Summary:
While I was working on fixing #604, I came across the rules
`ruleMilitarySpelledOutAMPM(2)`, which were actually capturing
some of my test phrases and confusing me.
This commit removes them because
- they aren't needed: the existing latent spelled-out hour + minute rules plus
the "(in the )?(am/pm)" rules together give the same behavior
- they are confusingly named - these aren't military times at all, they are
spelled-out civilian times
Reviewed By: haoxuany
Differential Revision: D27848485
fbshipit-source-id: ba1ed16ec22b5139b0b500b44dc91adb1b5e3d82
Summary:
This commit extend Spanish-language support for concatenations
of the form "<higher-order-of-magnitude> <lower>", e.g.
"doscientos tres" (203) or "cuatro mil ventiuno" (4022) to work
not just for hundreds but also for thousands and millions.
Reviewed By: chessai
Differential Revision: D27858133
fbshipit-source-id: 5c6b227ae7dad9009cd636e7ea49c209480c931a
Summary:
This commit adds two things to Spanish numeral support:
- support for millions
- support, via hooking into the `isMultipliable` logic used by EN, for
composing counts of 2-999 with either "mil" or "millones", which is
the standard way to say things like "tres mil" = 3000
Reviewed By: chessai
Differential Revision: D27858135
fbshipit-source-id: 980e95bd989f818c5ceaa2bb6c87fe81d3e08366
Summary:
This diff refactors our handling of "<hundreds> 0..99" numbers
to be more flexible by replacing `ruleNumeralthreePartHundreds`
with
- a rule for two-part hundreds like "dos cientos" (which is technically
incorrect grammar - doscientos is correct - but probably worth keeping) based
on a notion of multipliability like that used in EN rules
- a rule stating that we can compose hundreds with 0..99 additively
The resulting rules are more flexible, and they correctly parse not only
gramatically iffy phrases like "dos cientos tres", but also grammatically
correct phrases like "doscientos tres". This fixes#380.
Reviewed By: chessai
Differential Revision: D27858136
fbshipit-source-id: 4a918d84d93ac074f83f6947a8f80cfd11145115
Summary:
This fixes#592 in a very conservative way: the reason why `ruleIntersect` does
not detect "tonight 815" and "tonight eight fifteen" as it does "tonight 8:15"
is because it explicitly forbids the second part of the intersection from being
latent, unless it is a year.
I don't think it's a good idea to remove the restriction on latent inputs in
`ruleIntersect`, so instead I just made a new rule specifically for the
intersection of `<part-of-day> <time-of-day>`.
It also seems to me that there's a lot of room for this to be too aggressive,
for example if I say "tonight 500 people will laugh" the "tonight" and "500"
aren't really linked. So, I set the rule to be latent; this may be too conservative
to be useful though (do client libraries usually allow latent results?).
Reviewed By: chessai
Differential Revision: D27842596
fbshipit-source-id: 36ac59e31c632d4864241bce291147a46d52f780
Summary: Adds Catalan language and Numeral rules for it
Reviewed By: haoxuany
Differential Revision: D26518604
Pulled By: chessai
fbshipit-source-id: e6b4b0ceb9b7931d086c732dd03fb5cbbe062d5b
Summary:
I was testing an unrelated change (which doesn't change
classifier scores) and reran classifiers just to be safe, I noticed
that the scores changed.
This diff updates them.
Reviewed By: chessai
Differential Revision: D26892970
fbshipit-source-id: c7da3e3b7d01955f98b287a3ff4e7c1ff2837c7f
Summary: adds a rule for 'the day after tomorrow' in Romanian. regenerates classifiers.
Reviewed By: girifb
Differential Revision: D26155042
fbshipit-source-id: 80005ab94a10f9fbf242c9a712bd040e4f6bc477
Summary:
**2nd set of changes from pull request https://github.com/facebook/duckling/issues/516
Supporting Cantonese and more common expressions in Chinese.
Adding rules file for Duration/ZH.
Pull Request resolved: https://github.com/facebook/duckling/pull/523
Reviewed By: haoxuany
Differential Revision: D23428901
Pulled By: chessai
fbshipit-source-id: 6d04c97b63bac966eb61d77cab2f08f7543dbbf0
Summary:
**1st set of changes from pull request https://github.com/facebook/duckling/issues/516
Supporting more common expressions, such as fraction, half, dozen, in Chinese.
Pull Request resolved: https://github.com/facebook/duckling/pull/522
Reviewed By: patapizza
Differential Revision: D23428893
Pulled By: chessai
fbshipit-source-id: 3454ac70a4bfff90dc282560916a0fae9969f521
Summary:
* "at the moment" is considered identical to "now".
* "ASAP" is considered identical to "from now"
Pull Request resolved: https://github.com/facebook/duckling/pull/405
Reviewed By: patapizza
Differential Revision: D26009483
Pulled By: chessai
fbshipit-source-id: addf4c509e69d413cae279601c64f72710eba11f
Summary:
This pull request is to add support for Telugu language (Numerical Dimension) to Duckling
Pull Request resolved: https://github.com/facebook/duckling/pull/470
Differential Revision: D25546700
Pulled By: chessai
fbshipit-source-id: 1d88ee27da8a577a4a79ff31be8cb55ed6444c4e
Summary:
Improves the recognition of German time approximation language and removes a single error in the rule of <time-of-day> approximately.
Pull Request resolved: https://github.com/facebook/duckling/pull/435
Reviewed By: patapizza
Differential Revision: D24934281
Pulled By: chessai
fbshipit-source-id: 641bcb6a7e5c26e66c735fe13bccae9b7a8909ae
Summary:
Add support for additional Hindi numbers like 300, 81, 150, 1000, 1520. These are not supported in the current master version.
Pull Request resolved: https://github.com/facebook/duckling/pull/552
Reviewed By: ashwinp-fb, girifb
Differential Revision: D25072230
Pulled By: chessai
fbshipit-source-id: 35277a2349384bcf44a20e74852113f5c010e618
Summary:
Found a lacking frequent duration in German and a small typo in the existing one.
Pull Request resolved: https://github.com/facebook/duckling/pull/509
Reviewed By: patapizza
Differential Revision: D24690104
Pulled By: chessai
fbshipit-source-id: b49a7a636abf5b92f2fe7c0d5b2ca2fe64acbaa2
Summary:
Pull Request resolved: https://github.com/facebook/duckling/pull/533
In recent versions of Data.Some the name of the constructor, `This` has changed name to `Some`. This has become rather problematic for us to migrate so we're just going to remove the dependency. The meat of this diff is adding the type `Seal` to `Duckling.Types`. That type replaces `Some`.
Reviewed By: pepeiborra
Differential Revision: D23929459
fbshipit-source-id: 8ff4146ecba4f1119a17899961b2d877547f6e4f
Summary:
Current:
"seis cero cinco pm" [dimension Time] -> "cero cinco pm" or "5 pm"
here the term "seis" was dropped because it was treated as "6" in "Numeral" dimension.
Expected:
"seis cero cinco pm" -> "6:05 pm"
The root cause was that the rule "<hour-of-day> <integer> (as relative minutes)" dropped the first term "hour-of-day" if it was parsed as a latent token.
Reviewed By: chinmay87
Differential Revision: D22553028
fbshipit-source-id: abc92bb369c23d2b3084641eab2a2dabb87dbc66
Summary:
There are two rules for parsing "manana" (dimension: Time): one is resolved to "morning"; while the other is resolved to "tomorrow". And the first (or "morning") rule resolves to a LATENT result; while the second (or "tomorrow") rule resolves to a NON-LATENT result.
If the duckling is called with "latent" option turned off, the "tomorrow" rule prevails. However, if the duckling is invoked with "latent" option turned on, the "morning" rule is preferred.
The solution (for now) is to steer the classifier towards "tomorrow" rule by adding large number of (same) examples for "tomorrow" rule.
