luscenje_struktur/wani.py

from xml.etree import ElementTree
import re
from enum import Enum
from collections import defaultdict, namedtuple, Counter
import sys
import logging
import argparse
import pickle
import time
import subprocess
import concurrent.futures
import tempfile
from math import log2

from msd_translate import MSD_TRANSLATE

try:
    from tqdm import tqdm
except ImportError:
    tqdm = lambda x: x


MAX_NUM_COMPONENTS = -1

CODES = {
    "Noun": "N",
    "Verb": "V",
    "Adjective": "A",
    "Adverb": "R",
    "Pronoun": "P",
    "Numeral": "M",
    "Preposition": "S",
    "Conjunction": "C",
    "Particle": "Q",
    "Interjection": "I",
    "Abbreviation": "Y",
    "Residual": "X",

    'common': 'c',
    'proper': 'p',
    'masculine': 'm',
    'feminine': 'f',
    'neuter': 'n',
    "singular": "s",
    "dual": "d",
    "plural": "p",
    "nominative": "n",
    "genitive": "g",
    "dative": "d",
    "accusative": "a",
    "locative": "l",
    "instrumental": "i",
    "no": "n",
    "yes": "y",
    "main": "m",
    "auxiliary": "a",
    "perfective": "e",
    "progressive": "p",
    "biaspectual": "b",
    "infinitive": "n",
    "supine": "u",
    "participle": "p",
    "present": "r",
    "future": "f",
    "conditional": "c",
    "imperative": "m",
    "first": "1",
    "second": "2",
    "third": "3",
    "general": "g",
    "possessive": "s",
    "positive": "p",
    "comparative": "c",
    "superlative": "s",
    "personal": "p",
    "demonstrative": "d",
    "relative": "r",
    "reflexive": "x",
    "interrogative": "q",
    "indefinite": "i",
    "negative": "z",
    "bound": "b",
    "digit": "d",
    "roman": "r",
    "letter": "l",
    "cardinal": "c",
    "ordinal": "o",
    "pronominal": "p",
    "special": "s",
    "coordinating": "c",
    "subordinating": "s",
    "foreign": "f",
    "typo": "t",
    "program": "p",
}

TAGSET = {
    "N": ['type', 'gender', 'number', 'case', 'animate'],
    "V": ['type', 'aspect', 'vform', 'person', 'number', 'gender', 'negative'],
    "A": ['type', 'degree', 'gender', 'number', 'case', 'definiteness'],
    "R": ['type', 'degree'],
    "P": ['type', 'person', 'gender', 'number', 'case', 'owner_number', 'owned_gender', 'clitic'],
    "M": ['form', 'type', 'gender', 'number', 'case', 'definiteness'],
    "S": ['case'],
    "C": ['type'],
    "Q": [],
    "I": [],
    "Y": [],
    "X": ['type']
}

class ComponentType(Enum):
    Other = 0
    Core = 2
    Core2w = 3

class RestrictionType(Enum):
    Morphology = 0
    Lexis = 1
    MatchAll = 2

class Order(Enum):
    FromTo = 0
    ToFrom = 1
    Any = 2

    @staticmethod
    def new(order):
        if order is None:
            return Order.Any
        elif order == "to-from":
            return Order.ToFrom
        elif order == "from-to":
            return Order.FromTo
        else:
            raise NotImplementedError("What kind of ordering is: {}".format(order))


    def match(self, from_w, to_w):
        if self is Order.Any:
            return True

        fi = from_w.int_id
        ti = to_w.int_id

        if self is Order.FromTo:
            return fi < ti
        elif self is Order.ToFrom:
            return ti < fi
        else:
            raise NotImplementedError("Should not be here: Order match")


class ComponentRepresentation:
    def __init__(self, data, word_renderer):
        self.data = data
        self.word_renderer = word_renderer

        self.words = []
        self.rendition_text = None
        self.agreement = []

    def get_agreement(self):
        return []

    def add_word(self, word):
        self.words.append(word)

    def render(self):
        if self.rendition_text is None:
            self.rendition_text = self._render()

    def _render(self):
        raise NotImplementedError("Not implemented for class: {}".format(type(self)))

class LemmaCR(ComponentRepresentation):
    def _render(self):
        return self.words[0].lemma if len(self.words) > 0 else None

class LexisCR(ComponentRepresentation):
    def _render(self):
        return self.data['lexis']

class WordFormAllCR(ComponentRepresentation):
    def _render(self):
        if len(self.words) == 0:
            return None
        else:
            forms = [w.text.lower() for w in self.words]
            return "/".join(set(forms))

class WordFormAnyCR(ComponentRepresentation):
    def _render(self):
        text_forms = {}
        msd_lemma_txt_triplets = Counter([(w.msd, w.lemma, w.text) for w in self.words])
        for (msd, lemma, text), _n in reversed(msd_lemma_txt_triplets.most_common()):
            text_forms[(msd, lemma)] = text

        words_counter = []
        for word in self.words:
            words_counter.append((word.msd, word.lemma))
        sorted_words = sorted(set(words_counter), key=lambda x: -words_counter.count(x))

        for word_msd, word_lemma in sorted_words:
            for agr in self.agreement:
                if not agr.match(word_msd):
                    break
            else:
                for agr in self.agreement:
                    agr.confirm_match()

                if word_lemma is None:
                    return None
                else:
                    return text_forms[(word_msd, word_lemma)]

class WordFormMsdCR(WordFormAnyCR):
    def __init__(self, *args):
        super().__init__(*args)
        self.lemma = None
        self.msd = None

