The Role of Taxonomy and Ontology in Semantic Layers - Heather Hedden.pdf
Cross domainsc new
1. Background
Preliminary experiments
Modeling accuracy loss for cross-domain SC
Graph-based algorithms
Cross-domain Sentiment Classification: Resource
Selection and Algorithms
Natalia Ponomareva
Statistical Cybermetrics Research Group,
University of Wolverhampton, UK
December 17, 2011
Natalia Ponomareva Cross-domain Sentiment Classification
2. Background
Preliminary experiments
Modeling accuracy loss for cross-domain SC
Graph-based algorithms
Outline
1 Background
Introduction
State-of-the-art research
2 Preliminary experiments
In-domain study
Cross-domain experiments
3 Modeling accuracy loss for cross-domain SC
Domain similarity
Domain complexity
Model construction and validation
4 Graph-based algorithms
Comparison
Document similarity
Strategy for choosing the best parameters
Natalia Ponomareva Cross-domain Sentiment Classification
3. Background
Preliminary experiments Introduction
Modeling accuracy loss for cross-domain SC State-of-the-art research
Graph-based algorithms
What is Sentiment Classification?
Task within the research field of Sentiment Analysis.
It concerns classification of documents on the basis of overall
sentiments expressed by their authors.
Different scales can be used:
positive/negative;
positive, negative and neutral;
rating: 1*, 2*, 3*, 4*, 5*;
Example
“The film was fun and I enjoyed it.” ⇒ positive
“The film lasted too long and I got bored.” ⇒ negative
Natalia Ponomareva Cross-domain Sentiment Classification
4. Background
Preliminary experiments Introduction
Modeling accuracy loss for cross-domain SC State-of-the-art research
Graph-based algorithms
Applications:Business Intelligence
Natalia Ponomareva Cross-domain Sentiment Classification
5. Background
Preliminary experiments Introduction
Modeling accuracy loss for cross-domain SC State-of-the-art research
Graph-based algorithms
Applications: Event prediction
Natalia Ponomareva Cross-domain Sentiment Classification
6. Background
Preliminary experiments Introduction
Modeling accuracy loss for cross-domain SC State-of-the-art research
Graph-based algorithms
Applications: Opinion search
Natalia Ponomareva Cross-domain Sentiment Classification
7. Background
Preliminary experiments Introduction
Modeling accuracy loss for cross-domain SC State-of-the-art research
Graph-based algorithms
Why challenging?
Irony, humour.
Example
If you are reading this because it is your darling fragrance, please
wear it at home exclusively and tape the windows shut.
Generally positive words.
Example
This film should be brilliant. It sounds like a great plot, the actors
are fisrt grade, and the supporting cast is good as well, and
Stallone is attempting to deliver a good performance.
However, it cannot hold up.
Natalia Ponomareva Cross-domain Sentiment Classification
8. Background
Preliminary experiments Introduction
Modeling accuracy loss for cross-domain SC State-of-the-art research
Graph-based algorithms
Why challenging?
Context dependency.
Example
This is a great camera.
A great amount of money was spent for promoting this camera.
One might think this is a great camera. Well think again,
because.....
Rejection or advice?
Example
Go read the book.
Natalia Ponomareva Cross-domain Sentiment Classification
9. Background
Preliminary experiments Introduction
Modeling accuracy loss for cross-domain SC State-of-the-art research
Graph-based algorithms
Approaches to Sentiment Classification
Lexical approaches
Supervised machine learning
Semi-supervised and unsupervised approaches
Cross-domain Sentiment Classification (SC)
Natalia Ponomareva Cross-domain Sentiment Classification
10. Background
Preliminary experiments Introduction
Modeling accuracy loss for cross-domain SC State-of-the-art research
Graph-based algorithms
Lexical approaches
Natalia Ponomareva Cross-domain Sentiment Classification
11. Background
Preliminary experiments Introduction
Modeling accuracy loss for cross-domain SC State-of-the-art research
Graph-based algorithms
Lexical approaches
Use of dictionaries of sentiment words with a given semantic
orientation.
