sonatoki 0.1.0__py3-none-any.whl → 0.1.2__py3-none-any.whl

sonatoki/Preprocessors.py

@@ -121,6 +121,9 @@ class ArrowQuote(RegexPreprocessor):
 
 
 __all__ = [
+    "DiscordChannels",
+    "DiscordMentions",
+    "DiscordSpecial",
     "DiscordEmotes",
     "SingleQuotes",
     "DoubleQuotes",

sonatoki/Scorers.py

@@ -1,5 +1,6 @@
 # STL
 import math
+import logging
 from abc import ABC, abstractmethod
 from typing import Dict, List, Type, Union
 
@@ -9,24 +10,20 @@ from typing_extensions import override
 # LOCAL
 from sonatoki.Filters import Filter
 
+LOG = logging.getLogger(__name__)
+
 Number = Union[int, float]
 Weights = Dict[str, Number]
 
 
-class Scorer(ABC):
-    weights: Weights
-
-    # @classmethod
-    # def __score(cls, token: str, filters: List[Type[Filter]]) -> Tuple[int, Number]:
-    #     for filter in filters:
-    #         if not filter.filter(token):
-    #             continue
-    #         # NOTE: We assume the filters are ordered by their score
-    #         # Thus the first match is also the highest scoring
-    #         return filter.counts, cls.weights[filter.__name__]
-    #         # TODO: override weight if count is 0?
-    #     return 1, 0
+def sigmoid(n: int) -> Number:
+    return 1 / (1 + math.exp(-(0.30 * (n - 1))))
+    # n-1 makes sigmoid(1) == 0.5
+    # 0.30 softens scaling in favor of short input
+    # return n / (1+abs(n))   # too weak in 0.7+
+
 
+class Scorer(ABC):
     @classmethod
     @abstractmethod
     def score(cls, tokens: List[str], filters: List[Type[Filter]]) -> Number:
@@ -37,10 +34,15 @@ class PassFail(Scorer):
     """The token passes any filter or fails all of them, scoring 1 or 0 respectively."""
 
     @classmethod
-    def __score(cls, token: str, filters: List[Type[Filter]]) -> Number:
+    def score_token(cls, token: str, filters: List[Type[Filter]]) -> Number:
         for f in filters:
             if f.filter(token):
-                return 1
+                score = 1
+                LOG.debug(
+                    "%12s.%s('%s') = %.2f", cls.__name__, f.__name__, token, score
+                )
+                return score
+        LOG.debug("%12s('%s') = 0.00", cls.__name__, token)
         return 0
 
     @classmethod
@@ -52,10 +54,27 @@ class PassFail(Scorer):
         total_score = 0
         len_tokens = len(tokens)
         for token in tokens:
-            total_score += cls.__score(token, filters)
+            total_score += cls.score_token(token, filters)
         return total_score / len_tokens if len_tokens else 0
 
 
+class SoftPassFail(PassFail):
+    @classmethod
+    @override
+    def score(cls, tokens: List[str], filters: List[Type[Filter]]) -> Number:
+        if not tokens:
+            return 1
+
+        total_score = 0
+        len_tokens = len(tokens)
+        for token in tokens:
+            total_score += cls.score_token(token, filters)
+
+        percentage = total_score / len_tokens if len_tokens else 0
+        percentage **= sigmoid(len_tokens)
+        return percentage
+
+
 class Scaling(Scorer):
     """
     The sooner a token matches a filter, the higher its score.
@@ -67,7 +86,12 @@ class Scaling(Scorer):
     def score_token(cls, token: str, filters: List[Type[Filter]], scale: int):
         for i, f in enumerate(filters):
             if f.filter(token):
-                return scale - i
+                score = scale - i
+                LOG.debug(
+                    "%12s.%s('%s') = %.2f", cls.__name__, f.__name__, token, score
+                )
+                return score
+        LOG.debug("%12s('%s') = 0.00", cls.__name__, token)
         return 0
 
     @classmethod
@@ -91,13 +115,6 @@ class SoftScaling(Scaling):
     For example, a single token scoring 0.64 will now score 0.8.
     """
 
-    @staticmethod
-    def sigmoid(n: int) -> Number:
-        return 1 / (1 + math.exp(-(0.30 * (n - 1))))
-        # n-1 makes sigmoid(1) == 0.5
-        # 0.30 softens scaling against input
-        # return n / (1+abs(n))   # too weak in 0.7+
-
     @classmethod
     @override
     def score(cls, tokens: List[str], filters: List[Type[Filter]]) -> Number:
@@ -113,11 +130,11 @@ class SoftScaling(Scaling):
             total_score += cls.score_token(token, filters, len_filters)
 
         percentage = total_score / max_score if max_score else 0
-        percentage **= cls.sigmoid(len_tokens)
+        percentage **= sigmoid(len_tokens)
         return percentage
 
 
 class Logarithmic(Scorer): ...
 
