Class LmReaders

java.lang.Object
edu.berkeley.nlp.lm.io.LmReaders

public class LmReaders extends Object
This class contains a number of static methods for reading/writing/estimating n-gram language models. Since most uses of this software will use this class, I will use this space to document the software as a whole.

This software provides three main pieces of functionality:
(a) estimation of a language models from text inputs
(b) data structures for efficiently storing large collections of n-grams in memory
(c) an API for efficient querying language models derived from n-gram collections. Most of the techniques in the paper are described in "Faster and Smaller N-gram Language Models" (Pauls and Klein 2011).

This software supports the estimation of two types of language models: Kneser-Ney language models (Kneser and Ney, 1995) and Stupid Backoff language models (Brants et al. 2007). Kneser-Ney language models can be estimated from raw text by calling createKneserNeyLmFromTextFiles(List, WordIndexer, int, File, ConfigOptions). This can also be done from the command-line by calling main() in MakeKneserNeyArpaFromText. See the examples folder for a script which demonstrates its use. A Stupid Backoff language model can be read from a directory containing n-gram counts in the format used by Google's Web1T corpus by calling readLmFromGoogleNgramDir(String, boolean, boolean). Note that this software does not (yet) support building Google count directories from raw text, though this can be done using SRILM.

Loading/estimating language models from text files can be very slow. This software can use Java's built-in serialization to build language model binaries which are both smaller and faster to load. MakeLmBinaryFromArpa and MakeLmBinaryFromGoogle provide main() methods for doing this. See the examples folder for scripts which demonstrate their use.

Language models can be read into memory from ARPA formats using readArrayEncodedLmFromArpa(String, boolean) and readContextEncodedLmFromArpa(String). The "array encoding" versus "context encoding" distinction is discussed in Section 4.2 of Pauls and Klein (2011). Again, since loading language models from textual representations can be very slow, they can be read from binaries using readLmBinary(String). The interfaces for these language models can be found in ArrayEncodedNgramLanguageModel and ContextEncodedNgramLanguageModel. For examples of these interfaces in action, you can have a look at

invalid reference
PerplexityTest
.

We implement the HASH,HASH+SCROLL, and COMPRESSED language model representations described in Pauls and Klein (2011) in this release. The SORTED implementation may be added later. See HashNgramMap and CompressedNgramMap for the implementations of the HASH and COMPRESSED representations.

To speed up queries, you can wrap language models with caches ( ContextEncodedCachingLmWrapper and ArrayEncodedCachingLmWrapper). These caches are described in section 4.1 of Pauls and Klein (2011). You should more or less always use these caches, since they are faster and have modest memory requirements.

This software also support a java Map wrapper around an n-gram collection. You can read a map wrapper using readNgramMapFromGoogleNgramDir(String, boolean, WordIndexer).

ComputeLogProbabilityOfTextStream provides a main() method for computing the log probability of raw text.

Some example scripts can be found in the examples/ directory.

Author:
adampauls
  • Constructor Details

    • LmReaders

      public LmReaders()
  • Method Details

    • readContextEncodedLmFromArpa

      public static ContextEncodedProbBackoffLm<String> readContextEncodedLmFromArpa(String lmFile)
    • readContextEncodedLmFromArpa

      public static <W> ContextEncodedProbBackoffLm<W> readContextEncodedLmFromArpa(String lmFile, WordIndexer<W> wordIndexer)
    • readContextEncodedLmFromArpa

      public static <W> ContextEncodedProbBackoffLm<W> readContextEncodedLmFromArpa(String lmFile, WordIndexer<W> wordIndexer, ConfigOptions opts, int lmOrder)
      Reads a context-encoded language model from an ARPA lm file. Context-encoded language models allow faster queries, but require an extra 4-bytes of storage per n-gram for the suffix offsets (as compared to array-encoded language models).
      Type Parameters:
      W -
      Parameters:
      lmFile -
      wordIndexer -
      opts -
      lmOrder -
      compress -
      Returns:
    • readContextEncodedLmFromArpa

      public static <W> ContextEncodedProbBackoffLm<W> readContextEncodedLmFromArpa(LmReader<ProbBackoffPair,ArpaLmReaderCallback<ProbBackoffPair>> lmFile, WordIndexer<W> wordIndexer, ConfigOptions opts)
    • readArrayEncodedLmFromArpa

      public static ArrayEncodedProbBackoffLm<String> readArrayEncodedLmFromArpa(String lmFile, boolean compress)
    • readArrayEncodedLmFromArpa

      public static <W> ArrayEncodedProbBackoffLm<W> readArrayEncodedLmFromArpa(String lmFile, boolean compress, WordIndexer<W> wordIndexer)
    • readArrayEncodedLmFromArpa

      public static <W> ArrayEncodedProbBackoffLm<W> readArrayEncodedLmFromArpa(String lmFile, boolean compress, WordIndexer<W> wordIndexer, ConfigOptions opts, int lmOrder)
    • readArrayEncodedLmFromArpa

      public static <W> ArrayEncodedProbBackoffLm<W> readArrayEncodedLmFromArpa(LmReader<ProbBackoffPair,ArpaLmReaderCallback<ProbBackoffPair>> lmFile, boolean compress, WordIndexer<W> wordIndexer, ConfigOptions opts)
      Reads an array-encoded language model from an ARPA lm file.
      Type Parameters:
      W -
      Parameters:
      lmFile -
      compress - Compress the LM using block compression. This LM should be smaller but slower.
      wordIndexer -
      opts -
      lmOrder -
      Returns:
    • readNgramMapFromGoogleNgramDir