    def check_msd(self, word_msd):
        if 'msd' not in self.data:
            return True
        selectors = self.data['msd']

        for key, value in selectors.items():
            t = word_msd[0]
            v = TAGSET[t].index(key.lower())
            f1 = word_msd[v + 1]
            f2 = CODES[value]

            if '-' not in [f1, f2] and f1 != f2:
                return False

        return True

    def add_word(self, word):
        if self.lemma is None:
            self.lemma = word.lemma
            self.msd = word.msd

        if self.check_msd(word.msd):
            super().add_word(word)

    def _render(self):
        msd = self.word_renderer.get_lemma_msd(self.lemma, self.msd)
        self.words.append(WordMsdOnly(msd))

        return super()._render()

class WordFormAgreementCR(WordFormMsdCR):
    def __init__(self, data, word_renderer):
        super().__init__(data, word_renderer)
        self.rendition_candidate = None

    def get_agreement(self):
        return self.data['other']

    def match(self, word_msd):
        existing = [(w.msd, w.text) for w in self.words]

        lemma_available_words = self.word_renderer.available_words(self.lemma, existing)
        for candidate_msd, candidate_text in lemma_available_words:
            if self.msd[0] != candidate_msd[0]:
                continue

            if WordFormAgreementCR.check_agreement(word_msd, candidate_msd, self.data['agreement']):
                if self.check_msd(candidate_msd):
                    self.rendition_candidate = candidate_text
                    return True

        return False

    def confirm_match(self):
        self.rendition_text = self.rendition_candidate

    @staticmethod
    def check_agreement(msd1, msd2, agreements):
        for agr_case in agreements:
            t1 = msd1[0]
            # if not in msd, some strange msd was tries, skipping...
            if agr_case not in TAGSET[t1]:
                logging.warning("Cannot do agreement: {} for msd {} not found!"
                                .format(agr_case, msd1))
                return False

            v1 = TAGSET[t1].index(agr_case)
            # if none specified: nedolocnik, always agrees
            if v1 + 1 >= len(msd1):
                continue
            # first is uppercase, not in TAGSET
            m1 = msd1[v1 + 1]

            # REPEAT (not DRY!)
            t2 = msd2[0]
            if agr_case not in TAGSET[t2]:
                logging.warning("Cannot do agreement: {} for msd {} not found!"
                                .format(agr_case, msd2))
                return False
            v2 = TAGSET[t2].index(agr_case)
            if v2 + 1 >= len(msd2):
                continue
            m2 = msd2[v2 + 1]

            # match!
            if '-' not in [m1, m2] and m1 != m2:
                return False

        return True

    def render(self):
        pass


class ComponentRendition:
    def __init__(self):
        self.more = {}
        self.representation_factory = ComponentRepresentation

    def add_feature(self, feature):
        if 'rendition' in feature:
            if feature['rendition'] == "lemma":
                self.representation_factory = LemmaCR
            elif feature['rendition'] == "word_form":
                # just by default, changes with selection
                self.representation_factory = WordFormAnyCR
            elif feature['rendition'] == "lexis":
                self.representation_factory = LexisCR
                self.more['lexis'] = feature['string']
            else:
                raise NotImplementedError("Representation rendition: {}".format(feature))

        elif 'selection' in feature:
            if feature['selection'] == "msd":
                # could already be agreement
                if self.representation_factory != WordFormAgreementCR:
                    self.representation_factory = WordFormMsdCR
                self.more['msd'] = {k: v for k, v in feature.items() if k != 'selection'}
            elif feature['selection'] == "all":
                self.representation_factory = WordFormAllCR
            elif feature['selection'] == 'agreement':
                assert feature['head'][:4] == 'cid_'
                assert feature['msd'] is not None
                self.representation_factory = WordFormAgreementCR
                self.more['agreement'] = feature['msd'].split('+')
                self.more['other'] = feature['head'][4:]
            else:
                raise NotImplementedError("Representation selection: {}".format(feature))

    def cr_instance(self, word_renderer):
        return self.representation_factory(self.more, word_renderer)

    @staticmethod
    def set_representations(match, word_renderer):
        representations = {}
        for c in match.structure.components:
            representations[c.idx] = []
            for rep in c.representation:
                representations[c.idx].append(rep.cr_instance(word_renderer))

        for cid, reps in representations.items():
            for rep in reps:
                for agr in rep.get_agreement():
                    if len(representations[agr]) != 1:
                        n = len(representations[agr])
                        raise NotImplementedError(
                            "Structure {}: ".format(match.structure.id) +
                            "component {} has agreement".format(cid) +
                            " with component {}".format(agr) +
                            ", however there are {} (!= 1) representations".format(n) +
                            " of component {}!".format(agr))

                    representations[agr][0].agreement.append(rep)

        for words in match.matches:
            # first pass, check everything but agreements
            for w_id, w in words.items():
                component = match.structure.get_component(w_id)
                component_representations = representations[component.idx]
                for representation in component_representations:
                    representation.add_word(w)

        for cid, reps in representations.items():
            for rep in reps:
                rep.render()
        
        for cid, reps in representations.items():
            reps = [rep.rendition_text for rep in reps]
            if reps == []:
                pass
            elif all(r is None for r in reps):
                match.representations[cid] = None
            else:
                match.representations[cid] = " ".join(("" if r is None else r) for r in reps)


class ComponentStatus(Enum):
    Optional = 0
    Required = 1
    Forbidden = 2


def get_level(restriction):
    for feature in restriction:
        if "level" in feature.keys():
            lvl = feature.get("level")
        else:
            continue

    raise RuntimeError("Unreachable!")


