more frequent.)

"seek" entry commented out and replaced with "try". Francis Tyers

Second, we can assume that a verbal noun phrase will occur after an inflected verb (mostly forms of "bod"). So we might try expanding the above to say:

For Welsh pattern "verb_inflected + verb<vblex><inf> + [noun phrase]"
output English "to + verb + [noun phrase] + verb_inflected"

for the first two sentences ("roedd gwerthu pethau rhad" and "roedd palu"), and:

For Welsh pattern "verb_inflected + subject + verb<vblex><inf> + noun phrase"
output English "subject + verb_inflected + to + verb + noun phrase"

for the third ("ceisiodd y dyn agor y bocs").

This is not perfect, and I am not sure how it would cut across the existing rule for subject shift.

There are also interesting issues with nesting of infinitival subject phrases:

roedd gwerthu pethau rhad yn hawdd yn neis -> *was sell cheap things easy nice - to sell cheap things easily was nice

roedd gwerthu pethau rhad yn hawdd yn beth neis -> *was sell cheap things easy in a nice thing - to sell cheap things easily was a nice thing

where I'm not sure how you specify the boundaries of the noun phrase. Any views, or is that too complex for the present iteration.

At the moment to define noun phrases we just define fixed length patterns of tags which are matched in left-to-right, longest-match way. So for example for Welsh to English:

NOUN → SN_(NOUN)

PRNSUBJ → SN_(PRNSUBJ)

DET → SN_(DET)

DET NOUN → SN_(DET NOUN)

NOUN1 NOUN2 → SN_(NOUN2-NOUN1)

NOUN ADJ → SN_(ADJ NOUN)

yn ADJ → SN_(ADJ)

DET NOUN ADJ → SN_(DET ADJ NOUN)

NOUN ADJ1 ADJ2 → SN_(ADJ2 ADJ1 NOUN)

The first is the pattern we detect in Welsh, and the second is the "chunk" that we output in English. Any suggestions on defining more of these (the most frequently occurring), or changing them would be appreciated. - Francis Tyers

Note also in the above that we have the adverb problem from 1.3.4.

Infinitive after "yn lle"

For Welsh pattern "yn lle + verb<vblex><inf>"
output English "instead of + verb + ing"

For English pattern "instead of + verb + ing"
output Welsh "yn lle + verb<vblex><inf>"

Would you say "yn lle" is a multi-word preposition? - Francis Tyers

Yes, it is a compound preposition. "yn ei le" - instead of him (lit. in his place), "yn eu lle" - instead of them. But I don't want those included here, because there are places where you might want to translate them "in his place". - Donnek

yn lle mynd dros y ffordd -> *instead of go *dros the road - instead of going over the road

Incidentally, why is "dros" coming up as unknown here? I remember sweating over putting it in the dictionary (cy.dix line 46972, cy-en.dix line 49) :-)

It is on there, but under the paradigm /tros__pr, which means that it will never get detected... is 'dros' a mutation of 'tros' or a separate preposition, the same oddness goes for trwy and drwy. - Francis Tyers

Yes, they occur in both mutated and unmutated forms. I would think the mutated forms are more common. Hmm. I didn't realise the paradigms overwrote the cited form like that. In that case, we either need to do some [t,d] substitution, or (perhaps simpler) replicate the entire "tros" paradigm for "dros", replacing the t with d. Same for "trwy, drwy". - Donnek

This seems to be done. - Francis Tyers

"i gyd"

This mean "all", and occurs after the noun:

roedd y cwningod i gyd yn ddiogel -> *the rabbits were I joint safe - all the rabbits were safe

For Welsh pattern "det.def + noun + [qualifiers] + i gyd"
output English "all + det.def + [qualifiers] + noun"

Is 'i gyd' considered an adjective or pronoun? (or something else?) :)- Francis Tyers

An adjective, I suppose. Certainly a qualifier of some sort. - Donnek

Making progress. We now get:

roedd y cwningod i gyd yn ddiogel -> *the all rabbits were safe

Just need to massage that slightly. - Donnek

Non-compositional multiword phrases

This section is for phrases that have to be scoped as a whole, rather than broken down to their constituent parts.

ar werth - on sale

ar ôl y cwbl - after all

erbyn hyn - by now

doed a ddelo - come what may

hyd yn oed - even (adv)

pan fo angen - as needed

wrth gwrs - of course (adv)

yn weddill i - remaining to

Superlative adjective + "oll"

y rhai lleiaf oll -> *the some smallest *oll - the smallest ones of all

yn gyntaf oll -> *first *oll - first of all

For Welsh pattern "adj.super + oll"
output English "adj.super + of all"

We need to add "oll" (all) to the dictionary, but this would still be a necessary rule.

"rhai"

y rhai bach -> *the some small - the small ones

rhai mawr -> *some big - big ones

"rhai" (some) can be considered the plural of "un" (one).

For Welsh pattern "rhai + adj"
output English "adj + ones"

This also applies to phrases like "y rhai lleiaf oll" above. We need to convert "oll" first on the basis that is follows an adj, and then we need to convert "rhai" on the basis that it precedes an adj.

Dictionary errors (refs to cy-en.dix and cy.dix)

"hefyd" (152) is correctly listed as "also", but is wrongly coming up as "then".

Should be fixed - Francis Tyers

"da" (5318), is correctly listed as "good", but is coming up as unknown.

