文中の主題とそれぞれの従属句を抽出する方法は？

Question

主題に応じた感情を得ることができるように、文の主題抽出に取り組んでいます。この目的のために、python2.7でnltkを使用しています。例として次の文を取り上げます。

Donald Trump is the worst president of USA, but Hillary is better than him

彼は、Donald TrumpとHillaryが2つの主題であり、Donald Trumpに関連する感情は否定的ですが、Hillaryに関連する感情は肯定的であることがわかります。今まで、私はこの文を名詞句のチャンクに分割することができ、次のものを得ることができます：

(S
  (NP Donald/NNP Trump/NNP)
  is/VBZ
  (NP the/DT worst/JJS president/NN)
  in/IN
  (NP USA,/NNP)
  but/CC
  (NP Hillary/NNP)
  is/VBZ
  better/JJR
  than/IN
  (NP him/PRP))

では、これらの名詞句から主題を見つけるにはどうすればよいでしょうか。次に、両方の主題を対象としたフレーズをグループ化するにはどうすればよいですか？両方の主題を別々に意味するフレーズを取得したら、両方の感情分析を別々に実行できます。

[〜＃〜]編集[〜＃〜]

@Krzysiek（spacy）によって言及されたライブラリを調べたところ、文の中に依存関係ツリーも表示されました。

コードは次のとおりです。

from spacy.en import English
parser = English()

example = u"Donald Trump is the worst president of USA, but Hillary is better than him"
parsedEx = parser(example)
# shown as: original token, dependency tag, head Word, left dependents, right dependents
for token in parsedEx:
    print(token.orth_, token.dep_, token.head.orth_, [t.orth_ for t in token.lefts], [t.orth_ for t in token.rights])

依存関係ツリーは次のとおりです。

(u'Donald', u'compound', u'Trump', [], [])
(u'Trump', u'nsubj', u'is', [u'Donald'], [])
(u'is', u'ROOT', u'is', [u'Trump'], [u'president', u',', u'but', u'is'])
(u'the', u'det', u'president', [], [])
(u'worst', u'amod', u'president', [], [])
(u'president', u'attr', u'is', [u'the', u'worst'], [u'of'])
(u'of', u'prep', u'president', [], [u'USA'])
(u'USA', u'pobj', u'of', [], [])
(u',', u'punct', u'is', [], [])
(u'but', u'cc', u'is', [], [])
(u'Hillary', u'nsubj', u'is', [], [])
(u'is', u'conj', u'is', [u'Hillary'], [u'better'])
(u'better', u'acomp', u'is', [], [u'than'])
(u'than', u'prep', u'better', [], [u'him'])
(u'him', u'pobj', u'than', [], [])

これにより、文のさまざまなトークンの依存関係について詳細な洞察が得られます。これは、異なるペア間の依存関係を説明する論文への リンク です。このツリーを使用して、さまざまな主題の文脈上の単語をそれらに添付するにはどうすればよいですか？

Answer 1

私は最近、非常によく似た問題を解決していました-サブジェクト、アクション、オブジェクトを抽出する必要がありました。そして、私は自分の作品をオープンソース化したので、このライブラリをチェックできます： https://github.com/krzysiekfonal/textpipeliner

これはspacy（nltkの反対）に基づいていますが、センテンスツリーにも基づいています。

たとえば、このドキュメントを例としてspacyに埋め込んでみましょう。

import spacy
nlp = spacy.load("en")
doc = nlp(u"The Empire of Japan aimed to dominate Asia and the " \
               "Pacific and was already at war with the Republic of China " \
               "in 1937, but the world war is generally said to have begun on " \
               "1 September 1939 with the invasion of Poland by Germany and " \
               "subsequent declarations of war on Germany by France and the United Kingdom. " \
               "From late 1939 to early 1941, in a series of campaigns and treaties, Germany conquered " \
               "or controlled much of continental Europe, and formed the Axis alliance with Italy and Japan. " \
               "Under the Molotov-Ribbentrop Pact of August 1939, Germany and the Soviet Union partitioned and " \
               "annexed territories of their European neighbours, Poland, Finland, Romania and the Baltic states. " \
               "The war continued primarily between the European Axis powers and the coalition of the United Kingdom " \
               "and the British Commonwealth, with campaigns including the North Africa and East Africa campaigns, " \
               "the aerial Battle of Britain, the Blitz bombing campaign, the Balkan Campaign as well as the " \
               "long-running Battle of the Atlantic. In June 1941, the European Axis powers launched an invasion " \
               "of the Soviet Union, opening the largest land theatre of war in history, which trapped the major part " \
               "of the Axis' military forces into a war of attrition. In December 1941, Japan attacked " \
               "the United States and European territories in the Pacific Ocean, and quickly conquered much of " \
               "the Western Pacific.")