Reviewed By: chinmay87
Differential Revision: D22425277
fbshipit-source-id: 2f139eec0c38b9b5227f27d9f09f6264e7cf86cd
Summary:
The root cause is this lacking of support for the composition of numerals in ES.
For example, "mil novecientos noventa" is parsed 3 individual numbers: 1000, 900 and 90 correspondingly. Instead, the expected result is a single numeral value that is the sum of aforementioned three numbers. The same expection can be extended to the composition with arbitrary number of numeral values.
Reviewed By: chinmay87
Differential Revision: D22192034
fbshipit-source-id: 476489145b83297b82d88f3451020c867e2d08aa
Summary:
Current:
"first monday of last month" -> the date of first monday starting from current time. Note here the term "last month" is dropped
Expected:
"first monday of last month" -> the date of first monday of previous month.
Reviewed By: chinmay87
Differential Revision: D22300243
fbshipit-source-id: 16622860c52ec2ce9c7a7bcd6094192255aa5a0b
Summary:
Current:
"twelve zero three" -> 12:00pm
Expected:
"twelve zero three" -> 12:03pm
The root cause was that duckling doesn't support this kind of pattern for timestamp. The uniqueness here was that the number "three" was spelled as "zero three" that Duckling failed to understand.
Reviewed By: chinmay87
Differential Revision: D22313140
fbshipit-source-id: 9e481a142a16b94c61b1770e7f8be036497419f8
Summary:
current:
last friday in october -> the date of Friday of previous week
expected:
last friday in october -> the data of last Friday of month october
Reviewed By: chinmay87
Differential Revision: D22201326
fbshipit-source-id: 1983c1b9c24aa356977af7def42d5ba07c7f08be
Summary:
Current:
"seis dos de lar tarde" -> "dos de lar tarde" or 2pm; note
that the term "seis" is dropped.
Expected:
"seis dos de lar tarde" -> "seis dos de lar tarde"
or 6:02pm
Pull Request resolved: https://github.com/facebook/duckling/pull/496
Test Plan: H.io $ debug (makeLocale ES Nothing) "seis dos de la tarde" [This Time]
Reviewed By: chinmay87
Differential Revision: D22054328
Pulled By: yuanbing
fbshipit-source-id: 1ecb05885fc506176cc04768aa158279c7e7fd4f
Summary:
There are two types of ES phrases for timestamp to support:
1. "para las seis cero dos pm"
2. "para las 6 0 2 pm"
The solution is to:
1. added a new rule to parse two-digit number between 1 and 9 (inclusive);
2. modified the regex pattern to support additional optional phrase "para" in front of "las".
Reviewed By: chinmay87
Differential Revision: D22218800
fbshipit-source-id: 58f692beb6f10834c0ab639b31bf239bf4a1970e
Summary:
This fix is to add support to parse alternative phrase, in ES, for "noon".
Currently the supported ES phrase for "noon" is "mediodia", the alternative form is "medio<whitespace*>dia".
Reviewed By: chinmay87
Differential Revision: D22188049
fbshipit-source-id: 798b83be75798f3b0d695a0f01a65dc84af98e22
Summary:
the rule is updated to conform with natural expression of "ordinal day of month".
Pull Request resolved: https://github.com/facebook/duckling/pull/495
Differential Revision: D22054297
Pulled By: yuanbing
fbshipit-source-id: d9d8e00311d4d3121685ab5b09f6c1f52f3077c9
Summary:
Please note that the major diff with the
existing rule for next week is that the new
phrase doesn't have the leading "la" or anything with
similar meaning.
Pull Request resolved: https://github.com/facebook/duckling/pull/493
Test Plan: Imported from GitHub, without a Test Plan: line.
Reviewed By: patapizza
Differential Revision: D21981169
Pulled By: yuanbing
fbshipit-source-id: 7478d1262c3a4599d359b485b28a547ad5f44b76
Summary:
The root cause was the error in parsing the ES numeral value [1-9] that spelled with two words instead of one.