def determine_ppb(rgx):
    if rgx[0] in ("A", "N", "R"):
        return 0
    elif rgx[0] == "V":
        if len(rgx) == 1:
            return 2
        elif 'a' in rgx[1]:
            return 3
        elif 'm' in rgx[1]:
            return 1
        else:
            return 2
    else:
        return 4

class MorphologyRegex:
    def __init__(self, restriction):
        self.min_msd_length = 1

        restr_dict = {}
        for feature in restriction:
            feature_dict = dict(feature.items())

            match_type = True
            if "filter" in feature_dict:
                assert feature_dict['filter'] == "negative"
                match_type = False
                del feature_dict['filter']

            assert len(feature_dict) == 1
            key, value = next(iter(feature_dict.items()))
            restr_dict[key] = (value, match_type)

        assert 'POS' in restr_dict
        category = restr_dict['POS'][0].capitalize()
        cat_code = CODES[category]
        rgx = [cat_code] + ['.'] * 10

        del restr_dict['POS']

        for attribute, (value, typ) in restr_dict.items():
            index = TAGSET[cat_code].index(attribute.lower())
            assert index >= 0

            if '|' in value:
                match = "".join(CODES[val] for val in value.split('|'))
            else:
                match = CODES[value]

            match = "[{}{}]".format("" if typ else "^", match)
            rgx[index + 1] = match

            if typ:
                self.min_msd_length = max(index + 1, self.min_msd_length)

        # strip rgx
        for i in reversed(range(len(rgx))):
            if rgx[i] == '.':
                rgx = rgx[:-1]
            else:
                break

        self.re_objects = [re.compile(r) for r in rgx]
        self.rgx = rgx
    
    def __call__(self, text):
        if len(text) <= self.min_msd_length:
            return False

        for c, r in zip(text, self.re_objects):
            if not r.match(c):
                return False
        return True


class LexisRegex:
    def __init__(self, restriction):
        restr_dict = {}
        for feature in restriction:
            restr_dict.update(feature.items())

        assert "lemma" in restr_dict
        self.match_list = restr_dict['lemma'].split('|')
    
    def __call__(self, text):
        return text in self.match_list


class Restriction:
    def __init__(self, restriction_tag):
        self.ppb = 4 # polnopomenska beseda (0-4)

        if restriction_tag is None:
            self.type = RestrictionType.MatchAll
            self.matcher = None
            self.present = None
            return

        restriction_type = restriction_tag.get('type')
        if restriction_type == "morphology":
            self.type = RestrictionType.Morphology
            self.matcher = MorphologyRegex(list(restriction_tag))
            self.ppb = determine_ppb(self.matcher.rgx)

        elif restriction_type == "lexis":
            self.type = RestrictionType.Lexis
            self.matcher = LexisRegex(list(restriction_tag))
        else:
            raise NotImplementedError()

    def match(self, word):
        if self.type == RestrictionType.Morphology:
            match_to = word.msd
        elif self.type == RestrictionType.Lexis:
            match_to = word.lemma
        elif self.type == RestrictionType.MatchAll:
            return True
        else:
            raise RuntimeError("Unreachable!")

        return self.matcher(match_to)


class Component:
    def __init__(self, info):
        idx = info['cid']
        name = info['name'] if 'name' in info else None
        typ = ComponentType.Core if info['type'] == "core" else ComponentType.Other

        if 'status' not in info:
            status = ComponentStatus.Required
        elif info['status'] == 'forbidden':
            status = ComponentStatus.Forbidden
        elif info['status'] == 'obligatory':
            status = ComponentStatus.Required
        elif info['status'] == 'optional':
            status = ComponentStatus.Optional
        else:
            raise NotImplementedError("strange status: {}".format(info['status']))

        self.status = status
        self.name = name
        self.idx = idx
        self.restrictions = []
        self.next_element = []
        self.representation = []
        self.selection = {}
        self.type = typ

        self.iter_ctr = 0

    def add_next(self, next_component, link_label, order):
        self.next_element.append((next_component, link_label, Order.new(order)))

    def set_restriction(self, restrictions_tag):
        if restrictions_tag is None:
            self.restrictions = [Restriction(None)]

        elif restrictions_tag.tag == "restriction":
            self.restrictions = [Restriction(restrictions_tag)]

        elif restrictions_tag.tag == "restriction_or":
            self.restrictions = [Restriction(el) for el in restrictions_tag]

        else:
            raise RuntimeError("Unreachable")

    def set_representation(self, representation):
        for rep in representation:
            crend = ComponentRendition()
            for feature in rep:
                crend.add_feature(feature.attrib)
            self.representation.append(crend)

    def find_next(self, deps, comps, restrs, reprs):
        to_ret = []
        for d in deps:
            if d[0] == self.idx:
                _, idx, dep_label, order = d

                next_component = Component(comps[idx])
                next_component.set_restriction(restrs[idx])
                next_component.set_representation(reprs[idx])
                to_ret.append(next_component)

                self.add_next(next_component, dep_label, order)
                others = next_component.find_next(deps, comps, restrs, reprs)
                to_ret.extend(others)

        return to_ret

    def name_str(self):
        return "_" if self.name is None else self.name

    def match(self, word):
        m1 = self._match_self(word)
        if m1 is None:
            return None

        mn = self._match_next(word)
        if mn is None:
            return None

        to_ret = [m1]
        for cmatch in mn:
            # if good match but nothing to add, just continue
            if len(cmatch) == 0:
                continue

            # if more than one match found for particular component
            elif len(cmatch) > 1:
                # if more than one match in multiple components, NOPE!
                if len(to_ret) > 1:
                    logging.warning("Strange multiple match: {}".format(
                        str([w.id for w in cmatch[0].values()])))

                    for tr in to_ret:
                        tr.update(cmatch[0])
                    continue

                # yeah, so we have found more than one match, =>
                # more than one element in to_ret
                to_ret = [{**dict(to_ret[0]), **m} for m in cmatch]

            else:
                for tr in to_ret:
                    tr.update(cmatch[0])

        return to_ret

    def _match_self(self, word):
        # matching
        for restr in self.restrictions:
            if restr.match(word): # match either
                return {self.idx: word}

    def _match_next(self, word):
        # matches for every component in links from this component
        to_ret = []

        # need to get all links that match
        for next, link, order in self.next_element:
            next_links = word.get_links(link)
            to_ret.append([])

            # good flag
            good = next.status != ComponentStatus.Required
            for next_word in next_links:
                if not order.match(word, next_word):
                    continue

                match = next.match(next_word)

                if match is not None:
                    # special treatement for forbidden
                    if next.status == ComponentStatus.Forbidden:
                        good = False
                        break

                    else:
                        assert type(match) is list
                        to_ret[-1].extend(match)
                        good = True

            # if none matched, nothing found!
            if not good:
                return None

        return to_ret


class SyntacticStructure:
    def __init__(self):
        self.id = None
        self.lbs = None
        self.components = []

    @staticmethod
    def from_xml(xml):
        st = SyntacticStructure()
        st.id = xml.get('id')
        st.lbs = xml.get('LBS')

        assert len(list(xml)) == 1
        system = next(iter(xml))

        assert system.get('type') == 'JOS'
        components, dependencies, definitions = list(system)

        deps = [(dep.get('from'), dep.get('to'), dep.get('label'), dep.get('order'))
                for dep in dependencies]
        comps = {comp.get('cid'): dict(comp.items()) for comp in components}

        restrs, forms = {}, {}

        for comp in definitions:
            n = comp.get('cid')
            restrs[n] = None
            forms[n] = []

            for el in comp:
                if el.tag.startswith("restriction"):
                    assert restrs[n] is None
                    restrs[n] = el
                elif el.tag.startswith("representation"):
                    st.add_representation(n, el, forms)
                else:
                    raise NotImplementedError("Unknown definition: {} in structure {}"
                                              .format(el.tag, st.id))

        fake_root_component = Component({'cid': '#', 'type': 'other'})
        st.components = fake_root_component.find_next(deps, comps, restrs, forms)

        st.determine_core2w()
        return st

    def determine_core2w(self):
        ppb_components = []
        for c in self.components:
            if c.type != ComponentType.Core:
                continue

            ppb = 4
            for r in c.restrictions:
                ppb = min(r.ppb, ppb)

            ppb_components.append((c, ppb))

        ppb_components = sorted(ppb_components, key=lambda c: c[1])
        if len(ppb_components) > 2 and ppb_components[1][1] == ppb_components[2][1]:
            raise RuntimeError("Cannot determine 2 'jedrna polnopomenska beseda' for", self.id)

        for c, _ in ppb_components[:2]:
            c.type = ComponentType.Core2w

    def add_representation(self, n, rep_el, forms):
        assert rep_el.tag == "representation"
        to_add = []
        for el in rep_el:
            assert el.tag == "feature"
            if 'rendition' in el.attrib or 'selection' in el.attrib:
                to_add.append(el)
            else:
                logging.warning("Strange representation feature in structure {}. Skipping"
                                .format(self.id))
                continue
        forms[n].append(to_add)

    def get_component(self, idx):
        for c in self.components:
            if c.idx == idx:
                return c
        raise RuntimeError("Unknown component id: {}".format(idx))

    def match(self, word):
        matches = self.components[0].match(word)
        return [] if matches is None else matches

def load_structures(filename):
    with open(filename, 'r') as fp:
        et = ElementTree.XML(fp.read())

    return build_structures(et), get_lemma_features(et)

def build_structures(et):
    global MAX_NUM_COMPONENTS

    structures = []
    for structure in et.iter('syntactic_structure'):
        to_append = SyntacticStructure.from_xml(structure)
        if to_append is None:
            continue

        structures.append(to_append)
        MAX_NUM_COMPONENTS = max(MAX_NUM_COMPONENTS, len(to_append.components))

    return structures

def get_lemma_features(et):
    lf = et.find('lemma_features')
    if lf is None:
        return {}

    result = {}
    for pos in lf.iter('POS'):
        rgx_list = MorphologyRegex(pos).rgx
        rgx_str = ""
        for position in rgx_list:
            if position == ".":
                rgx_str += " "
            elif len(position) == 1:
                rgx_str += position
            elif len(position) == 3 and position[0] == "[" and position[2] == "]":
                rgx_str += position[1]
            else:
                raise RuntimeError("Strange rgx for lemma_feature...")

        assert rgx_str[0].isupper()
        result[rgx_str[0]] = rgx_str.strip().replace(' ', '-')

    return result

def get_msd(comp):
    d = dict(comp.items())
    if 'msd' in d:
        return d['msd']
    elif 'ana' in d:
        return d['ana'][4:]
    else:
        logging.error(d, file=sys.stderr)
        raise NotImplementedError("MSD?")