これで、単純なパイプ構造を作成できます（このプロジェクトのreadmeにあるパイプの詳細）：

pipes_structure = [SequencePipe([FindTokensPipe("VERB/nsubj/*"),
                                 NamedEntityFilterPipe(),
                                 NamedEntityExtractorPipe()]),
                   FindTokensPipe("VERB"),
                   AnyPipe([SequencePipe([FindTokensPipe("VBD/dobj/NNP"),
                                          AggregatePipe([NamedEntityFilterPipe("GPE"), 
                                                NamedEntityFilterPipe("PERSON")]),
                                          NamedEntityExtractorPipe()]),
                            SequencePipe([FindTokensPipe("VBD/**/*/pobj/NNP"),
                                          AggregatePipe([NamedEntityFilterPipe("LOC"), 
                                                NamedEntityFilterPipe("PERSON")]),
                                          NamedEntityExtractorPipe()])])]

engine = PipelineEngine(pipes_structure, Context(doc), [0,1,2])
engine.process()

そして結果としてあなたは得るでしょう：

>>>[([Germany], [conquered], [Europe]),
 ([Japan], [attacked], [the, United, States])]

実際には、別のライブラリであるgrammaregexに強く基づいています（パイプの検索）。投稿からそれについて読むことができます： https://medium.com/@krzysiek89dev/grammaregex-library-regex-like-for-text-mining-49e5706c9c6d#.zgx7odhsc

[〜＃〜]編集済み[〜＃〜]

実際、readmeで示した例では、adjは破棄されますが、必要なのは、必要に応じてエンジンに渡されるパイプ構造を調整することだけです。たとえば、サンプル文の場合、文ごとに3つの要素（主語、動詞、形容詞）のタプルを提供するような構造/ソリューションを提案できます。

import spacy
from textpipeliner import PipelineEngine, Context
from textpipeliner.pipes import *

pipes_structure = [SequencePipe([FindTokensPipe("VERB/nsubj/NNP"),
                                 NamedEntityFilterPipe(),
                                 NamedEntityExtractorPipe()]),
                       AggregatePipe([FindTokensPipe("VERB"),
                                      FindTokensPipe("VERB/xcomp/VERB/aux/*"),
                                      FindTokensPipe("VERB/xcomp/VERB")]),
                       AnyPipe([FindTokensPipe("VERB/[acomp,amod]/ADJ"),
                                AggregatePipe([FindTokensPipe("VERB/[dobj,attr]/NOUN/det/DET"),
                                               FindTokensPipe("VERB/[dobj,attr]/NOUN/[acomp,amod]/ADJ")])])
                      ]

engine = PipelineEngine(pipes_structure, Context(doc), [0,1,2])
engine.process()

それはあなたに結果を与えるでしょう：

[([Donald, Trump], [is], [the, worst])]

少し複雑なのは、複合文があり、libが文ごとに1つのタプルを生成するという事実です-パイプ構造のリストをエンジンに渡してより多くのタプルを生成できるようにする可能性をすぐに追加します（私のプロジェクトでも必要です）文ごと。しかし今のところ、複合送信用の2番目のエンジンを作成するだけで解決できます。このエンジンの構造は、VERBではなくVERB/conj/VERBのみが異なります（これらのregexは常にROOTから始まるため、VERB/conj/VERBは2番目の動詞につながります。重文）：

pipes_structure_comp = [SequencePipe([FindTokensPipe("VERB/conj/VERB/nsubj/NNP"),
                                 NamedEntityFilterPipe(),
                                 NamedEntityExtractorPipe()]),
                   AggregatePipe([FindTokensPipe("VERB/conj/VERB"),
                                  FindTokensPipe("VERB/conj/VERB/xcomp/VERB/aux/*"),
                                  FindTokensPipe("VERB/conj/VERB/xcomp/VERB")]),
                   AnyPipe([FindTokensPipe("VERB/conj/VERB/[acomp,amod]/ADJ"),
                            AggregatePipe([FindTokensPipe("VERB/conj/VERB/[dobj,attr]/NOUN/det/DET"),
                                           FindTokensPipe("VERB/conj/VERB/[dobj,attr]/NOUN/[acomp,amod]/ADJ")])])
                  ]

engine2 = PipelineEngine(pipes_structure_comp, Context(doc), [0,1,2])

そして今、両方のエンジンを実行した後、期待される結果が得られます:)

engine.process()
engine2.process()
[([Donald, Trump], [is], [the, worst])]
[([Hillary], [is], [better])]

これがあなたに必要なことだと思います。もちろん、特定の例文のパイプ構造をすばやく作成しただけで、すべての場合に機能するわけではありませんが、多くの文構造を確認し、すでにかなりの割合を満たしていますが、FindTokensPipeなどを追加するだけです。現在は機能しないケースで、いくつかの調整を行うと、非常に多くの可能な文をカバーできると確信しています（英語はそれほど複雑ではないので... :)

