Dependency Parsing

This tutorial is available as an IPython notebook at Malaya/example/dependency.

This module only trained on standard language structure, so it is not save to use it for local language structure.

[1]:
%%time
import malaya
CPU times: user 4.67 s, sys: 658 ms, total: 5.33 s
Wall time: 4.62 s

Describe supported dependencies

[2]:
malaya.dependency.describe()
INFO:root:you can read more from https://universaldependencies.org/treebanks/id_pud/index.html
[2]:
Tag Description
0 acl clausal modifier of noun
1 advcl adverbial clause modifier
2 advmod adverbial modifier
3 amod adjectival modifier
4 appos appositional modifier
5 aux auxiliary
6 case case marking
7 ccomp clausal complement
8 advmod adverbial modifier
9 compound compound
10 compound:plur plural compound
11 conj conjunct
12 cop cop
13 csubj clausal subject
14 dep dependent
15 det determiner
16 fixed multi-word expression
17 flat name
18 iobj indirect object
19 mark marker
20 nmod nominal modifier
21 nsubj nominal subject
22 obj direct object
23 parataxis parataxis
24 root root
25 xcomp open clausal complement
[3]:
string = 'Dr Mahathir menasihati mereka supaya berhenti berehat dan tidur sebentar sekiranya mengantuk ketika memandu.'

List available transformer Dependency models

[4]:
malaya.dependency.available_transformer()
INFO:root:tested on 20% test set.
[4]:
Size (MB) Quantized Size (MB) Arc Accuracy Types Accuracy Root Accuracy
bert 426.0 112.0 0.855 0.848 0.920
tiny-bert 59.5 15.7 0.718 0.694 0.886
albert 50.0 13.2 0.811 0.793 0.879
tiny-albert 24.8 6.6 0.708 0.673 0.817
xlnet 450.2 119.0 0.931 0.925 0.947
alxlnet 50.0 14.3 0.894 0.886 0.942

Make sure you can check accuracy chart from here first before select a model, https://malaya.readthedocs.io/en/latest/Accuracy.html#dependency-parsing

The best model in term of accuracy is XLNET.

Load xlnet dependency model

[5]:
model = malaya.dependency.transformer(model = 'xlnet')
WARNING:tensorflow:From /Users/huseinzolkepli/Documents/Malaya/malaya/function/__init__.py:54: The name tf.gfile.GFile is deprecated. Please use tf.io.gfile.GFile instead.

WARNING:tensorflow:From /Users/huseinzolkepli/Documents/Malaya/malaya/function/__init__.py:55: The name tf.GraphDef is deprecated. Please use tf.compat.v1.GraphDef instead.

WARNING:tensorflow:From /Users/huseinzolkepli/Documents/Malaya/malaya/function/__init__.py:49: The name tf.InteractiveSession is deprecated. Please use tf.compat.v1.InteractiveSession instead.

Load Quantized model

To load 8-bit quantized model, simply pass quantized = True, default is False.

We can expect slightly accuracy drop from quantized model, and not necessary faster than normal 32-bit float model, totally depends on machine.

[10]:
quantized_model = malaya.dependency.transformer(model = 'xlnet', quantized = True)
WARNING:root:Load quantized model will cause accuracy drop.
[6]:
d_object, tagging, indexing = model.predict(string)
d_object.to_graphvis()
[6]:
_images/load-dependency_14_0.svg
[11]:
d_object, tagging, indexing = quantized_model.predict(string)
d_object.to_graphvis()
[11]:
_images/load-dependency_15_0.svg

Voting stack model

[8]:
alxlnet = malaya.dependency.transformer(model = 'alxlnet')
tagging, indexing = malaya.stack.voting_stack([model, alxlnet, model], string)
malaya.dependency.dependency_graph(tagging, indexing).to_graphvis()
downloading frozen /Users/huseinzolkepli/Malaya/dependency/alxlnet/base model
51.0MB [00:50, 1.01MB/s]
[8]:
_images/load-dependency_17_2.svg

Dependency graph object

To initiate a dependency graph from dependency models, you need to call malaya.dependency.dependency_graph.