class WordMsdOnly:
    def __init__(self, msd):
        self.msd = msd
        self.lemma = None
        self.text = None

    def most_frequent_text(self, _):
        return None


class Word:
    def __init__(self, xml, do_msd_translate):
        self.lemma = xml.get('lemma')
        self.msd = MSD_TRANSLATE[get_msd(xml)] if do_msd_translate else get_msd(xml)
        self.id = xml.get('id')
        self.text = xml.text
        self.links = defaultdict(list)

        last_num = self.id.split('.')[-1]
        if last_num[0] not in '0123456789':
            last_num = last_num[1:]
        self.int_id = int(last_num)

        assert None not in (self.id, self.lemma, self.msd)

    @staticmethod
    def pc_word(pc, do_msd_translate):
        pc.set('lemma', pc.text)
        pc.set('msd', "N" if do_msd_translate else "U")
        return Word(pc, do_msd_translate)

    def add_link(self, link, to):
        self.links[link].append(to)

    def get_links(self, link):
        if link not in self.links and "|" in link:
            for l in link.split('|'):
                self.links[link].extend(self.links[l])

        return self.links[link]

    def most_frequent_text(self, word_renderer):
        return word_renderer.render(self.lemma, self.msd)

class WordMsdRenderer:
    def __init__(self, lemma_features):
        self.all_words = []
        self.rendered_words = {}
        self.frequent_words = {}
        self.num_words = {}
        self.lemma_msd = {}
        self.lemma_features = lemma_features
        self.memoized_msd_merges = {}

    def add_words(self, words):
        self.all_words.extend(words)

    def num_all_words(self):
        return len(self.all_words)

    def generate_renders(self):
        num_words = defaultdict(int)
        data = defaultdict(lambda: defaultdict(list))
        for w in self.all_words:
            data[w.lemma][w.msd].append(w.text)

        for lemma, ld in data.items():
            self.rendered_words[lemma] = {}
            freq_words = defaultdict(int)
            common_msd = "*" * 10

            for msd, texts in ld.items():
                # TODO: this should be out of generate_renders...
                num_words[(lemma, msd[0])] += len(texts)

                rep = max(set(texts), key=texts.count)
                self.rendered_words[lemma][msd] = (rep, len(texts))

                for txt in texts:
                    freq_words[(msd, txt)] += 1

                common_msd = self.merge_msd(common_msd, msd)

            self.lemma_msd[lemma] = common_msd

            self.frequent_words[lemma] = []
            for (msd, txt), n in sorted(freq_words.items(), key=lambda x: -x[1]):
                self.frequent_words[lemma].append((msd, txt, n))

        lf = self.lemma_features
        for lemma in self.lemma_msd:
            cmsd = self.lemma_msd[lemma]
            if cmsd[0] in lf:
                self.lemma_msd[lemma] = "".join(
                    l1 if l1 != "-" else l2 for l1, l2 in zip(lf[cmsd[0]], cmsd)
                )
        
        self.num_words = dict(num_words)

    def merge_msd(self, common_msd, new_msd):
        key = (common_msd, new_msd)
        if key in self.memoized_msd_merges:
            return self.memoized_msd_merges[key]

        def merge_letter(l1, l2):
            if l1 == "*":
                return l2
            elif l1 != l2:
                return "-"
            else:
                return l1

        value = "".join(merge_letter(l1, l2) for l1, l2 in zip(common_msd, new_msd))
        self.memoized_msd_merges[key] = value
        return value

    def render(self, lemma, msd):
        if lemma in self.rendered_words:
            if msd in self.rendered_words[lemma]:
                return self.rendered_words[lemma][msd][0]

    def available_words(self, lemma, existing_texts):
        counted_texts = Counter(existing_texts)
        for (msd, text), _n in counted_texts.most_common():
            yield (msd, text)

        if lemma in self.frequent_words:
            for msd, text, _ in self.frequent_words[lemma]:
                if (msd, text) not in counted_texts:
                    yield (msd, text)

    def get_lemma_msd(self, lemma, word_msd):
        # should be here, since we collect every lemmas
        lemma_msd = self.lemma_msd[lemma]

        if lemma_msd[0] == '-':
            if word_msd[0] in self.lemma_features:
                return self.lemma_features[word_msd[0]]
            else:
                return '-'
        else:
            return lemma_msd

def is_root_id(id_):
    return len(id_.split('.')) == 3


def load_files(args):
    filenames = args.input
    skip_id_check = args.skip_id_check
    do_msd_translate = not args.no_msd_translate

    for n, fname in enumerate(filenames):
        if args.count_files:
            status = " :: {} / {}".format(n, len(filenames))
        else:
            status = ""
        yield load_tei_file(fname, skip_id_check, do_msd_translate, args.pc_tag, status)


def load_tei_file(filename, skip_id_check, do_msd_translate, pc_tag, status):
    logging.info("LOADING FILE: {}{}".format(filename, status))

    with open(filename, 'r') as fp:
        xmlstring = re.sub(' xmlns="[^"]+"', '', fp.read(), count=1)
        xmlstring = xmlstring.replace(' xml:', ' ')
        et = ElementTree.XML(xmlstring)

    words = {}
    for w in et.iter("w"):
        words[w.get('id')] = Word(w, do_msd_translate)
    for pc in et.iter(pc_tag):
        words[pc.get('id')] = Word.pc_word(pc, do_msd_translate)

