def _combine(self, registry: Registry, language: str, first: Message,
second: Message) -> Message:
log.debug("Combining two templates:")
log.debug("\t{}".format([c.value for c in first.template.components]))
log.debug("\t{}".format([c.value for c in second.template.components]))
shared_prefix = self._get_combinable_prefix(first, second)
log.debug(f"Shared prefix is {[e.value for e in shared_prefix]}")
combined = [c for c in first.template.components]
# TODO At the moment everything is considered either positive or negative, which is sometimes weird.
# Add neutral sentences.
conjunctions = registry.get("conjunctions").get(language, None)
if not conjunctions:
conjunctions = (defaultdict(lambda x: "NO-CONJUNCTION-DICT"), )
if first.polarity != first.polarity:
combined.append(
Literal(
conjunctions.get("inverse_combiner",
"MISSING-INVERSE-CONJUCTION")))
else:
combined.append(
Literal(
conjunctions.get("default_combiner",
"MISSING-DEFAULT-CONJUCTION")))
combined.extend(second.template.components[len(shared_prefix):])
log.debug("Combined thing is {}".format([c.value for c in combined]))
new_message = Message(
facts=first.facts +
[fact for fact in second.facts if fact not in first.facts],
importance_coefficient=first.importance_coefficient,
)
new_message.template = Template(combined)
new_message.prevent_aggregation = True
return new_message
def run(
self,
registry: Registry,
random: RandomState,
language: str,
location_query: str,
location_type_query: str,
dataset: str,
previous_location: str,
ignored_cols: Optional[List[str]] = None,
) -> Tuple[List[Message], List[Message], List[Message]]:
log.info(
"Generating messages with location={}, location_type={}, data={}, previous_location={}"
.format(location_query, location_type_query, dataset,
previous_location))
data_store: DataFrameStore = registry.get("{}-data".format(dataset))
log.debug("Underlying DataFrame is of size {}".format(
data_store.all().shape))
if ignored_cols is None:
ignored_cols = []
if location_query == "all":
core_df = data_store.all()
expanded_df = None
elif self.expand:
log.debug('Query: "{}"'.format(
"location=={!r}".format(location_query)))
core_df = data_store.query("location=={!r}".format(location_query))
expanded_df = data_store.query(
"location!={!r}".format(location_query))
else:
log.debug('Query: "{}"'.format(
"location=={!r}".format(location_query)))
core_df = data_store.query("location=={!r}".format(location_query))
expanded_df = None
log.debug("Resulting DataFrames are of sizes {} and {}".format(
core_df.shape,
"empty" if expanded_df is None else expanded_df.shape))
core_messages: List[Message] = []
expanded_messages: List[Message] = []
col_names = core_df
col_names = [
col_name for col_name in col_names if not (col_name in [
"location", "location_type", "timestamp", "timestamp_type",
"agent", "agent_type"
] or col_name in ignored_cols or ":outlierness" in col_name)
]
core_df.apply(self._gen_messages,
axis=1,
args=(col_names, core_messages))
if expanded_df is not None:
expanded_df.apply(self._gen_messages,
axis=1,
args=(col_names, expanded_messages))
if log.getEffectiveLevel() <= 5:
for m in core_messages:
log.debug("Extracted CORE message {}".format(m.main_fact))
for m in expanded_messages:
log.debug("Extracted EXPANDED message {}".format(m.main_fact))
log.info(
"Extracted total {} core messages and {} expanded messages".format(
len(core_messages), len(expanded_messages)))
if not core_messages:
raise NoMessagesForSelectionException("No core messages")
# Remove all but 10k most interesting expanded messages
expanded_messages = sorted(expanded_messages,
key=lambda msg: msg.score,
reverse=True)[:10_000]
log.info(f"Filtered expanded messages to top {len(expanded_messages)}")
if previous_location:
log.info("Have previous_location, generating stuff for that")
previous_location_messages = self._gen_messages_for_previous_location(
registry, language, location_query, dataset, previous_location)
log.info("Finished generating previous location related things")
else:
previous_location_messages = []
return core_messages, expanded_messages, previous_location_messages
class EUNlgService:
def __init__(
self,
random_seed: Optional[int] = None,
force_cache_refresh: bool = False,
nomorphi: bool = False,
planner: str = "full",
) -> None:
"""
:param random_seed: seed for random number generation, for repeatability
:param force_cache_refresh:
:param nomorphi: don't load Omorphi for morphological generation. This removes the dependency on Omorphi,
so allows easier setup, but means that no morphological inflection will be performed on the output,
which is generally a very bad thing for the full pipeline
"""
# New registry and result importer
self.registry = Registry()
# DataSets
DATA_ROOT = Path(__file__).parent.absolute() / ".." / "data"
self.datasets = [
"cphi",
"health_cost",
"health_funding",
]
for dataset in self.datasets:
cache_path: Path = (DATA_ROOT / "{}.cache".format(dataset)).absolute()
if not cache_path.exists():
raise IOError("No cached dataset found at {}. Datasets must be generated before startup.")
