translation_matrix.py

#!/usr/bin/env python
# encoding: utf-8

"""Produce a translation matrix to translate words from one language to another, using either
a standard nearest neighbour method or a globally corrected neighbour retrieval method [1]_.

This method can be used to augment the existing phrase tables with more candidate translations, or
filter out errors from the translation tables and known dictionaries [2]_. What's more, It also work
for any two sets of named-vectors where there are some paired-guideposts to learn the transformation.

Examples
--------

How to make translation between two set of word-vectors
=======================================================

Initialize a word-vector models

.. sourcecode:: pycon

    >>> from gensim.models import KeyedVectors
    >>> from gensim.test.utils import datapath
    >>>
    >>> model_en = KeyedVectors.load_word2vec_format(datapath("EN.1-10.cbow1_wind5_hs0_neg10_size300_smpl1e-05.txt"))
    >>> model_it = KeyedVectors.load_word2vec_format(datapath("IT.1-10.cbow1_wind5_hs0_neg10_size300_smpl1e-05.txt"))

Define word pairs (that will be used for construction of translation matrix

.. sourcecode:: pycon

    >>> word_pairs = [
    ...     ("one", "uno"), ("two", "due"), ("three", "tre"), ("four", "quattro"), ("five", "cinque"),
    ...     ("seven", "sette"), ("eight", "otto"),
    ...     ("dog", "cane"), ("pig", "maiale"), ("fish", "cavallo"), ("birds", "uccelli"),
    ...     ("apple", "mela"), ("orange", "arancione"), ("grape", "acino"), ("banana", "banana")
    ... ]

Fit :class:`~gensim.models.translation_matrix.TranslationMatrix`

.. sourcecode:: pycon

    >>> trans_model = TranslationMatrix(model_en, model_it, word_pairs=word_pairs)

Apply model (translate words "dog" and "one")

.. sourcecode:: pycon

    >>> trans_model.translate(["dog", "one"], topn=3)
    OrderedDict([('dog', [u'cane', u'gatto', u'cavallo']), ('one', [u'uno', u'due', u'tre'])])


Save / load model

.. sourcecode:: pycon

    >>> with temporary_file("model_file") as fname:
    ...     trans_model.save(fname)  # save model to file
    ...     loaded_trans_model = TranslationMatrix.load(fname)  # load model


How to make translation between two :class:`~gensim.models.doc2vec.Doc2Vec` models
==================================================================================

Prepare data and models

.. sourcecode:: pycon

    >>> from gensim.test.utils import datapath
    >>> from gensim.test.test_translation_matrix import read_sentiment_docs
    >>> from gensim.models import Doc2Vec
    >>>
    >>> data = read_sentiment_docs(datapath("alldata-id-10.txt"))[:5]
    >>> src_model = Doc2Vec.load(datapath("small_tag_doc_5_iter50"))
    >>> dst_model = Doc2Vec.load(datapath("large_tag_doc_10_iter50"))

Train backward translation

.. sourcecode:: pycon

    >>> model_trans = BackMappingTranslationMatrix(data, src_model, dst_model)
    >>> trans_matrix = model_trans.train(data)


Apply model

.. sourcecode:: pycon

    >>> result = model_trans.infer_vector(dst_model.dv[data[3].tags])


References
----------
.. [1] Dinu, Georgiana, Angeliki Lazaridou, and Marco Baroni. "Improving zero-shot learning by mitigating the
       hubness problem", https://arxiv.org/abs/1412.6568
.. [2] Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg Corrado, and Jeffrey Dean.
       "Distributed Representations of Words and Phrases and their Compositionality", https://arxiv.org/abs/1310.4546

"""

import warnings
from collections import OrderedDict

import numpy as np

from gensim import utils


class Space:
    """An auxiliary class for storing the the words space."""

    def __init__(self, matrix, index2word):
        """
        Parameters
        ----------
        matrix : iterable of numpy.ndarray
            Matrix that contains word-vectors.
        index2word : list of str
            Words which correspond to the `matrix`.

