The Dark Sky API has, since the very beginning, included a module for producing textual weather summaries from its weather data. These summaries have always been in English (since that's the only language we know) and have always been procedurally generated (since there are so many possible weather conditions). Procedural generation makes translating these summaries into other languages especially difficult, because the naive approach—using a table lookup to replace an English sentence with one of a different language—becomes impractical, requiring a very large table to support!
This software module was developed in order to work around these issues. We are modifying the Dark Sky API text summary code to generate a machine-readable format (described below) rather than it's usual English; summaries in this new format are then handed off to this module for translation into the desired language. Since this module is open-source, anyone may contribute additional language components to it, so that the desired language can be used in the Dark Sky API.
The API (and therefore this module as well) is written in JavaScript, and meant to be used as a Node.JS module. Knowledge of that environment is required in order to contribute to this software, but this document will do its best to provide a crash-course for developers that are not familiar with Node.
You will need to have Node.JS installed. You can check to see whether it is installed by running:
$ node -v
v8.9.4
If the command gives an error message (or a version below 8), you should install Node from the Node.JS website and try again.
While this package requires no dependencies to run in production, if you want to develop against it you will need the testing library Mocha. Installing it is simple:
$ cd /path/to/translations
$ npm install
NPM is the Node Package Manager, and is part of the Node software
distribution. The above command will create the directory node_modules which
will contain the testing library. After this, you can verify that everything is
working by running the tests:
$ npm test
> [email protected] test /path/to/translations
> mocha --reporter dot --check-leaks
․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․
[...]
2335 passing (732ms)
The data passed from the Dark Sky API to this translation module is a simple, structured format reminiscent of s-expressions, consisting only of numbers, strings, and arrays. Some examples produced by the API are below:
"heavy-rain"["and", "light-wind", "light-clouds"]["starting-in", "very-light-rain", ["minutes", 15]]
Each of these expressions corresponds to a text summary in English:
- "heavy rain"
- "breezy and partly cloudy"
- "drizzle starting in 15 minutes"
Generally speaking, numbers and strings represent specific terms, while arrays
represent templates dependent upon some number of arguments (with the first
element in that array representing the form of the template). For example, in
the sentence above, "light-wind" represents the English term "breezy",
"light-clouds" represents the English term "partly cloudy", and the array
["and", X, Y] represents the English phrase "X and Y".
In this way, the meaning (in English) of any given (machine-readable) expression is intended to be fairly intuitive. However, a complete description of the input format is given below in Appendix A anyway.
There is one translation submodule per language, all found in the /lib/lang
directory. (Any source files in that directory are automatically loaded by the
library at run-time, so nothing further needs to be done once the file is
created.) Each submodule exports a JavaScript object, representing a collection
of translation templates. An expression (as described above) will get looked
up in the object and translated as the object's value dictates, recursively as
necessary.
For example, suppose we have the following object:
{
"very-light-rain": "drizzle",
"minutes": function(n) {
return n + " minutes";
},
"starting-in": function(rain, time) {
return rain + " starting in " + time;
}
}
And the expression noted above:
["starting-in", "very-light-rain", ["minutes", 15]]
The following algorithm will be applied:
-
The function will first look at the expression
["starting-in", X, Y]and find that there is a corresponding function in the associative array with two arguments. It will then recursively apply the procedure on these two arguments. -
The function will then look at the expression
"very-light-rain". There is also a match in the associative array, so this expression will be replaced with"drizzle". -
The function will then look at the expression
["minutes", 15]. There is once again a match in the associative array, for a function with a single argument. The function will once again recursively apply the procedure on the argument. -
The function will look at the expression
15. Being a number, it will simply return it verbatim. -
Having it's single argument collected,
["minutes", 15]is now replaced with the expression"15 minutes", as per the code of the function for"minutes". -
Finally, with the two arguments of
["starting-in", X, Y]collected, they are substituted into the function and a final expression is returned:"drizzle starting in 15 minutes".
Any arbitrary JavaScript code may be used in a function, but in many templating scenarios, only simple string concatenation is necessary. In these cases, a shortcut syntax is also allowed:
{
"very-light-rain": "drizzle",
"minutes": "$1 minutes",
"starting-in": "$1 starting in $2"
}
The sigiled expressions are replaced with the numbered argument to the
function ($1 with the first argument, $2 with the second, and so on).