Natalia Ponomareva Cross-domain Sentiment Classification
12. Background
Preliminary experiments Introduction
Modeling accuracy loss for cross-domain SC State-of-the-art research
Graph-based algorithms
Lexical approaches
Use of dictionaries of sentiment words with a given semantic
orientation.
Dictionaries are built either manually or (semi-)automatically.
Natalia Ponomareva Cross-domain Sentiment Classification
13. Background
Preliminary experiments Introduction
Modeling accuracy loss for cross-domain SC State-of-the-art research
Graph-based algorithms
Lexical approaches
Use of dictionaries of sentiment words with a given semantic
orientation.
Dictionaries are built either manually or (semi-)automatically.
A special scoring function is applied in order to calculate the
final semantic orientation of a text.
Natalia Ponomareva Cross-domain Sentiment Classification
14. Background
Preliminary experiments Introduction
Modeling accuracy loss for cross-domain SC State-of-the-art research
Graph-based algorithms
Lexical approaches
Use of dictionaries of sentiment words with a given semantic
orientation.
Dictionaries are built either manually or (semi-)automatically.
A special scoring function is applied in order to calculate the
final semantic orientation of a text.
Example
lightweight +3, good +4, ridiculous -2
Lightweight, stores a ridiculous amount of books and good battery
life.
SO1 = 3+4−2 = 1 2
3 3
SO2 = max{|3|, |4|, |−2|} · sign(max{|3|, |4|, |−2|}) = 4
Natalia Ponomareva Cross-domain Sentiment Classification
15. Background
Preliminary experiments Introduction
Modeling accuracy loss for cross-domain SC State-of-the-art research
Graph-based algorithms
Supervised Machine Learning
Natalia Ponomareva Cross-domain Sentiment Classification
16. Background
Preliminary experiments Introduction
Modeling accuracy loss for cross-domain SC State-of-the-art research
Graph-based algorithms
Supervised Machine Learning
Learn sentiment phenomena from an annotated corpus.
Natalia Ponomareva Cross-domain Sentiment Classification
17. Background
Preliminary experiments Introduction
Modeling accuracy loss for cross-domain SC State-of-the-art research
Graph-based algorithms
Supervised Machine Learning
Learn sentiment phenomena from an annotated corpus.
Different Machine Learning methods were tested (NB, SVM,
ME). In the majority of cases SVM demonstrates the best
performance.
Natalia Ponomareva Cross-domain Sentiment Classification
18. Background
Preliminary experiments Introduction
Modeling accuracy loss for cross-domain SC State-of-the-art research
Graph-based algorithms
Supervised Machine Learning
Learn sentiment phenomena from an annotated corpus.
Different Machine Learning methods were tested (NB, SVM,
ME). In the majority of cases SVM demonstrates the best
performance.
For review data ML approach performs better than lexical one
when training and test data belong to the same domain.
Natalia Ponomareva Cross-domain Sentiment Classification
19. Background
Preliminary experiments Introduction
Modeling accuracy loss for cross-domain SC State-of-the-art research
Graph-based algorithms
Supervised Machine Learning
Learn sentiment phenomena from an annotated corpus.
Different Machine Learning methods were tested (NB, SVM,
ME). In the majority of cases SVM demonstrates the best
performance.
For review data ML approach performs better than lexical one
when training and test data belong to the same domain.
But it needs substantial amount of annotated data.
Natalia Ponomareva Cross-domain Sentiment Classification
20. Background
Preliminary experiments Introduction
Modeling accuracy loss for cross-domain SC State-of-the-art research
Graph-based algorithms
Semi-supervised and unsupervised approaches
Natalia Ponomareva Cross-domain Sentiment Classification
21. Background
Preliminary experiments Introduction
Modeling accuracy loss for cross-domain SC State-of-the-art research
Graph-based algorithms
Semi-supervised and unsupervised approaches
Require small amount of annotated data or no data at all.