[9]:
graph = malaya.dependency.dependency_graph(tagging, indexing)
graph
[9]:
<malaya.function.parse_dependency.DependencyGraph at 0x164e67e90>

generate graphvis

[10]:
graph.to_graphvis()
[10]:
_images/load-dependency_21_0.svg

Get nodes

[11]:
graph.nodes
[11]:
defaultdict(<function malaya.function.parse_dependency.DependencyGraph.__init__.<locals>.<lambda>()>,
            {0: {'address': 0,
              'word': None,
              'lemma': None,
              'ctag': 'TOP',
              'tag': 'TOP',
              'feats': None,
              'head': None,
              'deps': defaultdict(list, {'root': [3]}),
              'rel': None},
             1: {'address': 1,
              'word': 'Dr',
              'lemma': '_',
              'ctag': '_',
              'tag': '_',
              'feats': '_',
              'head': 3,
              'deps': defaultdict(list, {'flat': [2]}),
              'rel': 'nsubj'},
             3: {'address': 3,
              'word': 'menasihati',
              'lemma': '_',
              'ctag': '_',
              'tag': '_',
              'feats': '_',
              'head': 0,
              'deps': defaultdict(list,
                          {'nsubj': [1], 'obj': [4], 'ccomp': [6]}),
              'rel': 'root'},
             2: {'address': 2,
              'word': 'Mahathir',
              'lemma': '_',
              'ctag': '_',
              'tag': '_',
              'feats': '_',
              'head': 1,
              'deps': defaultdict(list, {}),
              'rel': 'flat'},
             4: {'address': 4,
              'word': 'mereka',
              'lemma': '_',
              'ctag': '_',
              'tag': '_',
              'feats': '_',
              'head': 3,
              'deps': defaultdict(list, {}),
              'rel': 'obj'},
             5: {'address': 5,
              'word': 'supaya',
              'lemma': '_',
              'ctag': '_',
              'tag': '_',
              'feats': '_',
              'head': 6,
              'deps': defaultdict(list, {}),
              'rel': 'case'},
             6: {'address': 6,
              'word': 'berhenti',
              'lemma': '_',
              'ctag': '_',
              'tag': '_',
              'feats': '_',
              'head': 3,
              'deps': defaultdict(list,
                          {'case': [5], 'ccomp': [7], 'conj': [9]}),
              'rel': 'ccomp'},
             7: {'address': 7,
              'word': 'berehat',
              'lemma': '_',
              'ctag': '_',
              'tag': '_',
              'feats': '_',
              'head': 6,
              'deps': defaultdict(list, {}),
              'rel': 'ccomp'},
             8: {'address': 8,
              'word': 'dan',
              'lemma': '_',
              'ctag': '_',
              'tag': '_',
              'feats': '_',
              'head': 9,
              'deps': defaultdict(list, {}),
              'rel': 'cc'},
             9: {'address': 9,
              'word': 'tidur',
              'lemma': '_',
              'ctag': '_',
              'tag': '_',
              'feats': '_',
              'head': 6,
              'deps': defaultdict(list,
                          {'cc': [8],
                           'advmod': [10],
                           'amod': [12],
                           'advcl': [14]}),
              'rel': 'conj'},
             10: {'address': 10,
              'word': 'sebentar',
              'lemma': '_',
              'ctag': '_',
              'tag': '_',
              'feats': '_',
              'head': 9,
              'deps': defaultdict(list, {}),
              'rel': 'advmod'},
             11: {'address': 11,
              'word': 'sekiranya',
              'lemma': '_',
              'ctag': '_',
              'tag': '_',
              'feats': '_',
              'head': 12,
              'deps': defaultdict(list, {}),
              'rel': 'advmod'},
             12: {'address': 12,
              'word': 'mengantuk',
              'lemma': '_',
              'ctag': '_',
              'tag': '_',
              'feats': '_',
              'head': 9,
              'deps': defaultdict(list, {'advmod': [11]}),
              'rel': 'amod'},
             13: {'address': 13,
              'word': 'ketika',
              'lemma': '_',
              'ctag': '_',
              'tag': '_',
              'feats': '_',
              'head': 14,
              'deps': defaultdict(list, {}),
              'rel': 'case'},
             14: {'address': 14,
              'word': 'memandu.',
              'lemma': '_',
              'ctag': '_',
              'tag': '_',
              'feats': '_',
              'head': 9,
              'deps': defaultdict(list, {'case': [13]}),
              'rel': 'advcl'}})