      public static NgramMapWrapper<String,LongRef> readNgramMapFromGoogleNgramDir(String dir, boolean compress)
    • readNgramMapFromGoogleNgramDir

      public static <W> NgramMapWrapper<W,LongRef> readNgramMapFromGoogleNgramDir(String dir, boolean compress, WordIndexer<W> wordIndexer)
    • readNgramMapFromBinary

      public static NgramMapWrapper<String,LongRef> readNgramMapFromBinary(String binary, String vocabFile)
    • readNgramMapFromBinary

      public static <W> NgramMapWrapper<W,LongRef> readNgramMapFromBinary(String binary, String sortedVocabFile, WordIndexer<W> wordIndexer)
      Parameters:
      sortedVocabFile - should be the vocab_cs.gz file from the Google n-gram corpus.
      Returns:
    • readLmFromGoogleNgramDir

      public static ArrayEncodedNgramLanguageModel<String> readLmFromGoogleNgramDir(String dir, boolean compress, boolean kneserNey)
    • readLmFromGoogleNgramDir

      public static <W> ArrayEncodedNgramLanguageModel<W> readLmFromGoogleNgramDir(String dir, boolean compress, boolean kneserNey, WordIndexer<W> wordIndexer, ConfigOptions opts)
      Reads a stupid backoff lm from a directory with n-gram counts in the format used by Google n-grams.
      Type Parameters:
      W -
      Parameters:
      dir -
      compress -
      wordIndexer -
      opts -
      Returns:
    • readContextEncodedKneserNeyLmFromTextFile

      public static <W> ContextEncodedProbBackoffLm<W> readContextEncodedKneserNeyLmFromTextFile(List<String> files, WordIndexer<W> wordIndexer, int lmOrder, ConfigOptions opts)
      Builds a context-encoded LM from raw text. This call first builds and writes a (temporary) ARPA file by calling
      invalid reference
      #createKneserNeyLmFromTextFiles(List, WordIndexer, int, File)
      , and the reads the resulting file. Since the temp file can be quite large, it is important that the temp directory used by java ( java.io.tmpdir).
      Type Parameters:
      W -
      Parameters:
      files -
      wordIndexer -
      lmOrder -
      opts -
      Returns:
    • readKneserNeyLmFromTextFile

      public static <W> ArrayEncodedProbBackoffLm<W> readKneserNeyLmFromTextFile(List<String> files, WordIndexer<W> wordIndexer, int lmOrder, ConfigOptions opts, boolean compress)
      Builds an array-encoded LM from raw text. This call first builds and writes a (temporary) ARPA file by calling
      invalid reference
      #createKneserNeyLmFromTextFiles(List, WordIndexer, int, File)
      , and the reads the resulting file. Since the temp file can be quite large, it is important that the temp directory used by java ( java.io.tmpdir).
      Type Parameters:
      W -
      Parameters:
      files -
      wordIndexer -
      lmOrder -
      opts -
      Returns:
    • readContextEncodedKneserNeyLmFromTextFile

      public static <W> ContextEncodedProbBackoffLm<W> readContextEncodedKneserNeyLmFromTextFile(List<String> files, WordIndexer<W> wordIndexer, int lmOrder, ConfigOptions opts, File tmpFile)
    • readKneserNeyLmFromTextFile

      public static <W> ArrayEncodedProbBackoffLm<W> readKneserNeyLmFromTextFile(List<String> files, WordIndexer<W> wordIndexer, int lmOrder, boolean compress, ConfigOptions opts, File tmpFile)
    • createKneserNeyLmFromTextFiles

      public static <W> void createKneserNeyLmFromTextFiles(List<String> files, WordIndexer<W> wordIndexer, int lmOrder, File arpaOutputFile, ConfigOptions opts)
      Estimates a Kneser-Ney language model from raw text, and writes a file (in ARPA format). Probabilities are in log base 10 to match SRILM.
      Type Parameters:
      W -
      Parameters:
      files - Files of raw text (new-line separated).
      wordIndexer -
      lmOrder -
      arpaOutputFile -
    • readGoogleLmBinary

      public static StupidBackoffLm<String> readGoogleLmBinary(String file, String sortedVocabFile)
    • readGoogleLmBinary

      public static <W> StupidBackoffLm<W> readGoogleLmBinary(String file, WordIndexer<W> wordIndexer, String sortedVocabFile)
      Reads in a pre-built Google n-gram binary. The user must supply the vocab_cs.gz file (so that the corpus cannot be reproduced unless the user has the rights to do so).
      Type Parameters:
      W -
      Parameters:
      file - The binary
      wordIndexer -
      sortedVocabFile - the vocab_cs.gz vocabulary file.
      Returns:
    • readLmBinary

      public static <W> NgramLanguageModel<W> readLmBinary(String file)
      Reads a binary file representing an LM. These will need to be cast down to either ContextEncodedNgramLanguageModel or ArrayEncodedNgramLanguageModel to be useful.
    • writeLmBinary

      public static <W> void writeLmBinary(NgramLanguageModel<W> lm, String file)
      Writes a binary file representing the LM using the built-in serialization. These binaries should load much faster than ARPA files.
      Type Parameters:
      W -
      Parameters:
      lm -
      file -