For example "cero dos" should be parsed the as "dos". Currently it's being as two numeral values: 0 and 3.
Reviewed By: chinmay87
Differential Revision: D22162804
fbshipit-source-id: 949956935a21e742f6788e7afa788ff728dd9a8d
Summary:
the new rules could parse phrases in the form of
xxx upcoming weeks
upcoming xxx weeks
Pull Request resolved: https://github.com/facebook/duckling/pull/491
Test Plan: Imported from GitHub, without a Test Plan: line.
Differential Revision: D21959647
Pulled By: chinmay87
fbshipit-source-id: a062a8c7a6c2e23b921b1099b886fa589c69c454
Summary:
while computing a score used to rank in Duckling, it currently sums up the log likelihoods learned during training. While ranking, the goal is to find the (same span) parse candidate which is _more_ likely to lead to a *correct* parse. However, the old logic was summing up the "more confident of the two classes" log likelihood.From what I understand this is the part which feels wrong.
I created an example of two rules:
#1. a rule where the classifier learns that the rule is very confidently NOT the correct parse.
- okdata (positive class) is very low confidence (high negative number prior)
- kodata (negative class) is very high confidence (low negative number prior)
#2. a rule where the classifier is confident that it is the correct parse, but not Very Confident.
- okdata (positive class) is high confidence (nonzero, but low negative number prior)
- kodata (negative class) is very low confidence (high negative number prior)
these two rules match the same regex, thus the same span. While duckling parses it, it turns out, that rule #1 ranks higher than rule #2. The reason why is because #1 is MORE confident that it is the INCORRECT (does not contribute to) parse than rule #2. Does this make sense?
to solve this problem, I changed the ranking score estimation to use only the positive class scores (okdata). In the example above, it fixes it so rule #2 would end up ranking higher because the positive class confidence is higher than #1's positive class confidence.
Would really love some deeper input from Duckling experts. I re-learned haskell and learned haxl to craft a small example here, and I am very new to Duckling (just started reading the ranking code on Friday). I know Duckling is battle-tested but I also don't believe that means a bug can't exist. And further, this specific bug may not happen a whole lot for 2 reasons:
- there are not a lot of rules which end up higher negative confidence than positive (requires enough negative corpus examples over positive ones)
- ranking uses span width first, and only when the spans are equivalent does the score based ranking come into play. So it requires that 2 rules match the same span before any actual score calculation even matters.
Reviewed By: patapizza
Differential Revision: D22009276
fbshipit-source-id: 13491689d39d810da526fa4bb8b6e526d4cafd35
Summary:
added new EN rule to parse the phrases that contain "midday".
Pull Request resolved: https://github.com/facebook/duckling/pull/490
Differential Revision: D21959562
Pulled By: chinmay87
fbshipit-source-id: f9ab45aecd551e8959d00b0025ed38b616ed6b14
Summary:
Current:
"el dia nueve" -> "9pm" of current day
Expected:
"el dia nueve" -> 9th of current or next month
Fix:
added new ES rule to handle the pattern like "el dia <day of month>"
Pull Request resolved: https://github.com/facebook/duckling/pull/487
Reviewed By: girifb
Differential Revision: D21850807
Pulled By: chinmay87
fbshipit-source-id: d8edd81273c7e5f700b440ccc8c7e7bded679051
Summary: Fix `ruleYearLatent` to be the same as the one in `en`. We don't want to match numerals that could have been hours.
Reviewed By: patapizza
Differential Revision: D20683975
fbshipit-source-id: cdef9b1b5f8a21dc5e207ed2a7afcad84c56a596
Summary:
Leveraging `predNthClosest` helper in English rules.
"the second closest monday to february 6"
"the closest tax day to boss day 2018"
Reviewed By: haoxuany
Differential Revision: D20214444
fbshipit-source-id: b6be32f63097d221aa7ccc6df4e3639e4deee4a9