self.registry.register("{}-data".format(dataset), DataFrameStore(str(cache_path)))
# Resources
self.resources = [
CPHIEnglishResource(),
CPHIFinnishResource(),
CPHICroatianResource(),
CPHIRussianResource(),
CPHIEstonianResource(),
CPHISlovenianResource(),
ENVEnglishResource(),
ENVFinnishResource(),
HealthCostEnglishResource(),
HealthCostFinnishResource(),
HealthFundingEnglishResource(),
HealthFundingFinnishResource(),
]
# Templates
self.registry.register("templates", self._load_templates())
# Slot Realizers:
realizers: List[SlotRealizerComponent] = []
for resource in self.resources:
for realizer in resource.slot_realizer_components():
realizers.append(realizer(self.registry))
self.registry.register("slot-realizers", realizers)
# Language metadata
self.registry.register("conjunctions", CONJUNCTIONS)
self.registry.register("errors", ERRORS)
# PRNG seed
self._set_seed(seed_val=random_seed)
def _get_components(headline=False, planner="full"):
# Put together the list of components
# This varies depending on whether it's for headlines and which language we are doing stuff in
yield EUMessageGenerator(expand=True)
yield EUImportanceSelector()
if planner == "random":
yield EURandomHeadlineDocumentPlanner() if headline else EURandomBodyDocumentPlanner()
elif planner == "score":
yield EUScoreHeadlineDocumentPlanner() if headline else EUScoreBodyDocumentPlanner()
elif planner == "earlystop":
yield EUEarlyStopHeadlineDocumentPlanner() if headline else EUEarlyStopBodyDocumentPlanner()
elif planner == "topicsim":
yield EUTopicSimHeadlineDocumentPlanner() if headline else EUTopicSimBodyDocumentPlanner()
elif planner == "contextsim":
yield EUContextSimHeadlineDocumentPlanner() if headline else EUContextSimBodyDocumentPlanner()
elif planner == "neuralsim":
if headline:
yield EUHeadlineDocumentPlanner()
else:
yield TemplateAttacher()
yield EUNeuralSimBodyDocumentPlanner()
yield EmbeddingRemover()
elif planner == "full":
yield EUHeadlineDocumentPlanner() if headline else EUBodyDocumentPlanner()
else:
raise ValueError("INCORRECT PLANNER SETTING")
yield TemplateSelector()
yield Aggregator()
yield SlotRealizer()
yield LanguageSplitComponent(
{
"en": EnglishEUDateRealizer(),
"fi": FinnishEUDateRealizer(),
"hr": CroatianEUDateRealizer(),
"de": GermanEUDateRealizer(),
"ru": RussianEUDateRealizer(),
"ee": EstonianEUDateRealizer(),
"sl": SlovenianEUDateRealizer(),
}
)
yield EUEntityNameResolver()
yield EUNumberRealizer()
yield MorphologicalRealizer(
{
"en": EnglishUralicNLPMorphologicalRealizer(),
"fi": FinnishUralicNLPMorphologicalRealizer(),
"hr": CroatianSimpleMorphologicalRealizer(),
"ru": RussianMorphologicalRealizer(),
"ee": EstonianUralicNLPMorphologicalRealizer(),
"sl": SlovenianSimpleMorphologicalRealizer(),
}
)
yield HeadlineHTMLSurfaceRealizer() if headline else BodyHTMLSurfaceRealizer()
log.info("Configuring Body NLG Pipeline (planner = {})".format(planner))
self.body_pipeline = NLGPipeline(self.registry, *_get_components(planner=planner))
self.headline_pipeline = NLGPipeline(self.registry, *_get_components(headline=True, planner=planner))
T = TypeVar("T")
def _get_cached_or_compute(
self, cache: str, compute: Callable[..., T], force_cache_refresh: bool = False, relative_path: bool = True
