aspect-based sentiment classification with aspect-specific ... · task 4, semeval2015 task 12 and...
TRANSCRIPT
![Page 1: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/1.jpg)
Aspect-based Sentiment Classification with Aspect-specific Graph Convolutional Networks
Chen Zhang1, Qiuchi Li2, Dawei Song1
1Beijing Institute of Technology, China. {czhang, dwsong}@bit.edu.cn2University of Padua, Italy. {qiuchili}@dei.unipd.it
EMNLP 2019, Hong Kong, China
![Page 2: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/2.jpg)
Content
1
• Problem Formulation
• Related Work
• Motivation
• Proposed Model – ASGCN
• Experiments
• Discussion
• Conclusions & Future Work
![Page 3: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/3.jpg)
Content
2
• Problem Formulation
• Related Work
• Motivation
• Proposed Model – ASGCN
• Experiments
• Discussion
• Conclusions & Future Work
![Page 4: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/4.jpg)
3
• Aspect-based Sentiment Classification (ABSC): aiming at identifying the
sentiment polarities of aspect terms explicitly given in sentences.
• Aspect terms (a.k.a. opinion targets): operating system, Windows
• Corresponding sentiment polarities: positive, negative
Problem Formulation
From the speed to the multi-touch gestures this operating system beats Windows easily.
![Page 5: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/5.jpg)
Content
4
• Problem Formulation
• Related Work
• Motivation
• Proposed Model – ASGCN
• Experiments
• Discussion
• Conclusions & Future Work
![Page 6: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/6.jpg)
5
• Feature-based approaches
• Manual features + SVM (Jiang et al., 2011)
• Neural network-based approaches
• Embedding oriented features (Vo and Zhang, 2015)
• Recursive neural networks (Dong et al., 2014)
• End-to-end approaches
• Mainly based on recurrent neural network and attention mechanism
• TD-LSTM (Tang et al., 2016a)
• MemNet (Tang et al., 2016b)
• TNet (Li et al., 2018)
• etc.
Related Work
![Page 7: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/7.jpg)
6
• Observation 1:
The attention mechanism commonly used in previous ABSC models may
result in a given aspect mistakenly attending to syntactically unrelated
context words as descriptors
“Its size is ideal but the weight is unacceptable.”
√
×
Limitations of the SoTA
![Page 8: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/8.jpg)
7
• Observation 2:
These models are inadequate to determine sentiments depicted by multiple
words with long-range dependencies
“The staff should be a bit more friendly.”
Long-range dependency
Limitations of the SoTA
![Page 9: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/9.jpg)
8
Sole attention mechanism (at semantic level) is not enough !
Limitations of the SoTA
![Page 10: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/10.jpg)
Content
9
• Problem Formulation
• Related Work
• Motivation
• Proposed Model – ASGCN
• Experiments
• Discussion
• Conclusions & Future Work
![Page 11: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/11.jpg)
10
Sole attention mechanism (at semantic level) is not enough !
How about incorporating with dependency trees (at syntax
level)?
Motivation
![Page 12: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/12.jpg)
“Its size is ideal but the weight is unacceptable.”
size
is
ideal
unacceptableFar away
Dependency Tree
Syntactically unrelated words can be ignored by the aspect terms via distance computing
11
Motivation
![Page 13: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/13.jpg)
12
“The staff should be a bit more friendly.”
staff
be
friendly
Dependency Tree
should
Long-range dependencies can be shortened
Motivation
![Page 14: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/14.jpg)
13
Motivation
• Dependency trees can
• draw long sequences of words that are syntactically relevant closer
• keep irrelevant component words far away from aspect terms
• but they are insufficient to
• capture words’ semantics
• We propose to build Graph Convolutional Networks (GCNs) over
dependency trees to
• draw syntactically relevant words a step closer to the aspect
• exploit both (latent) semantics and syntactic information
![Page 15: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/15.jpg)
14
Slides from “Structured deep models: Deep learning on graphs and beyond”, Thomas Kipf, 2018.
GNNs
![Page 16: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/16.jpg)
15
Slides from “Structured deep models: Deep learning on graphs and beyond”, Thomas Kipf, 2018.