        """
        self.mat = matrix
        self.index2word = index2word

        # build a dict to map word to index
        self.word2index = {}
        for idx, word in enumerate(self.index2word):
            self.word2index[word] = idx

    @classmethod
    def build(cls, lang_vec, lexicon=None):
        """Construct a space class for the lexicon, if it's provided.

        Parameters
        ----------
        lang_vec : :class:`~gensim.models.keyedvectors.KeyedVectors`
            Model from which the vectors will be extracted.
        lexicon : list of str, optional
            Words which contains in the `lang_vec`, if `lexicon = None`, the lexicon is all the lang_vec's word.

        Returns
        -------
        :class:`~gensim.models.translation_matrix.Space`
            Object that stored word-vectors

        """
        # `words` to store all the word that
        # `mat` to store all the word vector for the word in 'words' list
        words = []
        mat = []
        if lexicon is not None:
            # if the lexicon is not provided, using the all the Keyedvectors's words as default
            for item in lexicon:
                words.append(item)
                mat.append(lang_vec.vectors[lang_vec.get_index(item)])

        else:
            for item in lang_vec.index_to_key:
                words.append(item)
                mat.append(lang_vec.vectors[lang_vec.get_index(item)])

        return Space(mat, words)

    def normalize(self):
        """Normalize the word vector's matrix."""
        self.mat = self.mat / np.sqrt(np.sum(np.multiply(self.mat, self.mat), axis=1, keepdims=True))


class TranslationMatrix(utils.SaveLoad):
    """Objects of this class realize the translation matrix which map the source language to the target language.
    The main methods are:

    We map it to the other language space by computing z = Wx, then return the
    word whose representation is close to z.

    The details use seen the notebook [3]_

    Examples
    --------
    .. sourcecode:: pycon

        >>> from gensim.models import KeyedVectors
        >>> from gensim.test.utils import datapath
        >>> en = datapath("EN.1-10.cbow1_wind5_hs0_neg10_size300_smpl1e-05.txt")
        >>> it = datapath("IT.1-10.cbow1_wind5_hs0_neg10_size300_smpl1e-05.txt")
        >>> model_en = KeyedVectors.load_word2vec_format(en)
        >>> model_it = KeyedVectors.load_word2vec_format(it)
        >>>
        >>> word_pairs = [
        ...     ("one", "uno"), ("two", "due"), ("three", "tre"), ("four", "quattro"), ("five", "cinque"),
        ...     ("seven", "sette"), ("eight", "otto"),
        ...     ("dog", "cane"), ("pig", "maiale"), ("fish", "cavallo"), ("birds", "uccelli"),
        ...     ("apple", "mela"), ("orange", "arancione"), ("grape", "acino"), ("banana", "banana")
        ... ]
        >>>
        >>> trans_model = TranslationMatrix(model_en, model_it)
        >>> trans_model.train(word_pairs)
        >>> trans_model.translate(["dog", "one"], topn=3)
        OrderedDict([('dog', [u'cane', u'gatto', u'cavallo']), ('one', [u'uno', u'due', u'tre'])])


    References
    ----------
    .. [3] https://github.com/RaRe-Technologies/gensim/blob/3.2.0/docs/notebooks/translation_matrix.ipynb

    """
    def __init__(self, source_lang_vec, target_lang_vec, word_pairs=None, random_state=None):
        """
        Parameters
        ----------
        source_lang_vec : :class:`~gensim.models.keyedvectors.KeyedVectors`
            Word vectors for source language.
        target_lang_vec : :class:`~gensim.models.keyedvectors.KeyedVectors`
            Word vectors for target language.
        word_pairs : list of (str, str), optional
            Pairs of words that will be used for training.
        random_state : {None, int, array_like}, optional
            Seed for random state.

        """

        self.source_word = None
        self.target_word = None
        self.source_lang_vec = source_lang_vec
        self.target_lang_vec = target_lang_vec

        self.random_state = utils.get_random_state(random_state)
        self.translation_matrix = None
        self.source_space = None
        self.target_space = None

        if word_pairs is not None:
            if len(word_pairs[0]) != 2:
                raise ValueError("Each training data item must contain two different language words.")
            self.train(word_pairs)

    def train(self, word_pairs):
        """Build the translation matrix that mapping from source space to target space.