Where is it coming up as unknown? - Francis Tyers

In :"mae'r bachgen yn licio'r eneth sy'n dda". But that is now coming up OK as well. - Donnek

    <pardef n="anghydwel/d__vblex">
    <e lm="anghydweld"><i>anghydwel</i><par n="anghydwel/d__vblex"/></e>
    <e lm="cyfweld"><par n="initial-c"/><i>yfwel</i><par n="anghydwel/d__vblex"/></e>
    <e lm="gweld"><par n="initial-g"/><i>wel</i><par n="anghydwel/d__vblex"/></e>
    <e lm="rhagweld"><par n="initial-rh"/><i>agwel</i><par n="anghydwel/d__vblex"/></e>
    <e lm="ymweld"><i>ymwel</i><par n="anghydwel/d__vblex"/></e>

This paradigm appears to be broken in that some of the <r> sides are of different lengths (they should all be 'd' if /d and 'eld' if /eld) - Francis Tyers

I'm missing something, sorry. To me they are all segmented /d. - Donnek

See for example the entries in the paradigm here. - Francis Tyers

"sydd" / "sy"

This is a relative present form of "bod" - "who/which is/are". The elided form "sy" is more common in speech. "sydd" is not listed in the dictionary, but "sy" is.

mae'r dyn yn adeiladu gwesty sy'n darparu llawer o ystafelloedd -> *the man is building hotel is provide many of rooms

mae'r blaid yn gwneud rhywbeth sy'n cyfrannu at ennill yr etholiad -> *the party is doing something is contribute towards win the election - ... which contributes towards winning the election

mae'r dafad yn pori yn y maes sy'n cynnig bwyd da -> *the sheep is grazing in the field is offer good food - ... which offers good food

(Note: there are some frustrating shortcomings in the output. If we use a variant of the last sentence:

mae'r defaid yn pori yn y cae sy'n cynnig bwyd da -> *the sheep #be<vbser><pres><p3><pl> grazing in the closes is offer good food

it appears that the conversion can't handle the plural of "sheep", and tagger insists on choosing an inflected verb ("closes", from "cau") instead of the noun (cae - field) - 1.3.5 really needs to be implemented.)

For Welsh pattern "{sydd yn, sy'n} + verb_infin"
output English "that + verb.pres.3p.sing"

This should also broadly be fixed. Can you check the output below:

mae'r dyn yn adeiladu gwesty sy'n darparu llawer o ystafelloedd → the man is building hotel that provides a lot of rooms

mae'r blaid yn gwneud rhywbeth sy'n cyfrannu at ennill yr etholiad → the party is doing something that contributes towards win the election

mae'r dafad yn pori yn y maes sy'n cynnig bwyd da → the sheep is grazing in the field that good food offers

mae'r defaid yn pori yn y cae sy'n cynnig bwyd da → the sheep are grazing in the closes that good food offers.

- Francis Tyers

Terrific. The only thing is that the last two sentences have subject shift, even though "bwyd da" is an object. Would it be possible to ban subject shift after "sy[dd]"? Also:

For Welsh pattern "at + vb.infin"

output English "towards + verb +-ing"

Donnek

Done. There are a few regressions because of the new tagger, but I'm looking into them. - Francis Tyers

"llawer"

"llawer o" (a lot of, many) seems to be OK. But another rule would be useful to deal with the third coolness sentence:

For Welsh pattern "llawer + adj.comp"
output English "much + adj.comp"

dyn llawer hŷn -> *older many man - a much older man (see 1.3.10 for the "a")

Done. - Francis Tyers

bod (inf) + subject pronoun + cael (inf)

Should this be mangled into:

be(inf)  + subject pronoun + get(inf)
that     + subject pronoun + had

e.g.

honiadau ei bod hi'n cael perthynas â

his allegations be she getting relation with → his allegations that she had relation with

... also on this subject, we currently don't have a verb for "have" in the bidix, the grammar i have suggests that "cael" might be it in a modal sense. - Francis Tyers

Subordinate ("reported speech") clauses with "bod" + pronoun

The above (1.3.25) may be worth doing, but it may be better to deal with the more general construction.

In effect, this is the same construction as 1.3.15, but with the noun replaced by a pronoun. However, while in English the pronoun is a subject pronoun, in Welsh it is a possessive pronoun. This means that the "that" word ("bod") gets sandwiched by a the two parts of the possessive pronoun (either of which may not appear, depending on style), and is also mutated accordingly.

We're currently calling "ei" and friends possessive determiners, should we change this to possessive pronoun, or does it not make much of a difference? - Francis Tyers

Not much difference I think. Possessive determiner has the benefit that they can be considered to specify the noun in the same way as det.def. - Donnek

Thus, with a noun:

clywodd y dyn bod y trên yn cyrraedd yn hwyr (the man heard that the train was arriving late)

becomes, with a pronoun:

clywodd y dyn ei fod o'n cyrraedd yn hwyr (the man heard that it was arriving late)

For Welsh pattern "[pr.poss] + VBSER.inf_mutated + [pr.subj] + yn + verb"
output English "that + pr.subj + is + verb + -ing"

Hmm, this one is problematic as we throw away the "yn" in stage one transfer in the rule that turns "yn + vblex.inf" → "vblex.ger",

^det<SD><det><pos><sp>{^his<det><pos><2>$}$ ^verbinf<SV><vbser><inf>{^be<vbser><3>$}$ 
^prnsubj<SN><p3><m><sg>{^prpers<prn><subj><2><3><4>$}$ ^verbinf<SV><vblex><ger>{^arrive<vblex><3>$}$

Would it cause any problems if we made this rule:

For Welsh pattern "[pr.poss] + VBSER.inf_mutated + [pr.subj] + verb.ger"
output English "that + pr.subj + is + verb + -ing"

This would give: "the man heard that he is arriving late" (the "he" is an open issue)

- Francis Tyers

Done. - Francis Tyers

I was actually typing in the same suggestion, but you got there first! I don't think you can do much about the "he" without some sort of semantic check, which is not realistic at this stage. the only thing you could do would be to use "he/it", but that looks clumsy. - Donnek

For Welsh pattern "[pr.poss] + VBSER.inf_mutated + [pr.subj] + wedi + verb"
output English "that + pr.subj + have/s + verb.pp"

Done. - Francis Tyers

For Welsh pattern "[pr.poss] + VBSER.inf_mutated + [pr.subj] + am + SM_verb"
output English "that + pr.subj + will + verb"

Done. - Francis Tyers

For Welsh pattern "[pr.poss] + VBSER.inf_mutated + [pr.subj] + ar + SM_verb"
output English "that + pr.subj + is about to + verb"

Done. - Francis Tyers

For Welsh pattern "[pr.poss] + VBSER.inf_mutated + [pr.subj] + newydd + SM_verb"
output English "that + pr.subj + have/s just + verb.pp"

In the above Welsh patterns, at least one of pr.poss and pr.subj must be present.

mae hi'n dweud eu bod nhw wedi mynd -> *she is saying their that they have gone - fine apart from the redundant possessive

Now gives: she is saying that they have gone - Francis Tyers

Excellent. - Donnek

mae'n amlwg ei fod o'n dweud y gwir -> *is #obvious<adj><sint> his be hesaying the true - it is obvious that he is telling the truth

Now gives: is obvious that he is saying the true (perhaps "dweud y gwir" might be a good multiword verb → "tell the truth"?) - Francis Tyers

That would be a good shortcut. "gwir" is both adj (true) and noun.m (truth), but the second is not in Eurfa! - Donnek

cy-en    mae'n amlwg ei fod o'n dweud y gwir 
         is obvious that he is telling the truth 

dywedodd y bachgen ei bod hi newydd siarad â nhw -> *the boy said his be she new talk with they - the boy said that she had just talked to them

I'm leaving this one for now "... newydd ..", as it is harder to do as we don't mark adjective chunks with their lemma. - Francis Tyers

Subject pronoun + verb (marked construction?)

"Fe dagodd ei wraig ar eu gwely yn eu cartref yn Abercynon ym mis Ebrill y llynedd ar ôl iddi ddweud ei bod yn ei adael am ddyn arall."

Gives:

"He his wife choked on their bed in their home in Abercynon in the April last year after to her say his be in his leave for other man."

The "normal" (VSO) order ("Dagodd fe ei wraig") would give:

he Choked his wife

Should we re-order prnsubj + verbcj → verbcj + prnsubj in the initial stage in order to normalise the word order?

Interesting - "fe dagodd ei wraig" is actually ambiguous in Welsh. It could mean "his wife choked", "he choked his wife", or at a pinch "it was he who choked his wife" (which would have suprasegmental differences). The news item is carrying over the subject of the previous sentence (the man) into this one, so the second choice is the correct one. But in isolation we wouldn't know.

I would hold on your suggestion, because the sentence actually omits the pr.3p.sing which would disambiguate - the "fe" at the beginning is not the (southern) pr.3p.sing - it is a (mostly southern) preverbal affirmative particle (equivalent to the (mostly northern) "mi"), in the same class as preverbal interrogative particle "a" and preverbal negative particle "ni". In fact, we need a rule to delete "fe/mi" before a conjugated verb (not perfect, but better than default), and in this case the sentence would have come out as the first choice ("his wife choked").

This is wrong in this context, but I can't see what Apertium could do about this without some complex inter-sentence parsing (if it's any consolation, a human reading this sentence in isolation might also make the same mistake until he came to the latter part).

Incidentally, if the pr.poss were omitted from "fe dagodd ei wraig" it would not be ambiguous - "fe dagodd gwraig - a woman choked", but "fe dagodd wraig - (he) choked a woman".

(Another incidentally - I have asked around, and while people do seem to accept that "his wife choked" is a possible interpretation, they discount it as a likely one. They would either use a different construction for this ("naeth ei wraig dagu"), or expect some additional information before considering it ("tagodd ei wraig ar afal" - his wife choked on an apple). We can't replicate this feeling in Apertium, however, because the sequence is ubiquitous with other verbs. In this case, there may be something inherent in the transitive/intransitive meanings that make people interpret the sequence as one rather than the other. It's certainly an interesting issue - surely a journal article in there somewhere!)

The "i + infin" needs a separate section.

The last part of the sentence "... ei bod yn ei adael am ddyn arall - ... that she is leaving him for another man" is in fact a special case of 1.3.26, with an object pronoun on the verb. We can see this if we take out the pr.obj:

ei bod hi'n gadael am ddyn arall -> that she is leaving for other man

Correct apart from "other", and that may be fixed if the revised 1.3.10 can be addressed. But note:

ei bod yn gadael am ddyn arall -> *his be leaving for other man

where the omission of the pr.subj means that the 1.3.26 rule is not being applied - we need to allow for this. In informal Welsh, the usual thing is to omit the pr.poss; in formal Welsh the usual thing is to omit the pr.subj. That may not be easy to handle.

I'm not going to deal with pr.obj in this construction until I have a few more basic constructions flagged - perhaps at the end of the week. - Donnek

Aha, thanks. Regarding the pr.obj construction, no problem. - Francis Tyers

Lexis

• cwm → cirque (replace with: cwm → valley)

Yes, the Eurfa entry is tagged (geography) so this makes sense. But "cym" would be wrong - cwm (valley), cymau or cymoedd (valleys) - Donnek

Oops, typo :) - Francis Tyers

• Some verb lemmas are being generated incorrectly, (see here)

"cael" - get / have

"cael" is in the dix as "get", because in most cases that is the better gloss:

mi aeth o i'r banc i gael pres allan -> *mi he went in the bank I get money *allan - he went to the bank to get money out

Apertium's version shows some regressions (we can omit the "mi" - see comment to 1.3.27, and "allan", along with the other Eurfa adverbs, has yet to be added to the dix). "i" should not be glossed as "in", but "to" (see comment to 1.3.7). Second, the shortened form of pr.sub.1p.sing ("i") will *never* occur without a preceding conjugated verb, so that option has to be banned in this context.)