Answer 2

私はspacyライブラリをさらに調べていましたが、依存関係の管理を通じて最終的に解決策を見つけました。 this repoのおかげで、形容詞も主観的な動詞オブジェクトに含める方法（SVAOのものにする）と、クエリで複合主題を削除する方法を理解しました。これが私の解決策です：

from nltk.stem.wordnet import WordNetLemmatizer
from spacy.lang.en import English

SUBJECTS = ["nsubj", "nsubjpass", "csubj", "csubjpass", "agent", "expl"]
OBJECTS = ["dobj", "dative", "attr", "oprd"]
ADJECTIVES = ["acomp", "advcl", "advmod", "amod", "appos", "nn", "nmod", "ccomp", "complm",
              "hmod", "infmod", "xcomp", "rcmod", "poss"," possessive"]
COMPOUNDS = ["compound"]
PREPOSITIONS = ["prep"]

def getSubsFromConjunctions(subs):
    moreSubs = []
    for sub in subs:
        # rights is a generator
        rights = list(sub.rights)
        rightDeps = {tok.lower_ for tok in rights}
        if "and" in rightDeps:
            moreSubs.extend([tok for tok in rights if tok.dep_ in SUBJECTS or tok.pos_ == "NOUN"])
            if len(moreSubs) > 0:
                moreSubs.extend(getSubsFromConjunctions(moreSubs))
    return moreSubs

def getObjsFromConjunctions(objs):
    moreObjs = []
    for obj in objs:
        # rights is a generator
        rights = list(obj.rights)
        rightDeps = {tok.lower_ for tok in rights}
        if "and" in rightDeps:
            moreObjs.extend([tok for tok in rights if tok.dep_ in OBJECTS or tok.pos_ == "NOUN"])
            if len(moreObjs) > 0:
                moreObjs.extend(getObjsFromConjunctions(moreObjs))
    return moreObjs

def getVerbsFromConjunctions(verbs):
    moreVerbs = []
    for verb in verbs:
        rightDeps = {tok.lower_ for tok in verb.rights}
        if "and" in rightDeps:
            moreVerbs.extend([tok for tok in verb.rights if tok.pos_ == "VERB"])
            if len(moreVerbs) > 0:
                moreVerbs.extend(getVerbsFromConjunctions(moreVerbs))
    return moreVerbs

def findSubs(tok):
    head = tok.head
    while head.pos_ != "VERB" and head.pos_ != "NOUN" and head.head != head:
        head = head.head
    if head.pos_ == "VERB":
        subs = [tok for tok in head.lefts if tok.dep_ == "SUB"]
        if len(subs) > 0:
            verbNegated = isNegated(head)
            subs.extend(getSubsFromConjunctions(subs))
            return subs, verbNegated
        Elif head.head != head:
            return findSubs(head)
    Elif head.pos_ == "NOUN":
        return [head], isNegated(tok)
    return [], False

def isNegated(tok):
    negations = {"no", "not", "n't", "never", "none"}
    for dep in list(tok.lefts) + list(tok.rights):
        if dep.lower_ in negations:
            return True
    return False

def findSVs(tokens):
    svs = []
    verbs = [tok for tok in tokens if tok.pos_ == "VERB"]
    for v in verbs:
        subs, verbNegated = getAllSubs(v)
        if len(subs) > 0:
            for sub in subs:
                svs.append((sub.orth_, "!" + v.orth_ if verbNegated else v.orth_))
    return svs

def getObjsFromPrepositions(deps):
    objs = []
    for dep in deps:
        if dep.pos_ == "ADP" and dep.dep_ == "prep":
            objs.extend([tok for tok in dep.rights if tok.dep_  in OBJECTS or (tok.pos_ == "PRON" and tok.lower_ == "me")])
    return objs

def getAdjectives(toks):
    toks_with_adjectives = []
    for tok in toks:
        adjs = [left for left in tok.lefts if left.dep_ in ADJECTIVES]
        adjs.append(tok)
        adjs.extend([right for right in tok.rights if tok.dep_ in ADJECTIVES])
        tok_with_adj = " ".join([adj.lower_ for adj in adjs])
        toks_with_adjectives.extend(adjs)

    return toks_with_adjectives

def getObjsFromAttrs(deps):
    for dep in deps:
        if dep.pos_ == "NOUN" and dep.dep_ == "attr":
            verbs = [tok for tok in dep.rights if tok.pos_ == "VERB"]
            if len(verbs) > 0:
                for v in verbs:
                    rights = list(v.rights)
                    objs = [tok for tok in rights if tok.dep_ in OBJECTS]
                    objs.extend(getObjsFromPrepositions(rights))
                    if len(objs) > 0:
                        return v, objs
    return None, None