Natalia Ponomareva Cross-domain Sentiment Classification
22. Background
Preliminary experiments Introduction
Modeling accuracy loss for cross-domain SC State-of-the-art research
Graph-based algorithms
Semi-supervised and unsupervised approaches
Require small amount of annotated data or no data at all.
Different techniques were exploited:
Automatic extraction of sentiment words on the Web using
seed words (Turney, 2002).
Exploiting spectral clustering and active learning (Dasgupta et
al., 2009).
Applying co-training (Li et al., 2010)
Bootstrapping (Zagibalov, 2010)
Using graph-based algorithms (Goldberg et al., 2006)
Natalia Ponomareva Cross-domain Sentiment Classification
23. Background
Preliminary experiments Introduction
Modeling accuracy loss for cross-domain SC State-of-the-art research
Graph-based algorithms
Cross-domain SC
Main approaches:
Ensemble of classifiers (Read 2005, Aue and Gamon 2005);
Structural Correspondence Learning (Blitzer 2007);
Graph-based algorithms (Wu 2009).
Natalia Ponomareva Cross-domain Sentiment Classification
24. Background
Preliminary experiments Introduction
Modeling accuracy loss for cross-domain SC State-of-the-art research
Graph-based algorithms
Ensemble of classifiers
Classifiers are learned on data belonging to different source
domains.
Various methods can be used to combine classifiers:
Majority voting;
Weighted voting, where development data set is used to learn
credibility weights for each classifier.
Learning a meta-classifier on a small amount of target domain
data.
Natalia Ponomareva Cross-domain Sentiment Classification
25. Background
Preliminary experiments Introduction
Modeling accuracy loss for cross-domain SC State-of-the-art research
Graph-based algorithms
Structural Correspondence Learning
Blitzer et al., 2007:
Introduce pivot features that appear frequently in source and
target domains.
Find projections of source features the co-occur with pivots in
a target domain.
Example
The laptop is great, it is extremely fast.
The book is great, it is very engaging.
Natalia Ponomareva Cross-domain Sentiment Classification
26. Background
Preliminary experiments Introduction
Modeling accuracy loss for cross-domain SC State-of-the-art research
Graph-based algorithms
Structural Correspondence Learning
Blitzer et al., 2007:
Introduce pivot features that appear frequently in source and
target domains.
Find projections of source features the co-occur with pivots in
a target domain.
Example
The laptop is great, it is extremely fast.
The book is great, it is very engaging.
Natalia Ponomareva Cross-domain Sentiment Classification
27. Background
Preliminary experiments Introduction
Modeling accuracy loss for cross-domain SC State-of-the-art research
Graph-based algorithms
Structural Correspondence Learning
Blitzer et al., 2007:
Introduce pivot features that appear frequently in source and
target domains.
Find projections of source features the co-occur with pivots in
a target domain.
Example
The laptop is great, it is extremely fast.
The book is great, it is very engaging.
Natalia Ponomareva Cross-domain Sentiment Classification
28. Background
Preliminary experiments Introduction
Modeling accuracy loss for cross-domain SC State-of-the-art research
Graph-based algorithms
Structural Correspondence Learning
Blitzer et al., 2007:
Introduce pivot features that appear frequently in source and
target domains.
Find projections of source features the co-occur with pivots in
a target domain.
Example
The laptop is great, it is extremely fast.
The book is great, it is very engaging.
Natalia Ponomareva Cross-domain Sentiment Classification
29. Background
Preliminary experiments Introduction
Modeling accuracy loss for cross-domain SC State-of-the-art research
Graph-based algorithms
Discussion
Natalia Ponomareva Cross-domain Sentiment Classification
30. Background
Preliminary experiments Introduction
Modeling accuracy loss for cross-domain SC State-of-the-art research
Graph-based algorithms
Discussion
Machine learning methods demonstrate a
very good performance and when the size of
the data is substantial they outperform
lexical approaches.