Regarding the shortened prn.subj, yep, this is the biggest regression we've had and I'm currently trying to fix it. - Francis Tyers 14:14, 1 July 2008 (UTC)

mae o'n cael mynd ar ôl y cwbl -> *he is getting go after the all - he's getting to go after all

Pretty creditable attempt, but could be improved if 1.3.16 was applied (if infin is not preceded by "yn", translate as "to"+ infin).

mi gafodd yr athro radd dda o Brifysgol Bangor -> **mi the teacher got good degree of Bangor-University - the teacher got a good degree from the University of Bangor

"mi" again (1.3.27) and inserted "a" (1.3.10) would improve this, but choosing "of" or "from" for "o" will be difficult. For "Bangor-University" see 1.3.31 below.

Note that "cael" in this sentence focusses on the "getting" - focussing on possession uses a different construction:

mae gan yr athro radd dda o Brifysgol Bangor -> *is with the teacher good degree of Bangor-University - the teacher has a good degree from the University of Bangor

"cael"is also used to create a passive construction:

mi gafodd y dyn ei daro gan gar -> **mi the man got his strike with car - the man was struck by a car

Apertium's version is pretty close to the literal - "he got his striking by a car" - which is close to the English "he got himself struck by a car"

However, in the coolness sentence this gloss is inappropriate:

ei bod hi'n cael perthynas â dyn llawer hŷn -> *that she is getting relation with a much older man - that she is having a relationship with a much older man

Can we try the following rule:

For Welsh pattern "cael + NP + {â, gyda}"
output English "have + NP + with"

"i" + infin

After most time conjunctions (eg

• cyn - before,
• ar ôl - after,
• ers - since,
• nes - until,
• erbyn - by the time that,
• wedi - after,
• wrth - while),
• rhag ofn (in case),
• er mwyn (in order to),

the conjunctive clause uses "i" + SM_infin to express person (with "i" being conjugated as appropriate), with tense being dependent on the first part of the sentence.

In terms of conjunctions, we currently have:

<l>a</l><r>and</r>
<l>ond</l><r>but</r>
<l>neu</l><r>or</r>

I'll add these, although some of them already exist as prepositions (which will be fun for the tagger to distinguish between). Should I just add them all to the bidix and Welsh monodix as given above? - Francis Tyers

I tend to think of them just as prepositions with "i" after them, ie a special form of prepositional phrase. But you're right, it would probably be best to tag them as conjunctions. You could get the tagger to distinguish them from prepositions by specifying that the conjunctive version will always be follows by "i" (or a conjugated form of it). - Donnek

Ok, I've added some others I found in my grammar too... probably some of them are variants (north/south) maybe? e.g. it has "gan" for "since". A quick other question:

Until I get the tagger up to speed, will it make much difference if I identify these in the transfer rules based only on lemma as opposed to on lemma+POS ? e.g. this means that if the tagger accidentally tags "wrth" as a preposition, it could still apply it in the transfer rules below. - Francis Tyers

Yes, there are others I haven't got around to yet. "gan" (because, since) and "am" (because, since) are two of them, but they use a "bod" construction similar to 1.3.26, so that's why they're not in this section.

Re the tagger question, I think that should be cause no problems, because the key point is that all of these are followed with "i" - that's why I said above that I tend to think of them as "preposition + i construction". - Donnek

(Note in the examples below that "i'r" is still coming up as "in the", which needs to be addressed.)

Done. This was confusing me for a while, but turns out it was down as "yn+yr" not "i+yr"! - Francis Tyers

The little rotter. Hopefully it has been squashed now. - Donnek

aeth y bachgen wedi i'r bws ddod - *the boy went after in the bus come - the boy went after the bus came

aeth y bachgen wedi iddo fo ddod -> *the boy went after to him come - the boy went after he/it came

bydd o wedi gadael erbyn i'r llythyr gyrraedd -> *that he has left by in the letter arrive - he will have left by the time the letter arrives

bydd o wedi gadael erbyn iddo fo gyrraedd -> *that he has left by to him arrive - he will have left by the time he/it arrives

wrth i'r heddlu ddod i mewn, rhedodd y dyn i mewn i'r stryd -> *beside in the police come I in, the man ran intothe road - as the police came in, the man ran into the street

wrth iddyn nhw ddod i mewn, rhedodd y dyn i mewn i'r stryd -> *beside to them come I in, the man ran intothe road - as they came in, the man ran into the street

For Welsh pattern "cyn + {"i" + NP, "i".conj + [pr.subj]} + SM_verb.infin"
output English "before + {NP, pr.subj) + verb.present/verb.past (see below)

Done. - Francis Tyers

For Welsh pattern "wedi + {"i" + NP, "i".conj + [pr.subj]} + SM_verb.infin"
output English "after + {NP, pr.subj) + verb.present/verb.past

Done. - Francis Tyers

For Welsh pattern "ar ôl + {"i" + NP, "i".conj + [pr.subj]} + SM_verb.infin"
output English "after + {NP, pr.subj) + verb.present/verb.past

For Welsh pattern "nes + {"i" + NP, "i".conj + [pr.subj]} + SM_verb.infin"
output English "until + {NP, pr.subj) + verb.present/verb.past

For Welsh pattern "erbyn + {"i" + NP, "i".conj + [pr.subj]} + SM_verb.infin"
output English "by the time that + {NP, pr.subj) + verb.present/verb.past