Flat the graph

[12]:
list(graph.triples())
[12]:
[(('menasihati', '_'), 'nsubj', ('Dr', '_')),
 (('Dr', '_'), 'flat', ('Mahathir', '_')),
 (('menasihati', '_'), 'obj', ('mereka', '_')),
 (('menasihati', '_'), 'ccomp', ('berhenti', '_')),
 (('berhenti', '_'), 'case', ('supaya', '_')),
 (('berhenti', '_'), 'ccomp', ('berehat', '_')),
 (('berhenti', '_'), 'conj', ('tidur', '_')),
 (('tidur', '_'), 'cc', ('dan', '_')),
 (('tidur', '_'), 'advmod', ('sebentar', '_')),
 (('tidur', '_'), 'amod', ('mengantuk', '_')),
 (('mengantuk', '_'), 'advmod', ('sekiranya', '_')),
 (('tidur', '_'), 'advcl', ('memandu.', '_')),
 (('memandu.', '_'), 'case', ('ketika', '_'))]

Check the graph contains cycles

[13]:
graph.contains_cycle()
[13]:
False

Generate networkx

Make sure you already installed networkx,

pip install networkx
[14]:
digraph = graph.to_networkx()
digraph
[14]:
<networkx.classes.multidigraph.MultiDiGraph at 0x1a875a110>
[15]:
import networkx as nx
import matplotlib.pyplot as plt
nx.draw_networkx(digraph)
plt.show()
<Figure size 640x480 with 1 Axes>
[16]:
digraph.edges()
[16]:
OutMultiEdgeDataView([(1, 3), (2, 1), (4, 3), (5, 6), (6, 3), (7, 6), (8, 9), (9, 6), (10, 9), (11, 12), (12, 9), (13, 14), (14, 9)])
[17]:
digraph.nodes()
[17]:
NodeView((1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14))
[18]:
labels = {i:graph.get_by_address(i)['word'] for i in digraph.nodes()}
labels
[18]:
{1: 'Dr',
 2: 'Mahathir',
 3: 'menasihati',
 4: 'mereka',
 5: 'supaya',
 6: 'berhenti',
 7: 'berehat',
 8: 'dan',
 9: 'tidur',
 10: 'sebentar',
 11: 'sekiranya',
 12: 'mengantuk',
 13: 'ketika',
 14: 'memandu.'}
[19]:
plt.figure(figsize=(15,5))
nx.draw_networkx(digraph,labels=labels)
plt.show()
_images/load-dependency_34_0.png

Vectorize

Let say you want to visualize word level in lower dimension, you can use model.vectorize,

def vectorize(self, string: str):
    """
    vectorize a string.

    Parameters
    ----------
    string: List[str]

    Returns
    -------
    result: np.array
    """
[6]:
r = quantized_model.vectorize(string)
[7]:
x = [i[0] for i in r]
y = [i[1] for i in r]
[8]:
from sklearn.manifold import TSNE
import matplotlib.pyplot as plt

tsne = TSNE().fit_transform(y)
tsne.shape
[8]:
(14, 2)
[9]:
plt.figure(figsize = (7, 7))
plt.scatter(tsne[:, 0], tsne[:, 1])
labels = x
for label, x, y in zip(
    labels, tsne[:, 0], tsne[:, 1]
):
    label = (
        '%s, %.3f' % (label[0], label[1])
        if isinstance(label, list)
        else label
    )
    plt.annotate(
        label,
        xy = (x, y),
        xytext = (0, 0),
        textcoords = 'offset points',
    )
_images/load-dependency_39_0.png
[ ]: