class TestSearch(unittest.TestCase):
def setUp(self):
self.search = Search()
def test_search(self):
mylist = [("dLondon", "Location"), ("Obama", "Person"), ("UNICEF", "Organization"), ("jhfvjhegrfvh", "Location"), ("Noor_Nakhaei", "Person"), ("Center_For_Smart_Health", "Organization")]
new, existing = self.search.query(mylist)
assert existing == [('Obama', 'Person'), ('UNICEF', 'Organization')]
assert new == [("dLondon", "Location"), ("jhfvjhegrfvh", "Location"), ("Noor_Nakhaei", "Person"), ("Center_For_Smart_Health", "Organization")]
class Graph(object):
"""
Base class which represents the heterogeneous textual graph (undirected graph). G = <V, E>. V is the set of nodes (objects), including 3 types of objects (i.e. new entites, known entities, and contextual words). Entities are words (with label: PERSON, LOCATION, and ORGANIZATION) whereas contextual words are the remaining uni-gram words. New entities are the entities not in DBpedia, and Know entities are the entities in the DBpedia. E is a set of edges (co-occurrences) of entity-entity, entity-word, and word-word corrences. Words within every 5-word sliding window in a news sentence are considered to be co-occuring with each other. The weights are represented by adjacency matrix using dataframe.
Attributes:
nodes: dictionary of nodes {"N (new entity)": [(word, label)], "K (Known entity)": [(word, label)], "C (Contextual word)": [(word, label)]} in the graph; Includes 3 types of objects (i.e. e new entites, known entities, and contextual words).
edges: set contains the tuples. e.g. ("A", "B") indicates a link between node "A" and node "B".
weights: The weights are represented by adjacency matrix using dataframe.
news: list of news articles (articles are string type).
"""
def __init__(self, lst_news, window_size=5):
"""Inits Graph
Args:
lst_news: list of string. list of news articles.
"""
self.window_size = window_size
self.news = list(lst_news)
self.reader = Reader()
self.search = Search()
self.nodes = self.__create_nodes()
self.edges = self.__create_edges()
self.edge_weights = self.__create_weights()
def __create_nodes(self):
"""Private class method
Takes in a list of news articles (articles are string types):
1) tokenize the articles
2) remove stopwords
3) label words with 3 labels (i.e. PERSON, ORGANIZATION, LOCATION)
4) Match entities (i.e. person, org, loc) against DBpedia
Returns:
Returns a dictionary contains 3 types of objects (i.e. new entites, known entities, and contextual words). E.g. {"N": [("Washington", "LOCATION")], "K":[("Trump", "PERSON"), ("Hua Wei", "ORGANIZATION")], "C": [("the", "O"), ("am", "O")]}
"""
# parse news articles
tagged_words = self.reader.parse_news(self.news)
# seperate entities from contextual words
entities, cwords = self.__entities_words(tagged_words)
new_e, known_e = self.search.query(entities)
ret = dict()
ret["N"] = list(set(new_e))
ret["K"] = list(set(known_e))
ret["C"] = list(set(cwords))
return dict(ret)
def get_nodes(self):
"""
Getter method which returns all nodes from self.nodes.
"""
ret = set()
for i in self.nodes["N"]:
ret.add(i[0])
for i in self.nodes["K"]:
ret.add(i[0])
for i in self.nodes["C"]:
ret.add(i[0])
return list(ret)
def get_entities(self):
"""
Getter method which returns a list of entities (i.e. word tagged with "PERSON", "LOCATION", "ORGANIZATION") from self.nodes.
"""
ret = set()
for i in self.nodes["N"]:
ret.add(i[0])
for i in self.nodes["K"]:
ret.add(i[0])
return list(ret)
def get_words(self):
"""
Getter method which returns a list of contextual words from self.nodes.
"""
ret = set()
for i in self.nodes["C"]:
ret.add(i[0])
return list(ret)
def __create_edges(self, window_size=5):
"""Private class method
Takes in a list of news articles, and extract the co-occurring links between nodes. Nodes within 5-word sliding window in a news sentence are considered to be co-occuring with each other. The frequncies of nodes co-appearing in news sentences as weights of these links.
Returns:
Returns a set of links between nodes.
"""
e = set()
for article in self.news:
self.tokenized_text = word_tokenize(article)
self.tokenized_text = self.reader.filter_stop_words(
self.tokenized_text)
generator = self.sliding_window(self.tokenized_text,
self.window_size)
for t in generator:
e = e.union(set(itertools.combinations(t, 2)))
return set(e)
def get_edges(self):
"""
Getter method which returns a set of edges from self.edges.
"""
return set(self.edges)
def sliding_window(self, seq, n=5):
"""
Returns a sliding window (of width n) over data from the iterable
s -> (s0,s1,...s[n-1]), (s1,s2,...,sn), ...
Args:
seq: list of words; This is one news article splitted into a list of words.
n: int; size of the sliding window
Returns:
An iterator contains all the sliced window. See the test case in `tests/test_graph.py` for more details.
"""
it = iter(seq)
result = tuple(islice(it, n))
if len(result) <= n:
yield result
for elem in it:
result = result[1:] + (elem, )
yield result
def __create_weights(self):
"""Private class method
Create weights matrix using pandas dataframe. The value at ith row and jth row is the counts of links (undirected) between node i and node j.
Returns:
Return a copy of dataframe representing the weights matrix.
"""
words = self.get_nodes()
df = pd.DataFrame(index=words, columns=words).fillna(0)
for article in self.news:
self.tokenized_text = word_tokenize(article)
self.tokenized_text = self.reader.filter_stop_words(
self.tokenized_text)
generator = self.sliding_window(self.tokenized_text,
self.window_size)
for t in generator:
for tup in set(itertools.combinations(t, 2)):
if tup[0] != tup[1]:
df.loc[tup[0], tup[1]] += 1
df.loc[tup[1], tup[0]] += 1
return df.copy()
def get_weights(self):
return self.edge_weights.copy()
def __entities_words(self, tagged_words):
"""Private class method
Seperate the entity words from the comtextual words.
Args:
tagged_words: list of strings; a list of tuples (word, label)
Returns:
entities: words tagged with "PERSON", "LOCATION", "ORGANIZATION".
cwords: words tagged with "O"
"""
entities = list()
cwords = list()
for word in tagged_words:
if word[1] == "O":
# contextual words
cwords.append(word)
else:
entities.append(word)
assert len(entities) + len(cwords) == len(tagged_words)
return entities, cwords
def update_weight(self, e, w):
"""
Update the edge weight in the enternal weight matrix.
Args:
e: tuple; a tuple contains two nodes, e.g. ("A", "B")
w: int; The new weight associated with e
"""
if e[0] not in set(self.get_nodes()):
raise ValueError("Node {} is not in the graph".format(str(e[0])))
if e[1] not in set(self.get_nodes()):
raise ValueError("Node {} is not in the graph".format(str(e[1])))
if e in self.edges and w <= 0:
self.edge_weights.loc[e[0], e[1]] = w
self.edge_weights.loc[e[1], e[0]] = w
self.edges.remove(e)
self.edges.remove((e[1], e[0]))
elif e in self.edges and w > 0:
self.edge_weights.loc[e[0], e[1]] = w
self.edge_weights.loc[e[1], e[0]] = w
else:
self.edge_weights.loc[e[0], e[1]] = w
self.edge_weights.loc[e[1], e[0]] = w
self.edges.add(e)