def getObjFromXComp(deps):
    for dep in deps:
        if dep.pos_ == "VERB" and dep.dep_ == "xcomp":
            v = dep
            rights = list(v.rights)
            objs = [tok for tok in rights if tok.dep_ in OBJECTS]
            objs.extend(getObjsFromPrepositions(rights))
            if len(objs) > 0:
                return v, objs
    return None, None

def getAllSubs(v):
    verbNegated = isNegated(v)
    subs = [tok for tok in v.lefts if tok.dep_ in SUBJECTS and tok.pos_ != "DET"]
    if len(subs) > 0:
        subs.extend(getSubsFromConjunctions(subs))
    else:
        foundSubs, verbNegated = findSubs(v)
        subs.extend(foundSubs)
    return subs, verbNegated

def getAllObjs(v):
    # rights is a generator
    rights = list(v.rights)
    objs = [tok for tok in rights if tok.dep_ in OBJECTS]
    objs.extend(getObjsFromPrepositions(rights))

    potentialNewVerb, potentialNewObjs = getObjFromXComp(rights)
    if potentialNewVerb is not None and potentialNewObjs is not None and len(potentialNewObjs) > 0:
        objs.extend(potentialNewObjs)
        v = potentialNewVerb
    if len(objs) > 0:
        objs.extend(getObjsFromConjunctions(objs))
    return v, objs

def getAllObjsWithAdjectives(v):
    # rights is a generator
    rights = list(v.rights)
    objs = [tok for tok in rights if tok.dep_ in OBJECTS]

    if len(objs)== 0:
        objs = [tok for tok in rights if tok.dep_ in ADJECTIVES]

    objs.extend(getObjsFromPrepositions(rights))

    potentialNewVerb, potentialNewObjs = getObjFromXComp(rights)
    if potentialNewVerb is not None and potentialNewObjs is not None and len(potentialNewObjs) > 0:
        objs.extend(potentialNewObjs)
        v = potentialNewVerb
    if len(objs) > 0:
        objs.extend(getObjsFromConjunctions(objs))
    return v, objs

def findSVOs(tokens):
    svos = []
    verbs = [tok for tok in tokens if tok.pos_ == "VERB" and tok.dep_ != "aux"]
    for v in verbs:
        subs, verbNegated = getAllSubs(v)
        # hopefully there are subs, if not, don't examine this verb any longer
        if len(subs) > 0:
            v, objs = getAllObjs(v)
            for sub in subs:
                for obj in objs:
                    objNegated = isNegated(obj)
                    svos.append((sub.lower_, "!" + v.lower_ if verbNegated or objNegated else v.lower_, obj.lower_))
    return svos

def findSVAOs(tokens):
    svos = []
    verbs = [tok for tok in tokens if tok.pos_ == "VERB" and tok.dep_ != "aux"]
    for v in verbs:
        subs, verbNegated = getAllSubs(v)
        # hopefully there are subs, if not, don't examine this verb any longer
        if len(subs) > 0:
            v, objs = getAllObjsWithAdjectives(v)
            for sub in subs:
                for obj in objs:
                    objNegated = isNegated(obj)
                    obj_desc_tokens = generate_left_right_adjectives(obj)
                    sub_compound = generate_sub_compound(sub)
                    svos.append((" ".join(tok.lower_ for tok in sub_compound), "!" + v.lower_ if verbNegated or objNegated else v.lower_, " ".join(tok.lower_ for tok in obj_desc_tokens)))
    return svos

def generate_sub_compound(sub):
    sub_compunds = []
    for tok in sub.lefts:
        if tok.dep_ in COMPOUNDS:
            sub_compunds.extend(generate_sub_compound(tok))
    sub_compunds.append(sub)
    for tok in sub.rights:
        if tok.dep_ in COMPOUNDS:
            sub_compunds.extend(generate_sub_compound(tok))
    return sub_compunds

def generate_left_right_adjectives(obj):
    obj_desc_tokens = []
    for tok in obj.lefts:
        if tok.dep_ in ADJECTIVES:
            obj_desc_tokens.extend(generate_left_right_adjectives(tok))
    obj_desc_tokens.append(obj)

    for tok in obj.rights:
        if tok.dep_ in ADJECTIVES:
            obj_desc_tokens.extend(generate_left_right_adjectives(tok))

    return obj_desc_tokens

次のようなクエリを渡すと、次のようになります。

from spacy.lang.en import English
parser = English()

sentence = u"""
Donald Trump is the worst president of USA, but Hillary is better than him
"""

parse = parser(sentence)
print(findSVAOs(parse))

次のようになります。

[(u'donald trump', u'is', u'worst president'), (u'hillary', u'is', u'better')]

@Krzysiekにもソリューションをありがとうございます。実際、ライブラリを深く掘り下げて変更することはできませんでした。私はむしろ私の問題を解決するために上記のリンクを変更しようとしました。