For Welsh pattern "wrth + {"i" + NP, "i".conj + [pr.subj]} + SM_verb.infin"
output English "as + {NP, pr.subj) + verb.present/verb.past

Done. - Francis Tyers

For Welsh pattern "rhag ofn + {"i" + NP, "i".conj + [pr.subj]} + SM_verb.infin"
output English "in case + {NP, pr.subj) + verb.present/verb.past

For Welsh pattern "er mwyn + {"i" + NP, "i".conj + [pr.subj]} + SM_verb.infin"
output English "in order for + {NP, pr.subj) + verb.present/verb.past

For Welsh pattern "ers + {"i" + NP, "i".conj + [pr.subj]} + SM_verb"
output English "since + {NP, pr.subj) + verb.past

Note: the use of present or past tense depends on the main verb (except with "ers", when you really only use it in the past). If Apertium can only choose one, then use the present as default. But it would be nice to be able to switch based on the main verb - if this is past, use the past tense (preterite) in English; if it is non-past, use the present tense in English (see comments to 1.3.15).

I'll try setting a variable which tracks the tense of the last conjugated verb (e.g. no inf, ger, etc.) - Francis Tyers

That would be great, becasue it makes quite a difference in terms of how the translation reads to an English speaker. - Donnek

It's done, but could I get examples for each of the rules above so I can test them? - Francis Tyers

I'll do full versions for "cyn":

roedd y siop wedi cau cyn i'r bachgen orffen ei ginio -> *The shop was after close before the boy completed his dinner - the shop had closed before the boy finished his dinner

roedd y siop wedi cau cyn iddo orffen ei ginio -> *The shop was after close before him completed his dinner - the shop had closed before he finished his dinner

Great - the "after close" is because we don't have a rule yet for periphrastic tenses.

caeodd y siop cyn i'r bachgen orffen ei ginio -> The shop closed before the boy completed his dinner - the shop closed before the boy finished his dinner

caeodd y siop cyn iddo orffen ei ginio -> The shop closed before him completed his dinner - the shop closed before he finished his dinner

Great.

mae'r siop yn cau cyn i'r bachgen orffen ei ginio -> *The shop is closing before the boy complete his dinner - the shop closes before the boy finishes his dinner

mae'r siop yn cau cyn iddo orffen ei ginio -> *The shop is closing before him complete his dinner - the shop closes before he finishes his dinner

Fine, apart from the fact that the pr number is not being carried across to the verb.

bydd y siop yn cau cyn i'r bachgen orffen ei ginio -> *Are the shop closing before the boy #complete<vblex><imp> his dinner

bydd y siop yn cau cyn iddo orffen ei ginio -> *Are the shop closing before him #complete<vblex><imp> his dinner

Not so good. The main problem is that the imperative is being chosen instead of the future, and I suspect this then mucks up your tense choice.

Note that with the pronoun versions the pr.obj is being used instead of the pr.subj.

Similar sentences for "wedi" (after):

roedd y siop wedi cau wedi i'r bachgen orffen ei ginio

roedd y siop wedi cau wedi iddo orffen ei ginio

caeodd y siop wedi i'r bachgen orffen ei ginio

caeodd y siop wedi iddo orffen ei ginio

mae'r siop yn cau wedi i'r bachgen orffen ei ginio

mae'r siop yn cau wedi iddo orffen ei ginio

bydd y siop yn cau wedi i'r bachgen orffen ei ginio

bydd y siop yn cau wedi iddo orffen ei ginio

Similar sentences for "wrth" (as, while):

roedd y siop wedi cau wrth i'r bachgen orffen ei ginio

roedd y siop wedi cau wrth iddo orffen ei ginio

caeodd y siop wrth i'r bachgen orffen ei ginio

caeodd y siop wrth iddo orffen ei ginio

mae'r siop yn cau wrth i'r bachgen orffen ei ginio

mae'r siop yn cau wrth iddo orffen ei ginio

bydd y siop yn cau wrth i'r bachgen orffen ei ginio

bydd y siop yn cau wrth iddo orffen ei ginio

- Donnek

Conjunctive genitive with proper names

The sentence from 1.3.29 above:

mi gafodd yr athro radd dda o Brifysgol Bangor -> **mi the teacher got good degree of Bangor-University - the teacher got a good degree from the University of Bangor

suggests the need to deal separately with proper nouns; a useful shorthand would be "any words not coming at the beginning of a sentence which are capitalised".

The "Bangor-University" is, I believe, an earlier rule intended to deal with compound nouns ( get "*woods-things" for "pethau pren" - wooden things, for instance), and I wonder whether it might be commented out for the present?

Yep, actually I haven't yet implemented the rule at the very top (in fact the first rule, as I'm still considering how to do it most effectively. I will make this a priority. - Francis Tyers

Ah, I see. Yes, it would be nice to get that sorted, because it's probably the most distinctive Welsh construction vis-a-vis English. - Donnek

For more discussion, see 1.3.1.1. - Francis Tyers

In this and similar instances it would be preferable for variations of the normal genitive rule (1.3.1) to apply (a capitalised noun is by definition definite), and we can therefore suggest an additional rule:

For a Welsh phrase of the type "!det.def + noun1 + noun_capitalised"
output English to "det.def + noun1 + of + noun_capitalised"

Should we make this noun_capitalised or "proper name" (we mark proper names separately in the dictionary and separate them into "anthroponyms", "cognomens" and "toponyms" - Francis Tyers

Ah, excellent. Would this offer a way of dealing with items like "Red Branch" below? Any of these three categories would use the rule above, but capitalised words that are not marked as one of the above would instead get the det.def added before the capitalised_noun. - Donnek

The following is another example:

Flwyddyn neu ddwy ar ôl priodas Cwchwlin, paratodd Bricriw Dafod Ffals wledd yn ei gastell a gwahoddodd y Brenin Conor ac arwyr y Gangen Goch.