Natalia Ponomareva Cross-domain Sentiment Classification
31. Background
Preliminary experiments Introduction
Modeling accuracy loss for cross-domain SC State-of-the-art research
Graph-based algorithms
Discussion
Machine learning methods demonstrate a
very good performance and when the size of
the data is substantial they outperform
lexical approaches.
On the other hand, there is a plethora of
annotated resources on the Web and the
possibility to re-use them would be very
beneficial.
Natalia Ponomareva Cross-domain Sentiment Classification
32. Background
Preliminary experiments Introduction
Modeling accuracy loss for cross-domain SC State-of-the-art research
Graph-based algorithms
Discussion
Machine learning methods demonstrate a
very good performance and when the size of
the data is substantial they outperform
lexical approaches.
On the other hand, there is a plethora of
annotated resources on the Web and the
possibility to re-use them would be very
beneficial.
Structural Correspondence Learning and
similar approaches are good for binary
classification but difficult to be applied for
multi-class problem.
Natalia Ponomareva Cross-domain Sentiment Classification
33. Background
Preliminary experiments Introduction
Modeling accuracy loss for cross-domain SC State-of-the-art research
Graph-based algorithms
Discussion
Machine learning methods demonstrate a
very good performance and when the size of
the data is substantial they outperform
lexical approaches.
On the other hand, there is a plethora of
annotated resources on the Web and the
possibility to re-use them would be very
beneficial.
Structural Correspondence Learning and
similar approaches are good for binary
classification but difficult to be applied for
multi-class problem.
That motivates us to exploit graph-based
cross-domain algorithms.
Natalia Ponomareva Cross-domain Sentiment Classification
34. Background
Preliminary experiments In-domain study
Modeling accuracy loss for cross-domain SC Cross-domain experiments
Graph-based algorithms
Data
Natalia Ponomareva Cross-domain Sentiment Classification
35. Background
Preliminary experiments In-domain study
Modeling accuracy loss for cross-domain SC Cross-domain experiments
Graph-based algorithms
Data
Data represent the corpus consist of Amazon product reviews
on 7 different topics: books (BO), electronics (EL),
kitchen&housewares (KI), DVDs (DV), music (MU),
health&personal care (HE) and toys&games(TO).
Natalia Ponomareva Cross-domain Sentiment Classification
36. Background
Preliminary experiments In-domain study
Modeling accuracy loss for cross-domain SC Cross-domain experiments
Graph-based algorithms
Data
Data represent the corpus consist of Amazon product reviews
on 7 different topics: books (BO), electronics (EL),
kitchen&housewares (KI), DVDs (DV), music (MU),
health&personal care (HE) and toys&games(TO).
Reviews are rated either as positive or negative.
Natalia Ponomareva Cross-domain Sentiment Classification
37. Background
Preliminary experiments In-domain study
Modeling accuracy loss for cross-domain SC Cross-domain experiments
Graph-based algorithms
Data
Data represent the corpus consist of Amazon product reviews
on 7 different topics: books (BO), electronics (EL),
kitchen&housewares (KI), DVDs (DV), music (MU),
health&personal care (HE) and toys&games(TO).
Reviews are rated either as positive or negative.
Data within each domain are balanced, they contain 1000
positive and 1000 negative reviews.
Natalia Ponomareva Cross-domain Sentiment Classification
38. Background
Preliminary experiments In-domain study
Modeling accuracy loss for cross-domain SC Cross-domain experiments
Graph-based algorithms
Data statistics
corpus num words mean words vocab size vocab size (>= 3)
BO 364k 181.8 23k 8 256
DV 397k 198.7 24k 8 632
MU 300k 150.1 19k 6 163
EL 236k 117.9 12k 4 465
KI 198k 98.9 11k 4 053
TO 206k 102.9 11k 4 018
HE 188k 93.9 11k 4 022
Natalia Ponomareva Cross-domain Sentiment Classification
39. Background
Preliminary experiments In-domain study
Modeling accuracy loss for cross-domain SC Cross-domain experiments
Graph-based algorithms
Data statistics
corpus num words mean words vocab size vocab size (>= 3)
BO 364k 181.8 23k 8 256
DV 397k 198.7 24k 8 632
MU 300k 150.1 19k 6 163
EL 236k 117.9 12k 4 465
KI 198k 98.9 11k 4 053
TO 206k 102.9 11k 4 018
HE 188k 93.9 11k 4 022
BO, DV, MU - longer reviews, richer vocabularies.
