Use lingua for language detection

libretranslate/detect.py

@@ -1,83 +1,36 @@
-# Originally adapted from https://github.com/aboSamoor/polyglot/blob/master/polyglot/base.py
-
-import unicodedata
-
-import pycld2 as cld2
-
-
-class UnknownLanguageError(Exception):
-  pass
+import linguars
+from functools import lru_cache
 
 class Language:
-  def __init__(self, choice):
-    name, code, confidence, bytesize = choice
+  def __init__(self, code, confidence):
     self.code = code
-    self.name = name
     self.confidence = float(confidence)
-    self.read_bytes = int(bytesize)
 
   def __str__(self):
-    return ("name: {:<12}code: {:<9}confidence: {:>5.1f} "
-            "read bytes:{:>6}".format(self.name, self.code,
-                                    self.confidence, self.read_bytes))
+    return ("code: {:<9} confidence: {:>5.1f} ".format(
+                self.code,
+                self.confidence))
 
-  @staticmethod
-  def from_code(code):
-    return Language(("", code, 100, 0))
+@lru_cache(maxsize=None)
+def load_detector(langcodes = ()):
+  languages = []
+  for lc in langcodes:
+    try:
+      languages.append(linguars.Language.from_iso_code_639_1(lc))
+    except:
+      pass # Not supported
+  
+  return linguars.LanguageDetector(languages=languages)
 
 
 class Detector:
-  """ Detect the language used in a snippet of text."""
-
-  def __init__(self, text, quiet=False):
-    """ Detector of the language used in `text`.
-    Args:
-      text (string): unicode string.
-    """
-    self.__text = text
-    self.reliable = True
-    """False if the detector used Best Effort strategy in detection."""
-    self.quiet = quiet
-    """If true, exceptions will be silenced."""
-    self.detect(text)
-
-  @staticmethod
-  def supported_languages():
-    """Returns a list of the languages that can be detected by pycld2."""
-    return [name.capitalize() for name,code in cld2.LANGUAGES if not name.startswith("X_")]
+  def __init__(self, langcodes = ()):
+    self.detector = load_detector(langcodes)
 
   def detect(self, text):
-    """Decide which language is used to write the text.
-    The method tries first to detect the language with high reliability. If
-    that is not possible, the method switches to best effort strategy.
-    Args:
-      text (string): A snippet of text, the longer it is the more reliable we
-                     can detect the language used to write the text.
-    """
-    try:
-      reliable, index, top_3_choices = cld2.detect(text, bestEffort=False)
-    except cld2.error as e:
-      if "input contains invalid UTF-8" in str(e):
-        # Fix for https://github.com/LibreTranslate/LibreTranslate/issues/514
-        # related to https://github.com/aboSamoor/polyglot/issues/71#issuecomment-707997790
-        text = ''.join([l for l in text if unicodedata.category(str(l))[0] not in ('S', 'M', 'C')])
-        reliable, index, top_3_choices = cld2.detect(text, bestEffort=False)
-      else:
-        raise e
+    top_3_choices = self.detector.confidence(text)[:3]
+    print(top_3_choices)
+    if top_3_choices[0][1] == 0:
+      return [Language("en", 0)]
+    return [Language(lang.iso_code_639_1, round(conf * 100)) for lang, conf in top_3_choices]
 
-    if not reliable:
-      self.reliable = False
-      reliable, index, top_3_choices = cld2.detect(text, bestEffort=True)
-
-      if not self.quiet and not reliable:
-        raise UnknownLanguageError("Try passing a longer snippet of text")
-
-    self.languages = [Language(x) for x in top_3_choices]
-    self.language = self.languages[0]
-    return self.language
-
-  def __str__(self):
-    text = f"Prediction is reliable: {self.reliable}\n"
-    text += "\n".join([f"Language {i+1}: {str(l)}"
-                        for i,l in enumerate(self.languages)])
-    return text

libretranslate/language.py

@@ -1,7 +1,8 @@
 
 from argostranslate import translate
+from functools import lru_cache
 
-from libretranslate.detect import Detector, UnknownLanguageError
+from libretranslate.detect import Detector
 
 __languages = None
 
@@ -13,6 +14,11 @@ def load_languages():
 
     return __languages
 
+@lru_cache(maxsize=None)
+def load_lang_codes():
+    languages = load_languages()
+    return (l.code for l in languages)
+
 def detect_languages(text):
     # detect batch processing
     if isinstance(text, list):
@@ -21,31 +27,24 @@ def detect_languages(text):
         is_batch = False
         text = [text]
 
+    lang_codes = load_lang_codes()
+
     # get the candidates
     candidates = []
     for t in text:
         try:
-            d = Detector(t).languages
+            d = Detector(lang_codes).detect(t)
             for i in range(len(d)):
                 d[i].text_length = len(t)
             candidates.extend(d)
-        except UnknownLanguageError:
+        except:
             pass
 
     # total read bytes of the provided text
     text_length_total = sum(c.text_length for c in candidates)
 
-    # Load language codes
-    languages = load_languages()
-    lang_codes = [l.code for l in languages]
-
-    # only use candidates that are supported by argostranslate
-    candidate_langs = list(
-        filter(lambda l: l.text_length != 0 and l.code in lang_codes, candidates)
-    )
-
     # this happens if no language could be detected
-    if not candidate_langs:
+    if not candidates:
         # use language "en" by default but with zero confidence
         return [{"confidence": 0.0, "language": "en"}]
 
@@ -55,7 +54,7 @@ def detect_languages(text):
         temp_average_list = []
         for lang_code in lang_codes:
             # get all candidates for a specific language
-            lc = list(filter(lambda l: l.code == lang_code, candidate_langs))
+            lc = list(filter(lambda l: l.code == lang_code, candidates))
             if len(lc) > 1:
                 # if more than one is present, calculate the average confidence
                 lang = lc[0]
@@ -68,14 +67,14 @@ def detect_languages(text):
 
         if temp_average_list:
             # replace the list
-            candidate_langs = temp_average_list
+            candidates = temp_average_list
 
     # sort the candidates descending based on the detected confidence
-    candidate_langs.sort(
+    candidates.sort(
         key=lambda l: (l.confidence * l.text_length) / text_length_total, reverse=True
     )
 
-    return [{"confidence": l.confidence, "language": l.code} for l in candidate_langs]
+    return [{"confidence": l.confidence, "language": l.code} for l in candidates]
 
 
 def improve_translation_formatting(source, translation, improve_punctuation=True):

pyproject.toml

@@ -42,7 +42,7 @@ dependencies = [
     "Flask-Session ==0.4.0",
     "waitress ==2.1.2",
     "expiringdict ==1.2.2",
-    " LTpycld2==0.42",
+    "linguars==0.4.0",
     "morfessor ==2.0.6",
     "appdirs ==1.4.4",
     "APScheduler ==3.9.1",