Finally, if you need the extra power, each function's this parameter is set
to an array representing the called function's position in the expression tree.
For example, in the example above, the "minutes" function is passed this
with a value of ["starting-in", "minutes"] since "minutes" is a child of
the "starting-in" template. (This is handy for languages like Dutch or German
where, I hear, that the ordering of words are important.)
Please see /lib/lang/en.js for an example of this in action.
Once a new translation module has been created, it is advisable to write tests
for it to ensure its correctness. (In fact, it may be advisable to write the
tests beforehand!) Much of the work of this has been done for you; simply
create the file /test-cases/<language>.json. This file should contain an
associative array of translated sentences to the expression used to generate
them:
{
"drizzle starting in 15 minutes":
["starting-in", "very-light-rain", ["minutes", 15]]
}
The English test cases (/test-cases/en.json) may be used as an example
and starting place. As noted above, you can verify your tests by running
npm test. Pull requests without a full suite of passing tests will not
be accepted. Please make every effort to ensure that your tests provide
as full code coverage as possible.
When translating text summaries, please keep the following in mind:
- Text summaries are often used by API consumers in headings: be as brief as possible, and use abbreviations where appropriate.
- It is simpler to maintain one version of a language than two: avoid dialectal or regional variations if at all possible. (For example, we try to maintain one version of English, despite the several major, distinct English variants—American English, British English, etc. We have had to alter terminology a few times to avoid generating insulting summaries!)
- Try to keep the text as natural as possible, so that it is easily intelligible to an average reader. (Yes, we know this conflicts with brevity, but try your best!)
Below is a listing of every possible machine-readable summary produced by Dark Sky. The listing is recursive so as to better describe how the various structural components interact.
Instead of producing a summary of weather information, we may sometimes generate a status message indicating an error of some sort. Such messages may take one of the following forms:
["sentence", [STATUS_MESSAGE, STATUS_TYPE, REASON]]
STATUS_MESSAGE may be one of the following:
"next-hour-forecast-status": we have information to convey about our hyperlocal next-hour forecasts
STATUS_TYPE may be one of the following:
"unavailable": no forecast is available for this request"partially-unavailable": only a partial forecast is available for this request"temporarily-unavailable": no forecast is available for this request, but we expect it to be available again in the future
REASON may be one of the following:
"station-offline": we cannot generate a forecast because all nearby weather stations are offline (e.g. for maintenance)"station-incomplete": we cannot generate a forecast because of gaps in the coverage of all nearby weather stations (e.g. radar beams are blocked by local terrain)
"next-hour-forecast-status", "unavailable", "partially-unavailable", "temporarily-unavailable", "station-offline", and "station-incomplete" are not used in any other forms.
When the API is producing a text summary for a single moment in time (that is,
currently.summary and hourly.data[N].summary), summaries of the following
structure are produced:
["title", WEATHER_CONDITION]["title", ["and", WEATHER_CONDITION, WEATHER_CONDITION]]
The "title" component is never used in any other situation, and signifies
that (in English, anyway) these conditions represent phrases rather than
complete sentences; as such, they are capitalized like a title (that is, each
word is capitalized and there is no punctuation). For all below cases, the
"summary" component wraps the construction (signifying that the summary
represents a full, English sentence, meaning that only the first word is
capitalized, and the sentence is to end with a period).
"and" is used all over the place. Sorry.
For text summaries for the next hour (that is, minutely.summary), summaries
of the following formats are produced:
["sentence", ["for-hour", WEATHER_CONDITION]]["sentence", ["starting-in", PRECIPITATION_TYPE, DURATION]]["sentence", ["stopping-in", PRECIPITATION_TYPE, DURATION]]["sentence", ["starting-then-stopping-later", PRECIPITATION_TYPE, DURATION, DURATION]]["sentence", ["stopping-then-starting-later", PRECIPITATION_TYPE, DURATION, DURATION]]
Except for the first case, each such summary only takes precipitation into account, and tells how the intensity of precipitation will vary over the next hour or so.