Natalia Ponomareva Cross-domain Sentiment Classification
40. Background
Preliminary experiments In-domain study
Modeling accuracy loss for cross-domain SC Cross-domain experiments
Graph-based algorithms
Feature selection
We compared several characteristics of features:
words vs. stems and lemmas;
unigrams vs. unigrams + bigrams;
binary weights vs. frequency, idf and tfidf;
features filtered by presence of verbs, adjectives, adverbs and
modal verbs vs. unfiltered features.
Natalia Ponomareva Cross-domain Sentiment Classification
41. Background
Preliminary experiments In-domain study
Modeling accuracy loss for cross-domain SC Cross-domain experiments
Graph-based algorithms
Feature selection
Natalia Ponomareva Cross-domain Sentiment Classification
42. Background
Preliminary experiments In-domain study
Modeling accuracy loss for cross-domain SC Cross-domain experiments
Graph-based algorithms
Feature selection
Filtering of features worsen the accuracy for all domains.
Natalia Ponomareva Cross-domain Sentiment Classification
43. Background
Preliminary experiments In-domain study
Modeling accuracy loss for cross-domain SC Cross-domain experiments
Graph-based algorithms
Feature selection
Filtering of features worsen the accuracy for all domains.
Unigrams + bigrams generally perform significantly much
better then unigrams alone.
Natalia Ponomareva Cross-domain Sentiment Classification
44. Background
Preliminary experiments In-domain study
Modeling accuracy loss for cross-domain SC Cross-domain experiments
Graph-based algorithms
Feature selection
Filtering of features worsen the accuracy for all domains.
Unigrams + bigrams generally perform significantly much
better then unigrams alone.
Binary, idf and delta idf weights generally give better results
than frequency, tfidf and delta tfidf weights.
Natalia Ponomareva Cross-domain Sentiment Classification
45. Background
Preliminary experiments In-domain study
Modeling accuracy loss for cross-domain SC Cross-domain experiments
Graph-based algorithms
Feature selection
domain features preference confidence interval, α = 0.01
BO word ≈ lemma ≈ stem inside
DV word ≈ lemma ≈ stem inside
MU lemma > stem > word boundary
EL word > lemma ≈ stem inside
KI word ≈ lemma > stem inside
TO word ≈ stem > lemma boundary
HE stem > lemma > word inside
Natalia Ponomareva Cross-domain Sentiment Classification
46. Background
Preliminary experiments In-domain study
Modeling accuracy loss for cross-domain SC Cross-domain experiments
Graph-based algorithms
Feature selection
domain features preference confidence interval, α = 0.01
BO word ≈ lemma ≈ stem inside
DV word ≈ lemma ≈ stem inside
MU lemma > stem > word boundary
EL word > lemma ≈ stem inside
KI word ≈ lemma > stem inside
TO word ≈ stem > lemma boundary
HE stem > lemma > word inside
Natalia Ponomareva Cross-domain Sentiment Classification
47. Background
Preliminary experiments In-domain study
Modeling accuracy loss for cross-domain SC Cross-domain experiments
Graph-based algorithms
Feature selection
domain features preference confidence interval, α = 0.01
BO word ≈ lemma ≈ stem inside
DV word ≈ lemma ≈ stem inside
MU lemma > stem > word boundary
EL word > lemma ≈ stem inside
KI word ≈ lemma > stem inside
TO word ≈ stem > lemma boundary
HE stem > lemma > word inside
Natalia Ponomareva Cross-domain Sentiment Classification
48. Background
Preliminary experiments In-domain study
Modeling accuracy loss for cross-domain SC Cross-domain experiments
Graph-based algorithms
10 most discriminative positive features
BO EL KI DV
highly recommend plenty perfect for album
concise plenty of be perfect magnificent
for anyone highly recommend favorite superb
i highly highly highly recommend debut
excellent ps NUM fiestaware wolf
my favorite please with be easy join
unique very happy easy to charlie
inspiring beat perfect love it
must read glad eliminate highly recommend
and also well as easy rare
Natalia Ponomareva Cross-domain Sentiment Classification
49. Background
Preliminary experiments In-domain study
Modeling accuracy loss for cross-domain SC Cross-domain experiments