This is very lightly edited actual text from p55 of "Cwchwlin" by Ivor Owen, a retelling of the Irish Cúchulain legend. Apertium's version of this is:

Year or two after marriage *Cwchwlin, prepared *Bricriw *Dafod *Ffals feast in his castle that invited the King *Conor and the heroes Red Branch.

This is actually pretty good.

The rule above would deal with the following:

ar ôl priodas Cwchwlin -> *after marriage Cwchwlin - after the marriage of Cwchwlin

This is kind of done, if we choose a "known" name, we get: "Year or two after the marriage of Elin." - Francis Tyers

Well done.- Donnek

The other example:

ac arwyr y Gangen Goch -> *and the heroes Red Branch - and the heroes of the Red Branch

should already be caught under the second rule under 1.3.1 above, but isn't. Unfortunately, the rule in this section would give the suboptimal "and the heroes of Red Branch". It may be possible to differentiate between different types of proper name (ie those which have an English gloss and those which don't), but in the meantime, on balance, it makes sense to implement the rule in this section.

The extended rule in 1.3.10 would deal with "(a) year" and "(a) feast".

There is the unsolved issue with word-sequence for unknown words (in my view, some pragmatic response to this has to be found, because it is unlikely that Apertium will ever be dealing with a text in which it knows every word). The worst infelicity in the above is that tagger has selected "a" as the relative particle "who, which, that" instead of the much more likely "and" (which is the correct one in this case). This should actually have been ruled out in this case because relative-a is followed by SM, while conjunctive-a is not.

Yep, this is a tagger error and I'm working to resolve it. Unfortunately the forbid/enforce rules don't appear to be applying themselves :/ - Francis Tyers

Subordinate "bod + pr" should not apply to inflected forms

mae o wedi dod -> *that he has come - he has come

mae o am adael -> *that he will leave - he will leave

I've restricted the "bod" to infinitive, now it gives:

He is after come

He is for leave

This does not seem to be a regression (I checked the tests), just that we haven't had a rule for verbcj{bod} SN{prnsubj} SP verbinf{SV} yet. - Francis Tyers

That's right - this is what I would have expected. We need to develop rules for periphrastic tenses. - Donnek

The rule in 1.3.26 above is being applied too broadly. As stated there, it should only apply when we have "bod" (ie the infinitive) in the verb place, eg:

ei fod o wedi dod -> that he has come

ei fod o am adael -> that he will leave

Elided pr.poss after a vowel

neidiodd Cwchwlin ar ei draed a'i wyneb yn fwgwd o waed, ac â'i gleddyf torrodd bob un o'r pennau sgrechlyd aflafar oddiar eu hysgwyddau

-> *Jumped *Cwchwlin on his #feet<n><sg> and his face in mask of blood, and with'to *gleddyf broke each one of the heads *sgrechlyd *aflafar *oddi on their *hysgwyddau

Cwchwlin jumped to his feet, his face a mask of blood, and with his sword cut each one of the shrieking, harsh heads from their shoulders. (Cwchwlin p110. Apologies to those of a nervous disposition!)

The key point here is "â'i" (we'll use a noun that is actually in the dix)

*â ei car -> *with his car

â ei gar -> with his car - with his car

â ei char -> *with his car - with her car

*â'i car -> With his car

â'i gar -> With his car - with his car

â'i char -> *With his car - with her car

Elided versions of the pr.poss are not being picked up, and no attention is being paid to the mutation of the following noun even with the non-elided pr.poss.

I'm not sure whether the elided (sometimes called infixed) pr.poss, which occur after vowels, should be entered separately in the dix, or handled with rules.

At the moment I'm handling them in the dictionary, making ad-hoc additions to the "postblank" section (right at the bottom of the file) as I go along. - Francis Tyers

1p.sing - 'm

fy mrawd a'm chwaer (my brother and sister) - see below

2p.sing - 'th + SM

gyda'th dad (with your dad) - see below

3p.sing.m - 'w + SM after "i", 'i + SM elsewhere

i'w dŷ (to his house), ei fam a'i dad (his mother and father), gyda'i arian (with his money)

3p.sing.f - 'w + {AM, h before a vowel} after "i", 'i + {AM, h before a vowel} elsewhere

i'w thŷ (to her house), ei mam a'i thad (her mother and father), gyda'i harian (with her money)

1p.pl - 'n + h before a vowel

o'n gwlad (from our country)

2p.pl - 'ch

cario'ch pethau (to carry your things)

3p.pl - 'w after "i", 'u + h before a vowel elsewhere

i'w tŷ (to their house), ei fam a'i dad (his mother and father), gyda'u harian (with their money) Note that this will deal with "ar eu hysgwyddau -> on their *hysgwyddau" in the sentence above. This works fine at present if we drop the h-mutation: "ar eu ysgwyddau -> on their shoulders".

Although these elided forms can be used after verbs etc, they are most likely after prepositions. All except 1p.sing and 2p.sing are used after any vowel, but these two ('m, 'th) can only be used after:

• a (and)
• â (with)
• gyda (with)
• efo (with)
• tua (towards)
• na (than, nor)
• i (to)
• o (from)

There are also forms that are used when the pronoun is a direct object. In most cases, these are the same as the above, but in 3p we have 's instead of 'w. There are also slight mutation changes. These forms are typical of old-fashioned written Welsh, so perhaps we could ignore them for now.