) -> T: # noqa: F821 -- Needed until https://github.com/PyCQA/pyflakes/issues/427 reaches a release
if relative_path:
cache = os.path.abspath(os.path.join(os.path.dirname(__file__), cache))
if force_cache_refresh:
log.info("force_cache_refresh is True, deleting previous cache from {}".format(cache))
if os.path.exists(cache):
os.remove(cache)
if not os.path.exists(cache):
log.info("No cache at {}, computing".format(cache))
result = compute()
with gzip.open(cache, "wb") as f:
pickle.dump(result, f)
return result
else:
log.info("Found cache at {}, decompressing and loading".format(cache))
with gzip.open(cache, "rb") as f:
return pickle.load(f)
def _load_templates(self) -> Dict[str, List[Template]]:
log.info("Loading templates")
templates: Dict[str, List[Template]] = defaultdict(list)
for resource in self.resources:
for language, new_templates in read_templates(resource.templates)[0].items():
templates[language].extend(new_templates)
log.debug("Templates:")
for lang, lang_templates in templates.items():
log.debug("\t{}".format(lang))
for templ in lang_templates:
log.debug("\t\t{}".format(templ))
return templates
def get_locations(self, dataset: str) -> List[str]:
return list(self.registry.get("{}-data".format(dataset)).all()["location"].unique()) + ["all"]
def get_datasets(self, language: Optional[str] = None) -> List[str]:
return list(
{
dataset
for resource in self.resources
for dataset in resource.supported_data
if (language is None or resource.supports(language, dataset)) and dataset in self.datasets
}
)
def get_languages(self):
return list({language for resource in self.resources for language in resource.supported_languages})
def run_pipeline(
self, language: str, dataset: str, location: str, location_type: str, previous_location: Optional[str]
) -> Tuple[str, str]:
log.info("Running headline NLG pipeline")
try:
headline_lang = "{}-head".format(language)
headline = self.headline_pipeline.run((location, location_type, dataset, previous_location), headline_lang,)
log.info("Headline pipeline complete")
except Exception as ex:
headline = location
log.error("%s", ex)
# TODO: Figure out what DATA is supposed to be here?!
log.info(
"Running Body NLG pipeline: "
"language={}, dataset={}, location={}, location_type={}, previous_location={}".format(
language, dataset, location, location_type, previous_location
)
)
try:
body = self.body_pipeline.run((location, location_type, dataset, previous_location), language)
log.info("Body pipeline complete")
except NoMessagesForSelectionException:
log.error("User selection returned no messages")
body = ERRORS.get(language, {}).get(
"no-messages-for-selection", "Something went wrong. Please try again later",
)
except Exception as ex:
log.error("%s", ex)
body = ERRORS.get(language, {}).get("general-error", "Something went wrong. Please try again later")
return headline, body
def _set_seed(self, seed_val: Optional[int] = None) -> None:
log.info("Selecting seed for NLG pipeline")
if not seed_val:
seed_val = randint(1, 10000000)
log.info("No preset seed, using random seed {}".format(seed_val))
else:
log.info("Using preset seed {}".format(seed_val))
self.registry.register("seed", seed_val)