 
-__all__ = ["PassFail", "Scaling", "SoftScaling"]
+__all__ = ["PassFail", "SoftPassFail", "Scaling", "SoftScaling"]

sonatoki/Tokenizers.py

@@ -4,6 +4,8 @@ from typing import List, Callable
 # PDM
 import regex as re
 
+# TODO: Entire module should be reworked to match the class scheme of the rest of the module, imo
+
 try:
     # PDM
     import nltk
@@ -15,18 +17,14 @@ except ImportError as e:
 
 LANGUAGE = "english"  # for NLTK
 
-SENT_DELIMS_RE = r"""(.*?[.?!;:])|(.+?$)"""
-SENT_DELIMS_RE = re.compile(SENT_DELIMS_RE)
-
-SENT_DELIMS_TOK = r"""(.*?[.?!;:-])|(.+?$)"""
-SENT_DELIMS_TOK = re.compile(SENT_DELIMS_TOK)
-
-
-WORD_DELIMS_RE = r"""\s+|(?=[.?!;:'"-])"""
-WORD_DELIMS_RE = re.compile(WORD_DELIMS_RE)
+SENT_DELIMS_RE = re.compile(r"""(.*?[.?!;:])|(.+?$)""")
+SENT_DELIMS_TOK = re.compile(r"""(?<=[.?!:;·…“”"'()\[\]\-]|$)""")
+# TODO: are <> or {} that common as *sentence* delims? [] are already a stretch
+# TODO: do the typography characters matter?
+# NOTE: | / and , are *not* sentence delimiters for my purpose
 
-WORD_DELIMS_TOK = r"([\p{Punctuation}\p{posix_punct}]+|\s+)"
-WORD_DELIMS_TOK = re.compile(WORD_DELIMS_TOK)
+WORD_DELIMS_RE = re.compile(r"""\s+|(?=[.?!;:'"-])""")
+WORD_DELIMS_TOK = re.compile(r"([\p{Punctuation}\p{posix_punct}]+|\s+)")
 
 Tokenizer = Callable[[str], List[str]]
 
@@ -53,11 +51,7 @@ def word_tokenize_re(s: str) -> List[str]:
 
 
 def sent_tokenize_tok(s: str) -> List[str]:
-    return [
-        clean
-        for sent in re.findall(SENT_DELIMS_TOK, s)
-        if (clean := sent[0].strip() or sent[1].strip())
-    ]
+    return [clean for sent in re.split(SENT_DELIMS_TOK, s) if (clean := sent.strip())]
 
 
 def word_tokenize_tok(s: str) -> List[str]:

sonatoki/ilo.py

@@ -1,5 +1,6 @@
 # STL
-from typing import List, Type
+import logging
+from typing import List, Type, Tuple
 
 # LOCAL
 from sonatoki.Filters import Filter
@@ -8,6 +9,8 @@ from sonatoki.Cleaners import Cleaner
 from sonatoki.Tokenizers import Tokenizer
 from sonatoki.Preprocessors import Preprocessor
 
+LOG = logging.getLogger(__name__)
+
 
 class Ilo:
     __preprocessors: List[Type[Preprocessor]]
@@ -17,7 +20,7 @@ class Ilo:
     __scorer: Type[Scorer]
     __tokenize: Tokenizer
     __passing_score: Number
-    debug: bool = False
+    logging_threshold: Number = 1.0
 
     def __init__(
         self,
@@ -83,19 +86,35 @@ class Ilo:
     def __score_tokens(self, tokens: List[str]) -> float:
         return self.__scorer.score(tokens, self.__scoring_filters)
 
-    def is_toki_pona(self, message: str) -> bool:
+    def _is_toki_pona(
+        self, message: str
+    ) -> Tuple[str, List[str], List[str], List[str], Number, bool]:
+        """Returns all components of the processing algorithm:
+        - Preprocessed message (str)
+        - Tokenized message (list[str])
+        - Filtered message (list[str])
+        - Cleaned message (list[str])
+        - Score (float)
+        - Result (bool)
+        """
         preprocessed = self.__preprocess(message)
         tokenized = self.__tokenize(preprocessed)
         filtered = self.__filter_tokens(tokenized)
         cleaned = self.__clean_tokens(filtered)
         score = self.__score_tokens(cleaned)
+        result = score >= self.__passing_score
 