The DURATIONs listed above may be either of:
["less-than", ["minutes", 1]]("less than a minute")["minutes", NUMBER]("X minutes")
"for-hour", "starting-in", "stopping-in",
"starting-then-stopping-later", "stopping-then-starting-later", and
"minutes" are only used as above. "less-than" is also used for snow
accumulation (see below).
Day summaries are produced by the API when a duration of 24 hours is under
consideration (that is, hourly.summary and daily.data[N].summary). They are
the most complex summaries in the API, owing to the number of possible
combinations of the various terms. They are of the following formats:
["sentence", DAY_CONDITION_SUMMARY]["sentence", ["and", DAY_CONDITION_SUMMARY, DAY_CONDITION_SUMMARY]]
A "day condition" represents a specific weather condition at a specific time of day. (Or a larger period of the day, as the case may be.)
["for-day", WEATHER_CONDITION]["during", WEATHER_CONDITION, TIME_OF_DAY]["during", WEATHER_CONDITION, ["and", TIME_OF_DAY, TIME_OF_DAY]]["starting", WEATHER_CONDITION, TIME_OF_DAY]["until", WEATHER_CONDITION, TIME_OF_DAY]["starting-continuing-until", WEATHER_CONDITION, TIME_OF_DAY, TIME_OF_DAY]["until-starting-again", WEATHER_CONDITION, TIME_OF_DAY, TIME_OF_DAY]
"for-day", "starting", "until", "starting-continuing-until", and
"until-starting-again" are only used in the above manner, and may be
considered analagous to the five similar cases in hourly summaries. "during"
is used both here and in weekly summaries, below.
Daily summaries covering a specific day use the following time periods:
"morning""afternoon""evening""night"
Daily summaries covering the next 24 hours (as in a forecast) use the following time periods instead:
"today-morning""today-afternoon""today-evening""today-night""later-today-morning""later-today-afternoon""later-today-evening""later-today-night""tomorrow-morning""tomorrow-afternoon""tomorrow-evening""tomorrow-night"
In general, the most specific case is used. (For example, if it is currently
afternoon and a weather condition would occur later in the afternoon,
"later-today-afternoon" would be used. If it was any other time of day,
"today-afternoon" would be used.)
The exact times that each duration begins or ends is not intended to be critical, and nonprecise terminology should be used if possible. However, for aid in translation, the time periods currently correspond to the following:
- morning: 04:00 (4am) to 12:00 (12pm)
- afternoon: 12:00 (12pm) to 17:00 (5pm)
- evening: 17:00 (5pm) to 22:00 (10pm)
- night: 22:00 (10pm) to 04:00 (4am)
For summaries spanning an entire week (daily.summary), the following format
is used:
["sentence", ["with", WEEKLY_PRECIPITATION_SUMMARY, WEEKLY_TEMPERATURE_SUMMARY]]
Since an entire week is a very broad span of time, we concern ourselves only with the most broadly applicable information: which days will have rain, and how the temperatures will fluctuate. The sentence is broken into two parts, which each comprise one of the above.
"with" is not used in any other way.
A "weekly precipitation summary" is used to describe which days of the week are expected to have rain, as compactly as possible.
["for-week", PRECIPITATION_TYPE]["over-weekend", PRECIPITATION_TYPE]["during", PRECIPITATION_TYPE, DAY_OF_WEEK]["during", PRECIPITATION_TYPE, ["and", DAY_OF_WEEK, DAY_OF_WEEK]]["during", PRECIPITATION_TYPE, ["through", DAY_OF_WEEK, DAY_OF_WEEK]]
"for-week", "over-weekend", and "through" are both only used as above.
"during" is used both here and in daily summaries.
A "weekly temperature summary" describes the general pattern of temperatures over the course of the next week: whether they'll get hotter, colder, hotter-then-colder, or colder-then-hotter.
["temperatures-rising", TEMPERATURE, DAY_OF_WEEK]["temperatures-falling", TEMPERATURE, DAY_OF_WEEK]["temperatures-peaking", TEMPERATURE, DAY_OF_WEEK]["temperatures-valleying", TEMPERATURE, DAY_OF_WEEK]
"temperatures-peaking", "temperatures-valleying", "temperatures-rising",
and "temperatures-falling" are all only used as above.
["fahrenheit", NUMBER]["celsius", NUMBER]
Every language should support both temperature units, as the choice of language and units are separate options in the API (and can be mixed-and-matched as desired).