Graph-based algorithms
10 most discriminative positive features
BO EL KI DV
highly recommend plenty perfect for album
concise plenty of be perfect magnificent
for anyone highly recommend favorite superb
i highly highly highly recommend debut
excellent ps NUM fiestaware wolf
my favorite please with be easy join
unique very happy easy to charlie
inspiring beat perfect love it
must read glad eliminate highly recommend
and also well as easy rare
Natalia Ponomareva Cross-domain Sentiment Classification
50. Background
Preliminary experiments In-domain study
Modeling accuracy loss for cross-domain SC Cross-domain experiments
Graph-based algorithms
10 most discriminative positive features
BO EL KI DV
highly recommend plenty perfect for album
concise plenty of be perfect magnificent
for anyone highly recommend favorite superb
i highly highly highly recommend debut
excellent ps NUM fiestaware wolf
my favorite please with be easy join
unique very happy easy to charlie
inspiring beat perfect love it
must read glad eliminate highly recommend
and also well as easy rare
Natalia Ponomareva Cross-domain Sentiment Classification
51. Background
Preliminary experiments In-domain study
Modeling accuracy loss for cross-domain SC Cross-domain experiments
Graph-based algorithms
10 most discriminative positive features
BO EL KI DV
highly recommend plenty perfect for album
concise plenty of be perfect magnificent
for anyone highly recommend favorite superb
i highly highly highly recommend debut
excellent ps NUM fiestaware wolf
my favorite please with be easy join
unique very happy easy to charlie
inspiring beat perfect love it
must read glad eliminate highly recommend
and also well as easy rare
Natalia Ponomareva Cross-domain Sentiment Classification
52. Background
Preliminary experiments In-domain study
Modeling accuracy loss for cross-domain SC Cross-domain experiments
Graph-based algorithms
10 most discriminative positive features
BO EL KI DV
highly recommend plenty perfect for album
concise plenty of be perfect magnificent
for anyone highly recommend favorite superb
i highly highly highly recommend debut
excellent ps NUM fiestaware wolf
my favorite please with be easy join
unique very happy easy to charlie
inspiring beat perfect love it
must read glad eliminate highly recommend
and also well as easy rare
Natalia Ponomareva Cross-domain Sentiment Classification
53. Background
Preliminary experiments In-domain study
Modeling accuracy loss for cross-domain SC Cross-domain experiments
Graph-based algorithms
10 most discriminative negative features
BO EL KI DV
poorly refund waste of your money
disappointing repair return it so bad
waste of do not buy it break ridiculous
your money waste of refund waste of
waste waste to return waste
annoying defective waste worst movie
bunch forum return pointless
boring junk very disappoint talk and
bunch of work worst pathetic
to finish worst I return horrible
Natalia Ponomareva Cross-domain Sentiment Classification
54. Background
Preliminary experiments In-domain study
Modeling accuracy loss for cross-domain SC Cross-domain experiments
Graph-based algorithms
10 most discriminative negative features
BO EL KI DV
poorly refund waste of your money
disappointing repair return it so bad
waste of do not buy it break ridiculous
your money waste of refund waste of
waste waste to return waste
annoying defective waste worst movie
bunch forum return pointless
boring junk very disappoint talk and
bunch of stop work worst pathetic
to finish worst I return horrible
Natalia Ponomareva Cross-domain Sentiment Classification
55. Background
Preliminary experiments In-domain study
Modeling accuracy loss for cross-domain SC Cross-domain experiments
Graph-based algorithms
10 most discriminative negative features
BO EL KI DV
poorly refund waste of your money
disappointing repair return it so bad
waste of do not buy it break ridiculous
your money waste of refund waste of
waste waste to return waste
annoying defective waste worst movie
bunch forum return pointless
boring junk very disappoint talk and