GCNs
![Page 17: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/17.jpg)
Content
16
• Problem Formulation
• Related Work
• Motivation
• Proposed Model – ASGCN
• Experiments
• Discussion
• Conclusions & Future Work
![Page 18: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/18.jpg)
17
• ASGCN is composed of
• Word embeddings
• Bidirectional LSTM
• GCNs
• Aspect-specific masking
• Attention
• Dependency tree is regarded
• As a graph (ASGCN-DG) or
• As a tree (ASGCN-DT)… … …
… … …
… … … … … …
… … …
… … …
Aspect-Specific GCN (ASGCN) - Overview
![Page 19: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/19.jpg)
18
• With the word embeddings of a
sentence, a bidirectional LSTM is
constructed to produce hidden state
vectors
• Following Zhang et al., 2018, we make
nodes on the graph aware of context by
initializing the nodes representation
to , i.e. … … …
… … …
… … … … … …
… … …
… … …
Embeddings & BiLSTM
![Page 20: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/20.jpg)
… … …
19
• Graph convolution at 𝑙𝑙–th layer (totally
𝐿𝐿 layers)
• We also incorporate position weights,
which is commonly used in ABSC
models, with GCNs
GCNs
… … …
… … …
… … …
… … …
… … …
![Page 21: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/21.jpg)
• Aspect-specific masking is proposed to
get aspect-oriented features
…… …
20
Aspect-specific Masking
… … …
… … …
… … …
… … …
… … …
![Page 22: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/22.jpg)
21
… … …
… … …
… … …
… … …
• Attention scores are computed based
on inner product (to facilitate the
masking mechanism)
• Final representation is computed as
Attention
… … …
… … …
![Page 23: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/23.jpg)
Content
22
• Problem Formulation
• Related Work
• Motivation
• Proposed Model – ASGCN
• Experiments
• Discussion
• Conclusions & Future Work
![Page 24: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/24.jpg)
23
Datasets & Settings
• TWITTER: built by Dong et al. (2014), containing twitter posts
• LAP14, REST14, REST15, REST16: respectively from SemEval 2014
task 4, SemEval 2015 task 12 and SemEval 2016 task 5 (Pontiki et al.,
2014, 2015, 2016), consisting of data from two categories, i.e. laptop
and restaurant
• The number of GCN layers is set to 2, which is the best-performing depth
in pilot studies (more illustration later)
• More parameter setting details could be found in paper
• Accuracy and Macro-Averaged F1
![Page 25: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/25.jpg)
24
Comparison
• Baselines
• SVM (Kiritchenko et al., 2014)
• LSTM (Tang et al., 2016a)
• MemNet (Tang et al., 2016b)
• AOA (Huang et al., 2018)
• IAN (Ma et al., 2017)
• TNet-LF (Li et al., 2018) (state-of-the-art)
• Variants
• ASCNN, which replaces 2-layer GCN with 2-layer CNN in ASGCN
![Page 26: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/26.jpg)
25
Comparison
![Page 27: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/27.jpg)
26
Comparison
![Page 28: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/28.jpg)
27
Comparison
![Page 29: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/29.jpg)
28
Comparison
• Results
• ASGCN shows a competitive performance against strong baselines
• ASGCN-DG is generally better than ASGCN-DT
• ASGCN is better at capturing long-range word dependencies than ASCNN
• ASGCN performs less well on less grammatical datasets such as
• Implications
• If we consider taking dependency trees as directed graph (e.g., ASGCN-
DT), we’d better also consider the edge label information
• We need more robust dependency parsers to reduce the effect of error
propagation
![Page 30: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/30.jpg)
29
Ablation Study
• w/o pos. : without position weight
• w/o mask: without aspect-specific masking
• w/o GCN: skip GCN layers
![Page 31: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/31.jpg)
30
Ablation Study
• w/o pos. : without position weight
• w/o mask: without aspect-specific masking
• w/o GCN: skip GCN layers
![Page 32: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/32.jpg)
31
Ablation Study
• Results
• Position weight is not working for all data
• Aspect-specific masking has positive effects
• Removal of GCN brings significant drops in performance (except for
TWITTER)
• Implications
• The integration of position weights is not crucial for less
grammatical sentences
• Aspect-specific masking mechanism is important in ASGCN
![Page 33: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/33.jpg)
32
Case Study
![Page 34: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/34.jpg)
Content
33
• Problem Formulation
• Related Work
• Motivation
• Proposed Model – ASGCN
• Experiments
• Discussion
• Conclusions & Future Work
![Page 35: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/35.jpg)
34
Impact of GCN Layers
• We could see that 2 is the best choice under our settings
![Page 36: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/36.jpg)
35
Effect of Multiple Aspects
• ASGCN shows a high variance with respect to sentences with different
number of aspect terms
![Page 37: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/37.jpg)
Content
36
• Problem Formulation
• Related Work
• Motivation
• Proposed Model – ASGCN
• Experiments
• Discussion
• Conclusions & Future Work
![Page 38: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/38.jpg)
37
Conclusions & Future Work
• Conclusions
• GCNs over dependency trees bring benefit to the overall
performance.
• ASGCN is less effective for ungrammatical contents owing to error
propagation of dependency trees
• GCNs with graphs are better than those with trees
• ASGCN is less robust to multi-aspect scenarios
• Future work
• Reduce errors of dependency parsers via joint modelling
• Incorporate edge information of dependency trees
• Reduce prediction variance via judging multiple aspects’ polarities
at the same time
![Page 39: Aspect-based Sentiment Classification with Aspect-specific ... · task 4, SemEval2015 task 12 and SemEval2016 task 5 (Pontiki et al., 2014, 2015, 2016), consisting of data from two](https://reader034.vdocument.in/reader034/viewer/2022051809/6010ec89649c387c916e0d4d/html5/thumbnails/39.jpg)
The end
Thanks!
Q&A
38
https://github.com/GeneZC/ASGCN
https://arxiv.org/abs/1909.03477