As regards rules:

For Welsh pattern "vowel_!i + 'i + {SM_noun, SM_infin}"
output "his {noun, infin}"

For Welsh pattern "vowel_i + 'w + {SM_noun, SM_infin}"
output "his {noun, infin}"

For Welsh pattern "vowel_!i + 'i + {AMH_noun, AMH_infin}"
output "her {noun, infin}"

For Welsh pattern "vowel_i + 'w + {AMH_noun, AMH_infin}"
output "her {noun, infin}"

For Welsh pattern "vowel_!i + 'u + {AMH_noun, AMH_infin}"
output "their {noun, infin}"

For Welsh pattern "vowel_i + 'w + {AMH_noun, AMH_infin}"
output "their {noun, infin}"

For Welsh pattern "vowel + 'ch + {noun, infin}"
output "your {noun, infin}"

For Welsh pattern "vowel + 'n + {noun, infin}"
output "our {noun, infin}"

This will clash with cases where "yn" becomes "'n" - you can implement it, or leave it for now.

For Welsh pattern "{a, â, gyda, efo, tua, na, i, o} + 'm + {noun, infin}"
output "my {noun, infin}"

For Welsh pattern "{a, â, gyda, efo, tua, na, i, o} + 'm + {SM_noun, SM_infin}"
output "your {noun, infin}"

Other issues:

"oddi ar" (from on - the opposite of "ar", using the ar paradigm) is not in the dix, although "oddi wrth" (from - the opposite of "at", using the wrth paradigm) is.

I'm not sure why "ar ei draed" (onto his feet) stumbles (if you'll pardon the pun).

We didn't have "feet" in the English dictionary, only foot sg/pl. I made an entry in the bidix that maps 'traed' → foot(pl), and now it seems to work. - Francis Tyers

Preverbal particles - negative

As noted in 1.3.27, preverbal particles need to be addressed. Note that this and the following sections apply mainly to formal, written Welsh.

ond ni threuliodd Cwchwlin ond ychydig o'r misoedd a oedd yn weddill iddo fo yn Nhŷ'r Bechgyn

->*but We spent *Cwchwlin but #a little<n><sg> of the months that was in remnant to him in the House of the Boys (p14)

but Cwchwlin spent only a few of the months remaining to him in the Boys' House

For Welsh pattern "ni + MM_verb.inflected + subject"
output English "subject + auxiliary + not + verb"

Without the auxiliary, this would produce archaic English ("he spent not"), which would be acceptable at a pinch, but Apertium may have standard rules for producing English negatives where a simple affirmative is replaced by a periphrastic negative (I go - I don't go; I went - I didn't go; etc).

MM is the "mixed mutation" - use AM where possible, and SM everywhere else:

• c -> ch
• p -> ph
• t -> th
• g -> ---
• b -> f
• d -> dd
• ll -> l
• m -> f
• rh -> r

ni phrynodd yr athro'r papur -> *We the teacher bought the paper - the teacher did not buy the paper

"ni" becomes "nid" before a vowel, but not before a vowel "exposed" by the soft mutation of a "g":

nid aeth i'r dref -> *Came went to the town - he did not go to the town

ni welodd y bachgen y gath -> *We the boy saw the cat - the boy did not see the cat

ni orffennodd fo'r dasg -> *We completed the task was - he did not complete the task

Note here that we seem to have a regression where the tagger is choosing an unlikely inflected form of "bod" instead of the pr.3p.sing.m. This may be related to the 1.3.29 issue with "i".

For Welsh pattern "verb_inflected + {fo, o, fe, e}"
output English "he + verb_inflected"

Can we also try a more general rule based on negation, namely:

For Welsh pattern "previous_negation + ond"
output "only"

That may not be possible, of course, in which case we'll have to try something else.

The "ni" particle is typical of formal or written Welsh. Spoken or informal Welsh tends to use "ddim" after the verb. But we can leave that till later.

Preverbal particles - interrogative

For Welsh pattern "a + SM_verb.inflected + subject"
output English "auxiliary + subject + verb"

Again, I assume that Apertium can deal with the technicalities of producing the proper English verbform (he went - did he go?; they see - do they see?; etc).

Note that "a" is often omitted, especially in spoken Welsh, but the soft mutation remains:

a gyrhaeddodd y llythyr? -> *And the letter arrived? - Did the letter arrive?

gyrhaeddodd y llythyr? -> *The letter arrived? - Did the letter arrive?

Introducing subordinate clauses, "a" means "if, whether":

For Welsh pattern "!noun + a + SM_verb.inflected + subject"
output English "!noun + if + subject + verb"

ewch i ofyn a fyddai hi'n dod - *Go I ask and she would be coming - go and ask if she will be coming

Note that we still have the frustrating issue of the tagger choosing the wrong "i" (see: http://wiki.apertium.org/wiki/Welsh_to_English#.22i.22_as_preposition).

Ok, after a morning of furious hacking, I've fixed the problem in the constraint grammar parser I was using with Apertium. Now this should be selected properly. You can see the file here... you might find it interesting, not least because it doesn't require messing about with any XML! :) If you fancy writing rules in that formalism (to improve tagger performance), I can include them directly. I've sent you an email with an example CG for faroese that you might find interesting. - Francis Tyers

That's good news. I'll do a bit of reading on CG, which looks promising, since it is largely bracketless! - Donnek

"oni" (with mixed mutation) is used when a positive answer is expected:

For Welsh pattern "oni+ MM_verb.inflected + subject"
output English "auxiliary + not + subject + verb"

oni phrynodd yr athro'r papur? -> **oni The teacher bought the paper? - didn't the teacher buy the paper?

"oni" becomes "onid" before a vowel, but not before a vowel "exposed" by the soft mutation of a "g":

onid aeth y bws heibio'r ysgol? -> **onid The bus went #past<pr>the school? - didn't the bus go past the school?

oni welsoch chi? -> **oni Dog saw? - didn't you see?