        Parameters
        ----------
        word_pairs : list of (str, str), optional
            Pairs of words that will be used for training.

        """
        self.source_word, self.target_word = zip(*word_pairs)

        self.source_space = Space.build(self.source_lang_vec, set(self.source_word))
        self.target_space = Space.build(self.target_lang_vec, set(self.target_word))

        self.source_space.normalize()
        self.target_space.normalize()

        m1 = self.source_space.mat[[self.source_space.word2index[item] for item in self.source_word], :]
        m2 = self.target_space.mat[[self.target_space.word2index[item] for item in self.target_word], :]

        self.translation_matrix = np.linalg.lstsq(m1, m2, -1)[0]

    def save(self, *args, **kwargs):
        """Save the model to a file. Ignores (doesn't store) the `source_space` and `target_space` attributes."""
        kwargs['ignore'] = kwargs.get('ignore', ['source_space', 'target_space'])
        super(TranslationMatrix, self).save(*args, **kwargs)

    def apply_transmat(self, words_space):
        """Map the source word vector to the target word vector using translation matrix.

        Parameters
        ----------
        words_space : :class:`~gensim.models.translation_matrix.Space`
            `Space` object constructed for the words to be translated.

        Returns
        -------
        :class:`~gensim.models.translation_matrix.Space`
            `Space` object constructed for the mapped words.

        """
        return Space(np.dot(words_space.mat, self.translation_matrix), words_space.index2word)

    def translate(self, source_words, topn=5, gc=0, sample_num=None, source_lang_vec=None, target_lang_vec=None):
        """Translate the word from the source language to the target language.

        Parameters
        ----------
        source_words : {str, list of str}
            Single word or a list of words to be translated
        topn : int, optional
            Number of words that will be returned as translation for each `source_words`
        gc : int, optional
            Define translation algorithm, if `gc == 0` - use standard NN retrieval,
            otherwise, use globally corrected neighbour retrieval method (as described in [1]_).
        sample_num : int, optional
            Number of word to sample from the source lexicon, if `gc == 1`, then `sample_num` **must** be provided.
        source_lang_vec : :class:`~gensim.models.keyedvectors.KeyedVectors`, optional
            New source language vectors for translation, by default, used the model's source language vector.
        target_lang_vec : :class:`~gensim.models.keyedvectors.KeyedVectors`, optional
            New target language vectors for translation, by default, used the model's target language vector.

        Returns
        -------
        :class:`collections.OrderedDict`
            Ordered dict where each item is `word`: [`translated_word_1`, `translated_word_2`, ...]

        """
        if isinstance(source_words, str):
            # pass only one word to translate
            source_words = [source_words]

        # If the language word vector not provided by user, use the model's
        # language word vector as default
        if source_lang_vec is None:
            warnings.warn(
                "The parameter source_lang_vec isn't specified, "
                "use the model's source language word vector as default."
            )
            source_lang_vec = self.source_lang_vec

        if target_lang_vec is None:
            warnings.warn(
                "The parameter target_lang_vec isn't specified, "
                "use the model's target language word vector as default."
            )
            target_lang_vec = self.target_lang_vec

        # If additional is provided, bootstrapping vocabulary from the source language word vector model.
        if gc:
            if sample_num is None:
                raise RuntimeError(
                    "When using the globally corrected neighbour retrieval method, "
                    "the `sample_num` parameter(i.e. the number of words sampled from source space) must be provided."
                )
            lexicon = set(source_lang_vec.index_to_key)
            addition = min(sample_num, len(lexicon) - len(source_words))
            lexicon = self.random_state.choice(list(lexicon.difference(source_words)), addition)
            source_space = Space.build(source_lang_vec, set(source_words).union(set(lexicon)))
        else:
            source_space = Space.build(source_lang_vec, source_words)
        target_space = Space.build(target_lang_vec, )