-        if self.debug:
-            print("msg: %.2f  %s" % (score, repr(message)))
-            print("Preproc:   %s" % repr(preprocessed))
-            print("Tokenized: %s" % tokenized)
-            print("Filtered:  %s" % filtered)
-            print("Cleaned:   %s" % cleaned)
-            print()
+        # NOTE: this method may break if above funcs start sharing a list
+        if score <= self.logging_threshold:
+            LOG.debug("Msg: %.2f  %s", score, repr(message))
+            LOG.debug("Preproc:   %s", repr(preprocessed))
+            LOG.debug("Tokenized: %s", tokenized)
+            LOG.debug("Filtered:  %s", filtered)
+            LOG.debug("Cleaned:   %s", cleaned)
+        # TODO: Move to each function? Loses ability to control when logging occurs by threshold
 
-        return score >= self.__passing_score
+        return preprocessed, tokenized, filtered, cleaned, score, result
+
+    def is_toki_pona(self, message: str) -> bool:
+        *_, result = self._is_toki_pona(message)
+        return result

{sonatoki-0.1.0.dist-info → sonatoki-0.1.2.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: sonatoki
-Version: 0.1.0
+Version: 0.1.2
 Summary: ilo li moku e toki li pana e sona ni: ni li toki ala toki pona?
 Author-Email: "jan Kekan San (@gregdan3)" <gregory.danielson3@gmail.com>
 License: AGPL-3.0-or-later
@@ -44,7 +44,7 @@ from sonatoki.Filters import (
     ProperName,
     Punctuations,
 )
-from sonatoki.Scorers import Scaling
+from sonatoki.Scorers import SoftScaling
 from sonatoki.Cleaners import ConsecutiveDuplicates
 from sonatoki.Tokenizers import word_tokenize_tok
 from sonatoki.Preprocessors import URLs, DiscordEmotes
@@ -55,22 +55,23 @@ def main():
         ignoring_filters=[Numerics, Punctuations],
         scoring_filters=[NimiLinku, Syllabic, ProperName, Alphabetic],
         cleaners=[ConsecutiveDuplicates],
-        scorer=Scaling,
+        scorer=SoftScaling,
         tokenizer=word_tokenize_tok,
     )
     ilo.is_toki_pona("imagine how is touch the sky")  # False
     ilo.is_toki_pona("o pilin insa e ni: sina pilin e sewi")  # True
+    ilo.is_toki_pona("I Think I Can Evade Detection")  # False
 
