def retrofitter():
return Retrofitter(max_iter=100,
alpha=None,
beta=None,
tol=1e-2,
verbose=False,
introspecting=False)
def system_2():
# Data------------
with open(wordentail_filename) as f:
wordentail_data = json.load(f)
X_glove = pd.DataFrame(GLOVE).T
print(X_glove.shape)
def convert_edges_to_indices(edges, Q):
lookup = dict(zip(Q.index, range(Q.shape[0])))
index_edges = defaultdict(set)
for start, finish_nodes in edges.items():
s = lookup.get(start)
if s:
f = {lookup[n] for n in finish_nodes if n in lookup}
if f:
index_edges[s] = f
return index_edges
wn_index_edges = convert_edges_to_indices(wn_edges, X_glove)
wn_retro = Retrofitter(verbose=True)
X_retro = wn_retro.fit(X_glove, wn_index_edges)
print(X_retro.shape)
def retro_vec(w):
"""Return `w`'s Retrofitted representation if available, else return
a random vector."""
return X_retro.loc[w].values if w in X_retro.index else randvec(
w, n=glove_dim)
# Model-----------
net = TorchShallowNeuralClassifier(hidden_dim=50, max_iter=100)
print(net)
# Exp-------------
result = nli.wordentail_experiment(
train_data=wordentail_data['word_disjoint']['train'],
assess_data=wordentail_data['word_disjoint']['dev'],
model=net,
vector_func=retro_vec,
vector_combo_func=vec_concatenate)
return result['macro-F1']
# ### As before, but now 2 has no outgoing edges
# In[7]:
edges_isolated = {0: {1, 2}, 1: {0, 2}, 2: set()}
_ = retrofitting.plot_retro_path(Q_hat, edges_isolated)
# ### All nodes connected to all others, but $\alpha = 0$
# In[8]:
_ = retrofitting.plot_retro_path(Q_hat,
edges_all,
retrofitter=Retrofitter(alpha=lambda x: 0))
# ## WordNet
#
# Faruqui et al. conduct experiments on three knowledge graphs: [WordNet](https://wordnet.princeton.edu), [FrameNet](https://framenet.icsi.berkeley.edu/fndrupal/), and the [Penn Paraphrase Database (PPDB)](http://paraphrase.org/). [The repository for their paper](https://github.com/mfaruqui/retrofitting) includes the graphs that they derived for their experiments.
#
# Here, we'll reproduce just one of the two WordNet experiments they report, in which the graph is formed based on synonymy.
# ### Background on WordNet
#
# WordNet is an incredible, hand-built lexical resource capturing a wealth of information about English words and their inter-relationships. ([Here is a collection of WordNets in other languages.](http://globalwordnet.org)) For a detailed overview using NLTK, see [this tutorial](http://compprag.christopherpotts.net/wordnet.html).
#
# The core concepts:
#
# * A __lemma__ is something like our usual notion of __word__. Lemmas are highly sense-disambiguated. For instance, there are six lemmas that are consistent with the string `crane`: the bird, the machine, the poets, ...
#
return edges
wn_edges = get_wordnet_edges()
data_home = 'data'
glove_dict = utils.glove2dict(
os.path.join(data_home, 'glove.6B', 'glove.6B.300d.txt'))
X_glove = pd.DataFrame(glove_dict).T
print(X_glove.shape)
def convert_edges_to_indices(edges, Q):
lookup = dict(zip(Q.index, range(Q.shape[0])))
index_edges = defaultdict(set)
for start, finish_nodes in edges.items():
s = lookup.get(start)
if s:
f = {lookup[n] for n in finish_nodes if n in lookup}
if f:
index_edges[s] = f
return index_edges
wn_index_edges = convert_edges_to_indices(wn_edges, X_glove)
wn_retro = Retrofitter(verbose=True)
X_retro = wn_retro.fit(X_glove, wn_index_edges)
print(X_retro.shape)