    for l in et.iter("link"):
        if 'dep' in l.keys():
            ana = l.get('afun')
            lfrom = l.get('from')
            dest = l.get('dep')
        else:
            ana = l.get('ana')
            if ana[:4] != 'syn:': # dont bother...
                continue
            ana = ana[4:]
            lfrom, dest = l.get('target').replace('#', '').split()

        if lfrom in words:
            if not skip_id_check and is_root_id(lfrom):
                logging.error("NOO: {}".format(lfrom))
                sys.exit(1)

            if dest in words:
                next_word = words[dest]
                words[lfrom].add_link(ana, next_word)
            else:
                logging.error("Unknown id: {}".format(dest))
                sys.exit(1)

        else:
            # strange errors, just skip...
            pass

    return list(words.values())


class Formatter:
    def __init__(self, colocation_ids, word_renderer):
        self.colocation_ids = colocation_ids
        self.word_renderer = word_renderer
        self.additional_init()
    
    def header_repeat(self):
        raise NotImplementedError("Header repeat formatter not implemented")
    def header_right(self):
        raise NotImplementedError("Header right formatter not implemented")
    def content_repeat(self, words, representations, idx, sidx):
        raise NotImplementedError("Content repeat formatter not implemented")
    def content_right(self, freq):
        raise NotImplementedError("Content right formatter not implemented")
    def group(self):
        raise NotImplementedError("Group for formatter not implemented")
    
    def additional_init(self):
        pass
    def length(self):
        return len(self.header_repeat())

    def set_structure(self, structure):
        pass
    def new_match(self, match):
        pass


class OutNoStatFormatter(Formatter):
    def additional_init(self):
        self.representation = ""

    def header_repeat(self):
        return ["Lemma", "Representative_form", "RF_scenario"]
    
    def header_right(self):
        return ["Joint_representative_form", "Frequency"]
    
    def content_repeat(self, words, representations, idx, _sidx):
        word = words[idx]
        if idx not in representations:
            return [word.lemma, "", ""]

        rep = representations[idx]
        if rep is None:
            self.representation += " " + word.lemma
            return [word.lemma, word.lemma, "lemma_fallback"]
        else:
            self.representation += " " + rep
            return [word.lemma, rep, "ok"]

    def content_right(self, freq):
        rep = re.sub(' +', ' ', self.representation.strip())
        result = [rep, str(freq)]
        self.representation = ""
        return result

    def group(self):
        return True

class AllFormatter(Formatter):
    def header_repeat(self):
        return ["Token_ID", "Word_form", "Lemma", "Msd"]
    
    def header_right(self):
        return []
    
    def content_repeat(self, words, _representations, idx, _sidx):
        word = words[idx]
        return [word.id, word.text, word.lemma, word.msd]
    
    def content_right(self, _freq):
        return []
    
    def group(self):
        return False

class StatsFormatter(Formatter):
    def additional_init(self):
        self.stats = None
        self.jppb = None
        self.corew = None
    
    @staticmethod
    def stat_str(num):
        return "{:.5f}".format(num) if isinstance(num, float) else str(num)
    
    def set_structure(self, structure):
        jppb = []
        corew = []

        for component in structure.components:
            if component.type == ComponentType.Core2w:
                jppb.append(component.idx)
            if component.type != ComponentType.Other:
                corew.append(component.idx)

        assert(len(jppb) == 2)
        self.jppb = tuple(jppb)
        self.corew = tuple(corew)
    
    def new_match(self, match):
        self.stats = {"freq": {}}

        for cid in self.corew:
            if cid not in match.matches[0]:
                freq = 0
            else:
                word = match.matches[0][cid]
                freq = self.word_renderer.num_words[(word.lemma, word.msd[0])]

            self.stats["freq"][cid] = freq

        fx = self.stats["freq"][self.jppb[0]]
        fy = self.stats["freq"][self.jppb[1]]
        freq = len(match)
        N = self.word_renderer.num_all_words()

        self.stats['d12'] = freq / fx - (fy - freq) / (N - fx)
        self.stats['d21'] = freq / fy - (fx - freq) / (N - fy)

        self.stats['df'] = match.distinct_forms()
        self.stats['freq_all'] = freq

    def header_repeat(self):
        return ["Distribution"]
    
    def header_right(self):
        return ["Delta_p12", "Delta_p21", "LogDice_core", "LogDice_all", "Distinct_forms"]
    
    def content_repeat(self, words, representations, idx, sidx):
        # not a core word
        if idx not in self.corew:
            return [""] * self.length()

        word = words[idx]
        key = (sidx, idx, word.lemma)
        distribution = self.colocation_ids.dispersions[key]
        return [self.stat_str(distribution)]
    
    def content_right(self, freq):
        fx = self.stats["freq"][self.jppb[0]]
        fy = self.stats["freq"][self.jppb[1]]
        freq = self.stats['freq_all']
        logdice_core = 14 + log2(2 * freq / (fx + fy))

        fi = [self.stats["freq"][idx] for idx in self.corew]
        fi = [f for f in fi if f > 0]
        logdice_all = 14 + log2(len(fi) * freq / sum(fi))

        return [self.stat_str(x) for x in (
            self.stats["d12"], self.stats["d21"], logdice_core, logdice_all, self.stats['df']
        )]
    
    def group(self):
        return True

class OutFormatter(Formatter):
    def additional_init(self):
        self.f1 = OutNoStatFormatter(self.colocation_ids, self.word_renderer)
        self.f2 = StatsFormatter(self.colocation_ids, self.word_renderer)

    def header_repeat(self):
        return self.f1.header_repeat() + self.f2.header_repeat()

    def header_right(self):
        return self.f1.header_right() + self.f2.header_right()

    def content_repeat(self, words, representations, idx, sidx):
        cr1 = self.f1.content_repeat(words, representations, idx, sidx)
        cr2 = self.f2.content_repeat(words, representations, idx, sidx)
        return cr1 + cr2

    def content_right(self, freq):
        return self.f1.content_right(freq) + self.f2.content_right(freq)

    def group(self):
        return self.f1.group() and self.f2.group()

    def set_structure(self, structure):
        self.f2.set_structure(structure)

    def new_match(self, match):
        self.f2.new_match(match)


class Writer:
    @staticmethod
    def other_params(args):
        return (args.multiple_output, int(args.sort_by), args.sort_reversed)

