def test_tf_metrics(self):
summarizer = LexRankSummarizer()
sentences = [
("this", "sentence", "is", "simple", "sentence"),
("this", "is", "simple", "sentence", "yes", "is", "too", "too",
"too"),
]
metrics = summarizer._compute_tf(sentences)
expected = [
{
"this": 1 / 2,
"is": 1 / 2,
"simple": 1 / 2,
"sentence": 1.0
},
{
"this": 1 / 3,
"is": 2 / 3,
"yes": 1 / 3,
"simple": 1 / 3,
"sentence": 1 / 3,
"too": 1.0
},
]
self.assertEqual(expected, metrics)
class LexRankSumy(Baseline):
""" Description from https://arxiv.org/pdf/1905.13164.pdf
LexRank (Erkan and Radev, 2004) is a widely-used graph-based extractive summarizer;
we build a graph with paragraphs as nodes andedges weighted by tf-idf cosine similarity;
we run a PageRank-like algorithm on this graph to rank and select paragraphs until
the length of the ground-truth summary is reached.
"""
""" Implementation
Wrapper of https://github.com/miso-belica/sumy
"""
def __init__(self, name, language):
super().__init__(name)
self.language = language
self.summarizer = LexRankSummarizer()
def rank_sentences(self, dataset, document_column_name, **kwargs):
all_sentences = []
all_scores = []
for document in tqdm(dataset[document_column_name]):
sentences, scores = self.run_single(document)
all_sentences.append(sentences)
all_scores.append(scores)
data = [{
"sentences": sentences,
"scores": scores
} for sentences, scores in zip(all_sentences, all_scores)]
return Baseline.append_column(dataset, data, self.name)
def run_single(self, document):
parser = PlaintextParser.from_string(document,
Tokenizer(self.language))
document = parser.document
self.summarizer._ensure_dependencies_installed()
sentences_words = [
self.summarizer._to_words_set(s) for s in document.sentences
]
if not sentences_words:
return tuple()
tf_metrics = self.summarizer._compute_tf(sentences_words)
idf_metrics = self.summarizer._compute_idf(sentences_words)
matrix = self.summarizer._create_matrix(sentences_words,
self.summarizer.threshold,
tf_metrics, idf_metrics)
scores = self.summarizer.power_method(matrix, self.summarizer.epsilon)
return list(map(str, document.sentences)), list(scores)
def test_tf_metrics(self):
summarizer = LexRankSummarizer()
sentences = [
("this", "sentence", "is", "simple", "sentence"),
("this", "is", "simple", "sentence", "yes", "is", "too", "too", "too"),
]
metrics = summarizer._compute_tf(sentences)
expected = [
{"this": 1/2, "is": 1/2, "simple": 1/2, "sentence": 1.0},
{"this": 1/3, "is": 2/3, "yes": 1/3, "simple": 1/3, "sentence": 1/3, "too": 1.0},
]
self.assertEqual(expected, metrics)
def lexrank_scoring(text: str) -> Tuple[List[str], np.ndarray]:
"""
LexRankアルゴリズムによって文に点数をつける。
この点数は文の重要度とみなすことができる。
Parameters
----------
text : str
分析対象のテキスト。
Returns
-------
List[str]
text を文のリストに分解したもの。
np.ndarray
文のリストに対応する重要度のリスト。
"""
doc = nlp(text)
# 文のリストと単語のリストをつくる
sentences = []
corpus = []
for sent in doc.sents:
sentences.append(sent.text)
tokens = []
for token in sent:
# 文に含まれる単語のうち, 名詞・副詞・形容詞・動詞に限定する
if token.pos_ in ('NOUN', 'ADV', 'ADJ', 'VERB'):
# ぶれをなくすため, 単語の見出し語 Token.lemma_ を使う
tokens.append(token.lemma_)
corpus.append(tokens)
# sentences = [文0, 文1, ...]
# corpus = [[文0の単語0, 文0の単語1, ...], [文1の単語0, 文1の単語1, ...], ...]
# sumyライブラリによるLexRankスコアリング
lxr = LexRankSummarizer()
tf_metrics = lxr._compute_tf(corpus)
idf_metrics = lxr._compute_idf(corpus)
matrix = lxr._create_matrix(corpus, lxr.threshold, tf_metrics, idf_metrics)
scores = lxr.power_method(matrix, lxr.epsilon)
# scores = [文0の重要度, 文1の重要度, ...]
return sentences, scores