bunch of work worst pathetic
to finish worst I return horrible
Natalia Ponomareva Cross-domain Sentiment Classification
56. Background
Preliminary experiments In-domain study
Modeling accuracy loss for cross-domain SC Cross-domain experiments
Graph-based algorithms
Results
Natalia Ponomareva Cross-domain Sentiment Classification
57. Background
Preliminary experiments In-domain study
Modeling accuracy loss for cross-domain SC Cross-domain experiments
Graph-based algorithms
Results for cross-domain SC
Accuracy Accuracy drop
Natalia Ponomareva Cross-domain Sentiment Classification
58. Background
Domain similarity
Preliminary experiments
Domain complexity
Modeling accuracy loss for cross-domain SC
Model construction and validation
Graph-based algorithms
Motivation
Usually cross-domain algorithms do not work well for very
different source and target domains.
Combinations of classifiers from different domains in some
cases perform much worse than a single classifier trained on
the closest domain (Blitzer et al. 2007)
Finding the closest domain can help to improve the results of
cross-domain sentiment classification.
Natalia Ponomareva Cross-domain Sentiment Classification
59. Background
Domain similarity
Preliminary experiments
Domain complexity
Modeling accuracy loss for cross-domain SC
Model construction and validation
Graph-based algorithms
How to compare data sets?
Natalia Ponomareva Cross-domain Sentiment Classification
60. Background
Domain similarity
Preliminary experiments
Domain complexity
Modeling accuracy loss for cross-domain SC
Model construction and validation
Graph-based algorithms
How to compare data sets?
Machine-learning techniques are based on the assumption that
training and test data are driven from the same probability
distribution, and, therefore, they perform much better when
training and test data sets are alike.
Natalia Ponomareva Cross-domain Sentiment Classification
61. Background
Domain similarity
Preliminary experiments
Domain complexity
Modeling accuracy loss for cross-domain SC
Model construction and validation
Graph-based algorithms
How to compare data sets?
Machine-learning techniques are based on the assumption that
training and test data are driven from the same probability
distribution, and, therefore, they perform much better when
training and test data sets are alike.
The task of finding the best training data transforms into the
task of finding data whose feature distribution is similar to the
test one.
Natalia Ponomareva Cross-domain Sentiment Classification
62. Background
Domain similarity
Preliminary experiments
Domain complexity
Modeling accuracy loss for cross-domain SC
Model construction and validation
Graph-based algorithms
How to compare data sets?
Machine-learning techniques are based on the assumption that
training and test data are driven from the same probability
distribution, and, therefore, they perform much better when
training and test data sets are alike.
The task of finding the best training data transforms into the
task of finding data whose feature distribution is similar to the
test one.
We propose two characteristics to model accuracy loss:
domain similarity and domain complexity or, more precisely,
domain complexity variance.
Natalia Ponomareva Cross-domain Sentiment Classification
63. Background
Domain similarity
Preliminary experiments
Domain complexity
Modeling accuracy loss for cross-domain SC
Model construction and validation
Graph-based algorithms
How to compare data sets?
Machine-learning techniques are based on the assumption that
training and test data are driven from the same probability
distribution, and, therefore, they perform much better when
training and test data sets are alike.
The task of finding the best training data transforms into the
task of finding data whose feature distribution is similar to the
test one.
We propose two characteristics to model accuracy loss:
domain similarity and domain complexity or, more precisely,
domain complexity variance.
Domain similarity approximate similarity between distributions
for frequent features.