LOL! Subject pronouns really need to be slapped into shape! That is a very existentialist translation, especially if you add "b" at the beginning.

Preverbal particles - affirmative

For Welsh pattern "{fe, mi} + SM_verb.inflected + subject"
output English "subject + verb"

"fe " is mainly southern, and "mi" is mainly northern. "fe" is the form usually used before impersonals (fe werthwyd y fferm -> *He the farm sold - the farm was sold).

There is another affirmative particle "y[r]" used before "bod":

yr ydych yn mynd -> *The are going - you are going

but that is only seen now in older written Welsh. The elided forms (eg "rydych") are already included in the "bod" paradigm, but to handle older Welsh, where "yr" is not elided, you could also have a rule:

For Welsh pattern "y[r] + VBSER.inflected + subject"
output English "subject + VBSER"

Incidentally, we get some oddities for variants of the fragment above, all of which mean "you are going":

yr ydych yn mynd -> *The are going

There is no attempt to mark the verb as 2p.pl.

yr ydych chi yn mynd -> *The dog are going

That dog is back again! "chi" meaning "dog" would never occur in this location, because there is nothing that could give it AM.

I think it's this one...

I made a rule: REMOVE N IF (-1 V) (0 N) (0 PrnSubj) (1 Pr);

It fixes that problem, but maybe it will cause others...

- Francis Tyers

LOL - catch the pigeon! If I interpret that rule correctly, it says "if a cohort contains a noun and a pr.subj, with a verb on the immediate left and a preposition on the immediate right, delete the noun"? (I see you've added quite a few rules since yesterday.) I think this should work OK, actually, but what I'm slightly nervous about is that we seem to be ignoring relevant data in the stream itself, namely (a) a 2p.pl verb is likely to be followed by a 2p.pl pr, and (b) an AM noun would only occur in an AM context, of which this is not one. Capturing (a) especially seems important - I know the constraint grammar (ch 11) did mention weighting choices by probability. I need to be more familiar with the CG syntax, and the Faroese file takes a lot of concentration! Talking of which, could you maybe do a comment below each of your Welsh rules as you do them, giving a Welsh example of what the rule acts on? This would make it easier to review them later. - Donnek

Yep, at the moment I'm just writing them as hacks. Hoping that as I add rules they'll become more clear and I'll be able to unify some. The probability stuff is for accepting prob info from an HMM tagger before (not generating it itself)... And, I'll go through and document the rules later this evening. Also, I can pass you a CG for Norwegian which uses more basic rules (but more). - Francis Tyers

yr ydych chi'n mynd -> *The you are going

This is the closest translation.

rydych chi yn mynd -> *Dog rust going

I think that's the name of a 70s album by Genesis ....

LOL!! :DD - Francis Tyers

rydych chi'n mynd -> *You rust going

I may do, but that is none of Apertium's business .... The above two are because cg-proc thinks "rydych" is a form of "rhydu (to rust)" instead of "bod".

This isn't cg-proc, its the apertium-tagger. Is there a rule I can try for disambiguating vblex.pres from vblex.prs? E.g. could the construction vblex.prs + prpers + yn + vblex.inf ever occur? - Francis Tyers

Although it could in theory, in practice it would be extremely unlikely. The subjunctive is more or less moribund now, except in stock phrases such as "doed a ddelo" (come what may), and in older written (and perhaps spoken) usage. The most common remnant is, unsurprisingly, for "bod", in the person of "bo" (3p.sing) (although other persons appear occasionally), and even then it could be argued that these are stock phrases, eg "pan fo angen" (as needed). I think that in the meantime, therefore, your rule would be OK. I also wonder about going farther, and automatically deleting vblex.prs from a cohort? This would not do in the longer run, but would cause few problems for 0.1. Phrases containing "bo" could be added to the multiword phrase section. - Donnek

The conclusion seems to be (apart from the "rhydu" problem) that the absence or non-elision of pr.subj can cause problems.

Regression tests

Main article: Welsh to English/Regression tests

Coolness factor

Disclaimer: We're not deliberately aiming the translator at crime texts, it just seems to work best with these — a subject for investigation perhaps?

Roedd y Comisiwn yn ymchwilio i'r honiadau bod yr AS wedi methu datgan £103,000 o roddion.

the Commission Was investigating the allegations that the MP has failed declare £103,000 of gifts.

"He was the Commission crookedly ymchwiliad I ' group claims be he drives ACE has failed declare he gifts." (InterTran)

Dywedodd yr heddlu fod y troseddau honedig wedi digwydd rhwng 2003 a 2007 yn Sir Benfro a Sir Gaerfyrddin.

the police Said that the alleged crimes have happened between 2003 and 2007 in Pembrokeshire and Carmarthenshire.

"He said he drives police force be the transgressions alleged has happened between 2003 I go 2007 crookedly Shire ble I go Shire Gaerfyrddin." (InterTran)

Mae'r heddlu hefyd yn ymchwilio i honiadau ei bod hi'n cael perthynas â dyn llawer hŷn.

the police Are also investigating his allegations be she getting relation with a much older man.

"He ' is being group police force also crookedly ymchwiliad I claims you go be she ' heartburn have relation he goes tight much hn & #375." (InterTran)

Mae Ymddiriedolaeth Caerdydd a'r Fro yn gwrthod dweud faint mae'r driniaeth yn costio, ond yn ôl papur newydd y Sun mae'r driniaeth cyffuriau yn costio £2,500 y mis.

Cardiff Trust and the Region Is refusing say size the treatment is costing , but according to the newspaper Sun the treatment is drugs costing £2,500 the month.

"He is being Trust Cardiff I ' go group Land refusing say as many he ' is being group treatment costing , except according to paper news the Sun he ' is being group treatment ingredients costing the month." (InterTran)