    @staticmethod
    def make_output_writer(args, colocation_ids, word_renderer):
        params = Writer.other_params(args)
        return Writer(args.out, OutFormatter(colocation_ids, word_renderer), params)

    @staticmethod
    def make_output_no_stat_writer(args, colocation_ids, word_renderer):
        params = Writer.other_params(args)
        return Writer(args.out_no_stat, OutNoStatFormatter(colocation_ids, word_renderer), params)

    @staticmethod
    def make_all_writer(args, colocation_ids, word_renderer):
        return Writer(args.all, AllFormatter(colocation_ids, word_renderer), None)

    @staticmethod
    def make_stats_writer(args, colocation_ids, word_renderer):
        params = Writer.other_params(args)
        return Writer(args.stats, StatsFormatter(colocation_ids, word_renderer), params)

    def __init__(self, file_out, formatter, params):
        if params is None:
            self.multiple_output = False
            self.sort_by = -1
            self.sort_order = None
        else:
            self.multiple_output = params[0]
            self.sort_by = params[1]
            self.sort_order = params[2]

        self.output_file = file_out
        self.formatter = formatter

    def header(self):
        repeating_cols = self.formatter.header_repeat()
        cols = ["C{}_{}".format(i + 1, thd) for i in range(MAX_NUM_COMPONENTS) 
                for thd in repeating_cols]

        cols = ["Structure_ID"] + cols + ["Colocation_ID"]
        cols += self.formatter.header_right()
        return cols

    def sorted_rows(self, rows):
        if self.sort_by < 0 or len(rows) < 2:
            return rows

        if len(rows[0]) <= self.sort_by:
            logging.warning("Cannot sort by column #{}: Not enough columns!".format(len(rows[0])))
            return rows

        try:
            int(rows[0][self.sort_by])
            def key(row): 
                return int(row[self.sort_by])
        except ValueError:
            def key(row): 
                return row[self.sort_by].lower()

        return sorted(rows, key=key, reverse=self.sort_order)

    def write_header(self, file_handler):
        file_handler.write(", ".join(self.header()) + "\n")

    def write_out_worker(self, file_handler, structure, colocation_ids):
        rows = []
        components = structure.components

        for match in colocation_ids.get_matches_for(structure):
            self.formatter.new_match(match)

            for words in match.matches:
                to_write = []

                for idx, _comp in enumerate(components):
                    idx = str(idx + 1)
                    if idx not in words:
                        to_write.extend([""] * self.formatter.length())
                    else:
                        to_write.extend(self.formatter.content_repeat(words, match.representations, idx, structure.id))

                # make them equal size
                to_write.extend([""] * (MAX_NUM_COMPONENTS * self.formatter.length() - len(to_write)))

                # structure_id and colocation_id
                to_write = [structure.id] + to_write + [match.match_id]

                # header_right
                to_write.extend(self.formatter.content_right(len(match)))
                rows.append(to_write)

                if self.formatter.group():
                    break

        if rows != []:
            rows = self.sorted_rows(rows)
            file_handler.write("\n".join([", ".join(row) for row in rows]) + "\n")
            file_handler.flush()

    def write_out(self, structures, colocation_ids):
        if self.output_file is None:
            return

        def fp_close(fp_):
            if fp_ != sys.stdout:
                fp_.close()

        def fp_open(snum=None):
            if snum is None:
                return open(self.output_file, "w")
            else:
                return open("{}.{}".format(self.output_file, snum), "w")

        if not self.multiple_output:
            fp = fp_open()
            self.write_header(fp)

        for s in structures:
            if self.multiple_output:
                fp = fp_open(s.id)
                self.write_header(fp)

            self.formatter.set_structure(s)
            self.write_out_worker(fp, s, colocation_ids)

            if self.multiple_output:
                fp_close(fp)

        if not self.multiple_output:
            fp_close(fp)

class StructureMatch:
    def __init__(self, match_id, structure):
        self.match_id = match_id
        self.structure = structure

        self.matches = []
        self.representations = {}

    def distinct_forms(self):
        dm = set()
        keys = list(self.matches[0].keys())
        for words in self.matches:
            dm.add(" ".join(words[k].text for k in keys))
        return len(dm)

    def append(self, match):
        self.matches.append(match)

    def __len__(self):
        return len(self.matches)

class ColocationIds:
    def __init__(self):
        self.data = {}
        self.min_frequency = args.min_freq
        self.dispersions = {}

    def _add_match(self, key, structure, match):
        if key not in self.data:
            self.data[key] = StructureMatch(str(len(self.data) + 1), structure)
        self.data[key].append(match)

    def get(self, key, n):
        return self.data[key][n]

    def add_matches(self, matches):
        for structure, nms in matches.items():
            for nm in nms:
                self._add_match(nm[1], structure, nm[0])

