MeTA v3.0.0 Release Notes

Release Date: 2017-02-13 // about 4 years ago
  • ๐Ÿ†• New features

    โž• Add an embedding_analyzer that represents documents with their averaged word vectors.

    โž• Add a parallel::reduction algorithm designed for parallelizing complex accumulation operations (like an E step in an EM algorithm)

    Parallelize feature counting in feature selector using the new parallel::reduction

    Add a parallel::for_each_block algorithm to run functions on (relatively) equal sub-ranges of an iterator range in parallel

    โž• Add a parallel merge sort as parallel::sort

    โž• Add a util/traits.h header for general useful traits

    โž• Add a Markov model implementation in sequence::markov_model

    โž• Add a generic unsupervised HMM implementation. This implementation supports HMMs with discrete observations (what is used most often) and sequence observations (useful for log mining applications). The forward-backward algorithm is implemented using both the scaling method and the log-space method. The scaling method is used by default, but the log-space method is useful for HMMs with sequence observations to avoid underflow issues when the output probabilities themselves are very small.

    Add the KL-divergence retrieval function using pseudo-relevance feedback with the two-component mixture-model approach of Zhai and Lafferty, called kl_divergence_prf. This ranker internally can use any language_model_ranker subclass like dirichlet_prior or jelinek_mercer to perform the ranking of the feedback set and the result documents with respect to the modified query.

    ๐Ÿ†“ The EM algorithm used for the two-component mixture model is provided as the index::feedback::unigram_mixture free function and returns the feedback model.

    โž• Add the Rocchio algorithm (rocchio) for pseudo-relevance feedback in the vector space model.

    ๐Ÿ’ฅ Breaking Change. To facilitate the above to changes, we have also broken the ranker hierarchy into one more level. At the top we have ranker, which has a pure virtual function rank() that can be overridden to provide entirely custom ranking behavior. This is the class the KL-divergence and Rocchio methods derive from, as we need to re-define what it means to rank documents (first retrieving a feedback set, then ranking documents with respect to an updated query).

    Most of the time, however, you will want to derive from the second level ranking_function, which is what was called ranker before. This class provides a definition of rank() to perform document-at-a-time ranking, and expects deriving classes to instead provide initial_score() and score_one() implementations to define the scoring function used for each document. Existing code that derived from ranker prior to this version of MeTA likely needs to be changed to instead derive from ranking_function.

    Add the util::transform_iterator class and util::make_transform_iterator function for providing iterators that transform their output according to a unary function.

    ๐Ÿ’ฅ Breaking Change. whitespace_tokenizer now emits only word tokens by default, suppressing all whitespace tokens. The old default was to emit tokens containing whitespace in addition to actual word tokens. The old behavior can be obtained by passing false to its constructor, or setting suppress-whitespace = false in its configuration group in config.toml. (Note that whitespace tokens are still needed if using a sentence_boundary filter but, in nearly all circumstances, icu_tokenizer should be preferred.)

    ๐Ÿ’ฅ Breaking Change. Co-occurrence counting for embeddings now uses history that crosses sentence boundaries by default. The old behavior (clearing the history when starting a new sentence) can be obtained by ensuring that a tokenizer is being used that emits sentence boundary tags and by setting break-on-tags = true in the [embeddings] table of config.toml.

    ๐Ÿ’ฅ Breaking Change. All references in the embeddings library to "coocur" are have changed to "cooccur". This means that some files and binaries have been renamed. Much of the co-occurrence counting part of the embeddings library has also been moved to the public API.

    ๐Ÿ”ง Co-occurrence counting now is performed in parallel. Behavior of its merge strategy can be configured with the new [embeddings] config parameter merge-fanout = n, which specifies the maximum number of on-disk chunks to allow before kicking off a multi-way merge (default 8).

    โœจ Enhancements

    • Add additional packed_write and packed_read overloads: for std::pair, stats::dirichlet, stats::multinomial, util::dense_matrix, and util::sparse_vector
    • Additional functions have been added to ranker_factory to allow construction/loading of language_model_ranker subclasses (useful for the kl_divergence_prf implementation)
    • ๐Ÿ›  Add a util::make_fixed_heap helper function to simplify the declaration of util::fixed_heap classes with lambda function comparators.
    • โž• Add regression tests for rankers MAP and NDCG scores. This adds a new dataset cranfield that contains non-binary relevance judgments to facilitate these new tests.
    • โฌ†๏ธ Bump bundled version of ICU to 58.2.

    ๐Ÿ› Bug Fixes

    • ๐Ÿ›  Fix bug in NDCG calculation (ideal-DCG was computed using the wrong sorting order for non-binary judgments)
    • ๐Ÿ›  Fix bug where the final chunks to be merged in index creation were not being deleted when merging completed
    • ๐Ÿ›  Fix bug where GloVe training would allocate the embedding matrix before starting the shuffling process, causing it to exceed the "max-ram" config parameter.
    • ๐Ÿ›  Fix bug with consuming MeTA from a build directory with cmake when building a static ICU library. meta-utf is now forced to be a shared library, which (1) should save on binary sizes and (2) ensures that the statically build ICU is linked into the library to avoid undefined references to ICU functions.
    • ๐Ÿ›  Fix bug with consuming Release-mode MeTA libraries from another project being built in Debug mode. Before, identifiers.h would change behavior based on the NDEBUG macro's setting. This behavior has been removed, and opaque identifiers are always on.

    ๐Ÿ—„ Deprecation

    • disk_index::doc_name and disk_index::doc_path have been deprecated in
      ๐Ÿš€ favor of the more general (and less confusing) metadata(). They will be removed in a future major release.
    • ๐Ÿ‘Œ Support for 32-bit architectures is provided on a best-effort basis. MeTA makes heavy use of memory mapping, which is best paired with a 64-bit address space. Please move to a 64-bit platform for using MeTA if at all possible (most consumer machines should support 64-bit if they were made in the last 5 years or so).

    Model File Checksums (sha256)

    d29bf8b4cbeef21db087cf8042efe5afe25c7bd3c460997728d58b92c24ec283 beam-search-constituency-parser-4.tar.gz
    ce44c7d96a8339ff4b597f35a35534ccf93ab99b7d45cbbdddffe7e362b9c20e crf.tar.gz
    672b10c398c1a193ba91dc8c0493d729ad3f73d9192ef33100baeb8afd4f5cde gigaword-embeddings-50d.tar.gz
    40cd87901eb29b69e57e4bca14bc2539d7d6b4ad5c186d6f3b1532a60c5163b0 greedy-constituency-parser.tar.gz
    a0a3814c1f82780f1296d600eba260f474420aa2d93f000e390c71a0ddac42d9 greedy-perceptron-tagger.tar.gz

    Please note that the embeddings model has changed. Please re-download.