        # Normalize the source vector and target vector
        source_space.normalize()
        target_space.normalize()

        # Map the source language to the target language
        mapped_source_space = self.apply_transmat(source_space)

        # Use the cosine similarity metric
        sim_matrix = -np.dot(target_space.mat, mapped_source_space.mat.T)

        # If `gc=1`, using corrected retrieval method
        if gc:
            srtd_idx = np.argsort(np.argsort(sim_matrix, axis=1), axis=1)
            sim_matrix_idx = np.argsort(srtd_idx + sim_matrix, axis=0)
        else:
            sim_matrix_idx = np.argsort(sim_matrix, axis=0)

        # Translate the words and for each word return the `topn` similar words
        translated_word = OrderedDict()
        for idx, word in enumerate(source_words):
            translated_target_word = []
            # Search the most `topn` similar words
            for j in range(topn):
                map_space_id = sim_matrix_idx[j, source_space.word2index[word]]
                translated_target_word.append(target_space.index2word[map_space_id])
            translated_word[word] = translated_target_word
        return translated_word


class BackMappingTranslationMatrix(utils.SaveLoad):
    """Realize the BackMapping translation matrix which map the source model's document vector
    to the target model's document vector(old model).

    BackMapping translation matrix is used to learn a mapping for two document vector space which we
    specify as source document vector and target document vector. The target document vector are trained
    on superset corpus of source document vector, we can incrementally increase the vector in
    the old model through the BackMapping translation matrix.

    the details use seen the notebook [3]_.

    Examples
    --------
    .. sourcecode:: pycon

        >>> from gensim.test.utils import datapath
        >>> from gensim.test.test_translation_matrix import read_sentiment_docs
        >>> from gensim.models import Doc2Vec, BackMappingTranslationMatrix
        >>>
        >>> data = read_sentiment_docs(datapath("alldata-id-10.txt"))[:5]
        >>> src_model = Doc2Vec.load(datapath("small_tag_doc_5_iter50"))
        >>> dst_model = Doc2Vec.load(datapath("large_tag_doc_10_iter50"))
        >>>
        >>> model_trans = BackMappingTranslationMatrix(src_model, dst_model)
        >>> trans_matrix = model_trans.train(data)
        >>>
        >>> result = model_trans.infer_vector(dst_model.dv[data[3].tags])

    """
    def __init__(self, source_lang_vec, target_lang_vec, tagged_docs=None, random_state=None):
        """

        Parameters
        ----------
        source_lang_vec : :class:`~gensim.models.doc2vec.Doc2Vec`
            Source Doc2Vec model.
        target_lang_vec : :class:`~gensim.models.doc2vec.Doc2Vec`
            Target Doc2Vec model.
        tagged_docs : list of :class:`~gensim.models.doc2vec.TaggedDocument`, optional.
            Documents that will be used for training, both the source language document vector and
            target language document vector trained on those tagged documents.
        random_state : {None, int, array_like}, optional
            Seed for random state.

        """
        self.tagged_docs = tagged_docs
        self.source_lang_vec = source_lang_vec
        self.target_lang_vec = target_lang_vec

        self.random_state = utils.get_random_state(random_state)
        self.translation_matrix = None

        if tagged_docs is not None:
            self.train(tagged_docs)

    def train(self, tagged_docs):
        """Build the translation matrix that mapping from the source model's vector to target model's vector

        Parameters
        ----------
        tagged_docs : list of :class:`~gensim.models.doc2vec.TaggedDocument`, Documents
            that will be used for training, both the source language document vector and
            target language document vector trained on those tagged documents.

        Returns
        -------
        numpy.ndarray
            Translation matrix that mapping from the source model's vector to target model's vector.

        """
        m1 = [self.source_lang_vec.dv[item.tags].flatten() for item in tagged_docs]
        m2 = [self.target_lang_vec.dv[item.tags].flatten() for item in tagged_docs]

        self.translation_matrix = np.linalg.lstsq(m2, m1, -1)[0]
        return self.translation_matrix

    def infer_vector(self, target_doc_vec):
        """Translate the target model's document vector to the source model's document vector

        Parameters
        ----------
        target_doc_vec : numpy.ndarray
            Document vector from the target document, whose document are not in the source model.

        Returns
        -------
        numpy.ndarray
            Vector `target_doc_vec` in the source model.

        """
        return np.dot(target_doc_vec, self.translation_matrix)