compact_lang_det_impl.h

// Copyright 2013 Google Inc. All Rights Reserved.

//

// Licensed under the Apache License, Version 2.0 (the "License");

// you may not use this file except in compliance with the License.

// You may obtain a copy of the License at

//

//     http://www.apache.org/licenses/LICENSE-2.0

//

// Unless required by applicable law or agreed to in writing, software

// distributed under the License is distributed on an "AS IS" BASIS,

// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.

// See the License for the specific language governing permissions and

// limitations under the License.

//

// Author: dsites@google.com (Dick Sites)

//

#ifndef I18N_ENCODINGS_COMPACT_LANG_DET_COMPACT_LANG_DET_IMPL_H_

#define I18N_ENCODINGS_COMPACT_LANG_DET_COMPACT_LANG_DET_IMPL_H_

#include <vector>

#include "../public/compact_lang_det.h"   // For CLDHints, ResultChunkVector

#include "integral_types.h"

#include "lang_script.h"

namespace CLD2 {

// Internal use flags

static const int kCLDFlagFinish = 1;

static const int kCLDFlagSqueeze = 2;

static const int kCLDFlagRepeats = 4;

static const int kCLDFlagTop40 = 8;

static const int kCLDFlagShort = 16;

static const int kCLDFlagHint = 32;

static const int kCLDFlagUseWords = 64;

static const int kCLDFlagUNUSED = 128;

// Public use flags, debug output controls, defined in compact_lang_det.h

// 0x0100 and above

/***

Flag meanings:

Flags are used in the context of a recursive call from Detect to itself,

trying to deal in a more restrictive way with input that was not reliably

identified in the top-level call.

Finish -- Do not further recurse; return whatever result ensues, even if it is

          unreliable. Typically set in any recursive call to take a second try

          on unreliable text.

Squeeze -- For each text run, do an inplace cheapsqueeze to remove chunks of

          highly repetitive text and chunks of text with too many 1- and

          2-letter words. This avoids scoring repetitive or useless non-text

          crap in large files such bogus JPEGs within an HTML file.

Repeats -- When scoring a text run, do a cheap prediction of each character

          and do not score a unigram/quadgram if the last character of same is

          correctly predicted. This is a slower, finer-grained form of

          cheapsqueeze, typically used when the first pass got unreliable

          results.

Top40 -- Restrict the set of scored languages to the Google "Top 40", which is

          actually 38 languages. This gets rid of about 110 languages that

          represent about 0.7% of the web. Typically used when the first pass

          got unreliable results.

Short -- DEPRICATED, unused

Hint -- EXPERIMENTAL flag for compact_lang_det_test.cc to indicate a language

          hint supplied in parameter plus_one.

UseWords -- In additon to scoring quad/uni/nil-grams, score complete words

Tentative decision logic:

In the middle of first pass -- After 4KB of text, look at the front 256 bytes

          of every full 4KB buffer. If it compresses very well (say 3:1) or has

          lots of spaces (say 1 of every 4 bytes), assume that the input is

          large and contains lots of bogus non-text. Recurse, passing the

          Squeeze flag to strip out chunks of this non-text.

At the end of the first pass --

          If the top language is reliable and >= 70% of the document, return.

          Else if the top language is reliable and top+2nd >= say 94%, return.

          Else, either the top language is not reliable or there is a lot of

          other crap.

***/

  // Scan interchange-valid UTF-8 bytes and detect most likely language,

  // or set of languages.

//

  // Design goals:

  //   Skip over big stretches of HTML tags

  //   Able to return ranges of different languages

  //   Relatively small tables and relatively fast processing

  //   Thread safe

//

  typedef struct {

    int perscript_count;

    const Language* perscript_lang;

  } PerScriptPair;

  typedef struct {

    // Constants for hashing 4-7 byte quadgram to 32 bits

    const int kQuadHashB4Shift;

    const int kQuadHashB4bShift;

    const int kQuadHashB5Shift;

    const int kQuadHashB5bShift;

    // Constants for hashing 32 bits to kQuadKeyTable subscript/key

    const int kHashvalToSubShift;

    const uint32 kHashvalToSubMask;

    const int kHashvalToKeyShift;

    const uint32 kHashvalToKeyMask;

    const int kHashvalAssociativity;

    // Pointers to the actual tables

    const PerScriptPair* kPerScriptPair;

    const uint16* kQuadKeyTable;

    const uint32* kQuadValueTable;

  } LangDetObj;

  // For HTML documents, tags are skipped, along with <script> ... </script>

  // and <style> ... </style> sequences, and entities are expanded.

//

  // We distinguish between bytes of the raw input buffer and bytes of non-tag

  // text letters. Since tags can be over 50% of the bytes of an HTML Page,

  // and are nearly all seven-bit ASCII English, we prefer to distinguish

  // language mixture fractions based on just the non-tag text.

//

  // Inputs: text and text_length

  //  is_plain_text if true says to NOT parse/skip HTML tags nor entities

  // Outputs:

  //  language3 is an array of the top 3 languages or UNKNOWN_LANGUAGE

  //  percent3 is an array of the text percentages 0..100 of the top 3 languages

  //  normalized_score3 is an array of internal scores, normalized to the

  //    average score for each language over a body of training text. A

  //    normalized score significantly away from 1.0 indicates very skewed text

  //    or gibberish.

//

  //  text_bytes is the amount of non-tag/letters-only text found

  //  is_reliable set true if the returned Language is at least 2**30 times more

  //  probable then the second-best Language

//

  // Return value: the most likely Language for the majority of the input text

  //  Length 0 input and text with no reliable letter sequences returns

  //  UNKNOWN_LANGUAGE

//

  // Subsetting: For fast detection over large documents, these routines will

  // only scan up to a fixed limit (currently 160KB of non-tag letters).

//

  Language DetectLanguageSummaryV2(

                        const char* buffer,

                        int buffer_length,

                        bool is_plain_text,

                        const CLDHints* cld_hints,

                        bool allow_extended_lang,

                        int flags,

                        Language plus_one,

                        Language* language3,

                        int* percent3,

                        double* normalized_score3,

                        ResultChunkVector* resultchunkvector,

                        int* text_bytes,

                        bool* is_reliable);

  // For unit testing:

  // Remove portions of text that have a high density of spaces, or that are

  // overly repetitive, squeezing the remaining text in-place to the front

  // of the input buffer.

  // Return the new, possibly-shorter length

  int CheapSqueezeInplace(char* isrc, int srclen, int ichunksize);

}       // End namespace CLD2

#endif  // I18N_ENCODINGS_COMPACT_LANG_DET_COMPACT_LANG_DET_IMPL_H_