 if __name__ == "__main__":
     main()
 ```
 
-`Ilo` is highly configurable by design, so I recommend exploring the `Preprocessors`, `Filters`, and `Scorers` modules. The `Cleaners` module only contains one cleaner, which I highly recommend. The `Tokenizers` module contains several other word tokenizers, but their performance will be worse than the
+`Ilo` is highly configurable by design, so I recommend exploring the `Preprocessors`, `Filters`, and `Scorers` modules. The `Cleaners` module only contains one cleaner, which I recommend using. The `Tokenizers` module contains several other word tokenizers, but their performance will be worse than the dedicated Toki Pona tokenizer `word_tokenize_tok`.
 
 ## Development
 
 1. Install [pdm](https://github.com/pdm-project/pdm)
-1. `pdm sync --dev`
+1. `pdm install --dev`
 1. Open any file you like!
 
 ## FAQ
@@ -81,4 +82,26 @@ The intent is to show our methodology to the Unicode Consortium, particularly to
 
 After our proposal has been examined and a result given by the committee, I will translate this file and library into Toki Pona, with a note left behind for those who do not understand it.
 
-### Why aren't any of the specific
+### What's the deal with the tokenizers?
+
+The Toki Pona tokenizer `word_tokenize_tok` is very specific in always separating writing characters from punctuation, and leaving contiguous punctuation as contiguous- this is a level of precision that NLTK's English tokenizer does not want for several reasons, such as that English words can have "punctuation" characters in them.
+
+Toki Pona doesn't have any mid-word symbols when rendered in the Latin alphabet, so a more aggressive tokenizer is highly desirable.
+
+The other tokenizers are provided as a comparison case more than anything. I do not recommend their use.
+
+### Aren't there a lot of false positives?
+
+Yes. It's up to you to use this tool responsibly on input you've done your best to clean, and better, use stronger filters before weaker ones. For now though, here's a list of relevant false positives:
+
+- `ProperName` will errantly match text in languages without a capital/lowercase distinction, artificially inflating the scores.
+- `Alphabetic` will match a _lot_ of undesirable text- it essentially allows 14 letters of the English alphabet.
+
+### Don't some of the cleaners/filters conflict?
+
+Yes. Some do so
+
+- `ConsecutiveDuplicates` may errantly change a word's validity. For example, "manna" is phonotactically invalid in Toki Pona, but would become "mana" which is valid.
+- `ConsecutiveDuplicates` will not work correctly with syllabaries (alphabets, but representing a pair of consonant and vowel).
+
+You'll notice a _lot_ of these are troubles regarding the application of latin alphabet filters to non-latin text. Working on it!

sonatoki-0.1.2.dist-info/RECORD

@@ -0,0 +1,14 @@
+sonatoki-0.1.2.dist-info/METADATA,sha256=3ccNKi5ODyxbkBvihKJ-XyXEDF5mzO0AzB0myhCMNMM,5020
+sonatoki-0.1.2.dist-info/WHEEL,sha256=vnE8JVcI2Wz7GRKorsPArnBdnW2SWKWGow5gu5tHlRU,90
+sonatoki-0.1.2.dist-info/licenses/LICENSE,sha256=DZak_2itbUtvHzD3E7GNUYSRK6jdOJ-GqncQ2weavLA,34523
+sonatoki/Cleaners.py,sha256=gTZ9dSsnvKVUtxM_ECSZ-_2heh--nD5A9dCQR1ATb1c,1160
+sonatoki/Filters.py,sha256=yzhYF79GX03cOwlR_-B8SPMQPZv4UpAPytH0fQwBE70,4093
+sonatoki/Preprocessors.py,sha256=OhjKcH_nlDznT1Y-ssdU1xH6_xbNrPDQEh8D7H3UTcs,3499
+sonatoki/Scorers.py,sha256=V293DBiupBiujzuc4yMrKOAiuNTLltIsiCzIAlLeokA,4129
+sonatoki/Tokenizers.py,sha256=lFvYS9SXij6gIk-CatKyFQtbRV0ska9FAgynEObxd-A,1741
+sonatoki/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+sonatoki/__main__.py,sha256=6xc-wIrrFo9wTyn4zRQNAmqwmJBtVvCMwV-CrM-hueA,82
+sonatoki/constants.py,sha256=h5rbCfu9YF76BsjQYud5d2wq1HODY05zOaw0Ir1cwjo,1320
+sonatoki/ilo.py,sha256=h3TYoqrjHxMGhB8ZJLVijVzy1AVCeWJk5x0q-bs4JMc,4278
+sonatoki/linku.json,sha256=MdFuFRIHniPDUVxKEKuUg1KyzPVgcCj4ZeyvburCwD0,270928
+sonatoki-0.1.2.dist-info/RECORD,,

sonatoki-0.1.0.dist-info/RECORD

@@ -1,14 +0,0 @@
-sonatoki-0.1.0.dist-info/METADATA,sha256=EQaB5tsicEQ4wYn5curehbhkzGF0qHqC1bnUbOVDCu0,3332
-sonatoki-0.1.0.dist-info/WHEEL,sha256=vnE8JVcI2Wz7GRKorsPArnBdnW2SWKWGow5gu5tHlRU,90
-sonatoki-0.1.0.dist-info/licenses/LICENSE,sha256=DZak_2itbUtvHzD3E7GNUYSRK6jdOJ-GqncQ2weavLA,34523
-sonatoki/Cleaners.py,sha256=gTZ9dSsnvKVUtxM_ECSZ-_2heh--nD5A9dCQR1ATb1c,1160
-sonatoki/Filters.py,sha256=yzhYF79GX03cOwlR_-B8SPMQPZv4UpAPytH0fQwBE70,4093
-sonatoki/Preprocessors.py,sha256=G2up2jKKSrHQtTQWYNWH_fkjgroL45ZeajVn1KUECt8,3431
-sonatoki/Scorers.py,sha256=X1vo-eIPbtl0IC5suIX6hu-4VG7NSzR90rkrLpep8WY,3690
-sonatoki/Tokenizers.py,sha256=epOG3jZHI3MSO_L_6Z3zsSkexDEMLVzA2ARg6EnPMO0,1628
-sonatoki/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-sonatoki/__main__.py,sha256=6xc-wIrrFo9wTyn4zRQNAmqwmJBtVvCMwV-CrM-hueA,82
-sonatoki/constants.py,sha256=h5rbCfu9YF76BsjQYud5d2wq1HODY05zOaw0Ir1cwjo,1320
-sonatoki/ilo.py,sha256=Uu0zipAF-L-5Wxw_EBB7-EMc40PM4WBa59Atq0zmYYE,3482
-sonatoki/linku.json,sha256=MdFuFRIHniPDUVxKEKuUg1KyzPVgcCj4ZeyvburCwD0,270928
-sonatoki-0.1.0.dist-info/RECORD,,