"today""tomorrow""sunday""monday""tuesday""wednesday""thursday""friday""saturday""next-sunday""next-monday""next-tuesday""next-wednesday""next-thursday""next-friday""next-saturday"
"today" and "tomorrow" are used in preference to the other cases. The
"next-*" cases are used when the day in question is a week from today (e.g.
if today is Wednesday, and we expect rain a week from today, then the summary
would be ["during", "rain", "next-wednesday"].
"no-precipitation": Represents no precipitation. Only used in "weekly precipitation summary" blocks. (This condition is in contrast to"clear", which represents no significant weather of any kind.)"mixed-precipitation": Represents multiple types of precipitation, such as both rain and snow. Only used in "weekly precipitation summary" blocks; in all other cases, the predominate type of precipitation is used.GENERIC_TYPERAIN_TYPESLEET_TYPESNOW_TYPE["parenthetical", EXPECTED_PRECIP_TYPE, SNOW_ACCUMULATION]: For daily or weekly summaries, if a significant amount of snow is expected, we will qualify it with the amount of expected snow accumulation. (For example, "snow (3-4 in.) throughout the day".) PLEASE NOTE that it is possible for a chance of snow accumulation to be forecasted even if the expected precipitation type is rain or sleet: this may occur if the forecasted temperature is right around the freezing point. Translations should clarify that the parenthetical refers to a chance of snow in such circumstances. (For example, "sleet (chance of 3-4 in. of snow) throughout the day".)
In each of the below precipitation types, the intensity of precipitation is (very approximately) as follows:
"very-light-X": 0–0.4 mm/hr"light-X": 0.4–2.5 mm/hr"medium-X": 2.5–10 mm/hr"heavy-X": 10 mm/hr
Snow intensities are (also very approximately) one-third of these. (That is,
"heavy-snow" is more like 3 mm/hr.) However, these are only intended as a
rough guide, as these values change over time as we fine-tune our system.
Generic precipitation forms are used when we don't have information regarding the exact type of precipitation expected. (This is a rare occurance.)
"possible-very-light-precipitation""very-light-precipitation""possible-light-precipitation""light-precipitation""medium-precipitation""heavy-precipitation"
Rain precipitation forms represent liquid precipitation.
"possible-very-light-rain""very-light-rain""possible-light-rain""light-rain""medium-rain""heavy-rain"
Sleet precipitation forms represent sleet, freezing rain, or ice pellets, of the sort that generally occur in winter when temperatures are around freezing.
"possible-very-light-sleet""very-light-sleet""possible-light-sleet""light-sleet""medium-sleet""heavy-sleet"
Snow precipitation forms represent solid precipitation in the form of snowflakes.
"possible-very-light-snow""very-light-snow""possible-light-snow""light-snow""medium-snow""heavy-snow"
Represents a distance measurement indicating the amount of snow accumulation is expected. These take the form of "N inches", "< N inches", or "M-N inches" in English, respectively.
["inches", NUMBER]["less-than", ["inches", 1]]["inches", ["range", NUMBER, NUMBER]]["centimeters", NUMBER]["less-than", ["centimeters", 1]]["centimeters", ["range", NUMBER, NUMBER]]
"clear": Represents the lack of any significant weather occurring."possible-thunderstorm": Represents a chance of thunderstorms occurring."thunderstorm": Represents thunderstorms occurring."light-wind": Represents light wind at a location. (3 or 4 on the Beaufort scale, but only when this is historically unusual.)"medium-wind": Represents moderate wind at a location. (5, 6, or 7 on the Beaufort scale, but only when this is historically unusual.)"heavy-wind": Represents strong wind at a location. (8+ on the Beaufort scale.)"low-humidity": Represents when the air is unusually dry."high-humidity": Represents when the air is unusually humid."fog": Represents when there is less than 1 mile (1.6 kilometers) of visibility."light-clouds": Represents when clouds cover less than half of the sky. (Usually called "partly cloudy" in English.)"medium-clouds": Represents when clouds cover more than half (but not all) of the sky. (Usually called "mostly cloudy" in English.)"heavy-clouds": Represents complete (or nearly-complete) cloud cover. (Usually called "overcast" in English.)