    def get_matches_for(self, structure):
        for _cid_tup, sm in self.data.items():
            if sm.structure != structure:
                continue

            yield sm

    def set_representations(self, word_renderer):
        for _1, sm in tqdm(self.data.items()):
            ComponentRendition.set_representations(sm, word_renderer)

    def determine_colocation_dispersions(self):
        dispersions = defaultdict(int)
        for (structure_id, *word_tups) in self.data.keys():
            for component_id, lemma in word_tups:
                dispersions[(structure_id, component_id, lemma)] += 1
        self.dispersions = dict(dispersions)


def match_file(words, structures):
    matches = {s: [] for s in structures}

    for s in tqdm(structures):
        for w in words:
            mhere = s.match(w)
            for match in mhere:
                colocation_id = [(idx, w.lemma) for idx, w in match.items()]
                colocation_id = [s.id] + list(sorted(colocation_id, key=lambda x: x[0]))
                colocation_id = tuple(colocation_id)

                matches[s].append((match, colocation_id))

    return matches


def main(structures_file, args):
    structures, lemma_msds = load_structures(structures_file)

    colocation_ids = ColocationIds()
    word_renderer = WordMsdRenderer(lemma_msds)

    if args.parallel:
        num_parallel = int(args.parallel)

        # make temporary directory to hold temporary files
        with tempfile.TemporaryDirectory() as tmpdirname:
            cmd = sys.argv
            for inpt in args.input:
                if inpt in cmd:
                    cmd.remove(inpt)

            # remove "--parallel X"
            pidx = cmd.index('--parallel')
            del cmd[pidx]
            del cmd[pidx]

            def func(n):
                cmdn = [sys.executable] + cmd + [args.input[n],
                                                 "--match-to-file", "{}/{}.p".format(tmpdirname, n)]
                subprocess.check_call(cmdn)
                return n

            # use ThreadPoolExecuter to run subprocesses in parallel using py threads
            with concurrent.futures.ThreadPoolExecutor(max_workers=num_parallel) as executor:
                # fancy interface to wait for threads to finish
                for id_input in executor.map(func, [i for i, _ in enumerate(args.input)]):
                    with open("{}/{}.p".format(tmpdirname, id_input), "rb") as fp:
                        words, matches = pickle.load(fp)

                    colocation_ids.add_matches(matches)
                    word_renderer.add_words(words)

    else:
        for words in load_files(args):
            matches = match_file(words, structures)
            # just save to temporary file, used for children of a parallel process
            # MUST NOT have more than one file
            if args.match_to_file is not None:
                with open(args.match_to_file, "wb") as fp:
                    pickle.dump((words, matches), fp)
                    return
            else:
                colocation_ids.add_matches(matches)
                word_renderer.add_words(words)

    # get word renders for lemma/msd
    word_renderer.generate_renders()
    colocation_ids.determine_colocation_dispersions()

    # figure out representations!
    if args.out or args.out_no_stat:
        colocation_ids.set_representations(word_renderer)

    Writer.make_output_writer(args, colocation_ids, word_renderer).write_out(
        structures, colocation_ids)
    Writer.make_output_no_stat_writer(args, colocation_ids, word_renderer).write_out(
        structures, colocation_ids)
    Writer.make_all_writer(args, colocation_ids, word_renderer).write_out(
        structures, colocation_ids)
    Writer.make_stats_writer(args, colocation_ids, word_renderer).write_out(
        structures, colocation_ids)

if __name__ == '__main__':
    parser = argparse.ArgumentParser(
        description='Extract structures from a parsed corpus.')
    parser.add_argument('structures',
                        help='Structures definitions in xml file')
    parser.add_argument('input',
                        help='input xml file in `ssj500k form`, can list more than one', nargs='+')
    parser.add_argument('--out',
                        help='Classic output file')
    parser.add_argument('--out-no-stat',
                        help='Output file, but without statistical columns')
    parser.add_argument('--all',
                        help='Additional output file, writes more data')
    parser.add_argument('--stats',
                        help='Output file for statistics')

    parser.add_argument('--no-msd-translate',
                        help='MSDs are translated from slovene to english by default',
                        action='store_true')
    parser.add_argument('--skip-id-check',
                        help='Skips checks for ids of <w> and <pc>, if they are in correct format',
                        action='store_true')
    parser.add_argument('--min_freq', help='Minimal frequency in output',
                        type=int, default=0, const=1, nargs='?')
    parser.add_argument('--verbose', help='Enable verbose output to stderr',
                        choices=["warning", "info", "debug"], default="info",
                        const="info", nargs='?')
    parser.add_argument('--count-files',
                        help="Count files: more verbose output", action='store_true')
    parser.add_argument('--multiple-output',
                        help='Generate one output for each syntactic structure',
                        action='store_true')

    parser.add_argument('--sort-by',
                        help="Sort by a this column (index)", type=int, default=-1)
    parser.add_argument('--sort-reversed',
                        help="Sort in reversed ored", action='store_true')

    parser.add_argument('--pc-tag',
                        help='Tag for separators, usually pc or c', default="pc")
    parser.add_argument('--parallel',
                        help='Run in multiple processes, should speed things up')
    parser.add_argument('--match-to-file', help='Do not use!')

    args = parser.parse_args()
    logging.basicConfig(stream=sys.stderr, level=args.verbose.upper())

    start = time.time()
    main(args.structures, args)
    logging.info("TIME: {}".format(time.time() - start))