wiki:private/AdvancedNlpCourse/LanguageModelling

Version 16 (modified by Vít Baisa, 4 years ago) (diff)

--

Language modelling

IA161 Advanced NLP Course, Course Guarantee: Aleš Horák

Prepared by: Vít Baisa

State of the Art

The goal of a language model is a) to predict a following word or phrase based on a given history and b) to assign a probability (= score) to any possible input sentence. In the past, this was achieved mainly by n-gram models known since WWII. But recently, the buzzword deep learning penetrated also into language modelling and it turned out to be substantially better than Markov's n-gram models.

References

  1. Bengio, Yoshua, et al. "A neural probabilistic language model." The Journal of Machine Learning Research 3 (2003): 1137-1155.
  2. Mikolov, Tomas, et al. "Efficient estimation of word representations in vector space." arXiv preprint arXiv:1301.3781 (2013).
  3. Mikolov, Tomas, et al. "Distributed representations of words and phrases and their compositionality." Advances in Neural Information Processing Systems. 2013.
  4. Chelba, Ciprian, et al. "One billion word benchmark for measuring progress in statistical language modeling." arXiv preprint arXiv:1312.3005 (2013).

Practical Session

Bůh požádal, aby nedošlo k případnému provádění stavby pro klasické použití techniky ve výši stovek milionů korun.

We will build a simple character-based language model and generate naturally-looking sentences. We need a plain text and fast suffix sorting algorithm (mksary).

http://corpora.fi.muni.cz/cblm/generate.cgi

Getting necessary data and tools

  • wget nlp.fi.muni.cz/~xbaisa/cblm.tar
  • tar xf cblm.tar in your directory
  • cd cblm
  • get a large Czech model (3 GB):
    scp anxur:/tmp/cztenten.trie .
    or scp aurora:/corpora/data/cblm/data/cztenten.trie .
    (or download from a stable copy)
  • get a large plain text:
    scp anxur:/tmp/cztenten.txt.xz .

or scp aurora:/corpora/data/cblm/data/cztenten_1M_sentences.txt.xz .

mksary

  • git clone https://github.com/lh3/libdivsufsort.git
  • cd libdivsufsort
  • cmake -DCMAKE_BUILD_TYPE="Release" -DCMAKE_INSTALL_PREFIX="/ABSOLUTE_PATH_TO_LIBDIVSUFSORT"
  • make
  • cd ..
  • ln -s libdivsufsort/examples/mksary mksary

Training data

To build a new model, we need

  • a plain text file (suffix .in) all in lowercase: xz cztenten_1M_sentences.txt.xz | python lower.py > input.in
  • to create a suffix array ./mksary INPUT.in OUTPUT.sa
  • and compute the prefix tree: python build_trie.py FILE_PREFIX [MINFREQ]

The model will be stored in FILE_PREFIX.trie file.

Generating text

To generate a random text, just run python alarm.py FILE_PREFIX.trie

You may try to generate a random sentence using the large 3G model: python alarm.py cztenten.trie

Task

Change the training process (build_trie.py) and the generating process (alarm.py) to generate the most naturally-looking sentences. Either by

  • pre-processing the input plain text or
  • setting training parameters or
  • changing the generating process
  • or all above.

Upload 10,000 random sentences to your vault together with the amended scripts. Describe your changes and tunings in README file where you can put some hilarious random sentence examples.