Natalia Ponomareva Cross-domain Sentiment Classification
64. Background
Domain similarity
Preliminary experiments
Domain complexity
Modeling accuracy loss for cross-domain SC
Model construction and validation
Graph-based algorithms
How to compare data sets?
Machine-learning techniques are based on the assumption that
training and test data are driven from the same probability
distribution, and, therefore, they perform much better when
training and test data sets are alike.
The task of finding the best training data transforms into the
task of finding data whose feature distribution is similar to the
test one.
We propose two characteristics to model accuracy loss:
domain similarity and domain complexity or, more precisely,
domain complexity variance.
Domain similarity approximate similarity between distributions
for frequent features.
Domain complexity compares tails of distributions.
Natalia Ponomareva Cross-domain Sentiment Classification
65. Background
Domain similarity
Preliminary experiments
Domain complexity
Modeling accuracy loss for cross-domain SC
Model construction and validation
Graph-based algorithms
Domain similarity
Natalia Ponomareva Cross-domain Sentiment Classification
66. Background
Domain similarity
Preliminary experiments
Domain complexity
Modeling accuracy loss for cross-domain SC
Model construction and validation
Graph-based algorithms
Domain similarity
We are not interested in all terms but rather on those bearing
sentiment.
Natalia Ponomareva Cross-domain Sentiment Classification
67. Background
Domain similarity
Preliminary experiments
Domain complexity
Modeling accuracy loss for cross-domain SC
Model construction and validation
Graph-based algorithms
Domain similarity
We are not interested in all terms but rather on those bearing
sentiment.
The study on SA suggested that adjectives, verbs and adverbs
are the main indicators of sentiment, so, we keep only
unigrams and bigrams that contain those POS as features.
Natalia Ponomareva Cross-domain Sentiment Classification
68. Background
Domain similarity
Preliminary experiments
Domain complexity
Modeling accuracy loss for cross-domain SC
Model construction and validation
Graph-based algorithms
Domain similarity
We are not interested in all terms but rather on those bearing
sentiment.
The study on SA suggested that adjectives, verbs and adverbs
are the main indicators of sentiment, so, we keep only
unigrams and bigrams that contain those POS as features.
We compare different weighting schemes: frequencies, TF-IDF
and IDF to compute corpus similarity.
Natalia Ponomareva Cross-domain Sentiment Classification
69. Background
Domain similarity
Preliminary experiments
Domain complexity
Modeling accuracy loss for cross-domain SC
Model construction and validation
Graph-based algorithms
Measures of domain similarity
Natalia Ponomareva Cross-domain Sentiment Classification
70. Background
Domain similarity
Preliminary experiments
Domain complexity
Modeling accuracy loss for cross-domain SC
Model construction and validation
Graph-based algorithms
Measures of domain similarity
χ2 taken from Corpus Linguistics where it was demonstrated
to have the best correlation with the gold standard.
Natalia Ponomareva Cross-domain Sentiment Classification
71. Background
Domain similarity
Preliminary experiments
Domain complexity
Modeling accuracy loss for cross-domain SC
Model construction and validation
Graph-based algorithms
Measures of domain similarity
χ2 taken from Corpus Linguistics where it was demonstrated
to have the best correlation with the gold standard.
Kullback-Leibler divergence (DKL ) and its symmetric analogue
Jensen-Shannon divergence (DJS ) were borrowed from
Information Theory.
Natalia Ponomareva Cross-domain Sentiment Classification
72. Background
Domain similarity
Preliminary experiments
Domain complexity
Modeling accuracy loss for cross-domain SC
Model construction and validation
Graph-based algorithms
Measures of domain similarity
χ2 taken from Corpus Linguistics where it was demonstrated
to have the best correlation with the gold standard.
Kullback-Leibler divergence (DKL ) and its symmetric analogue
Jensen-Shannon divergence (DJS ) were borrowed from
Information Theory.
Jaccard coefficient (Jaccard) and cosine similarity (cosine) are
well-known similarity measures
Natalia Ponomareva Cross-domain Sentiment Classification
73. Background
Domain similarity
Preliminary experiments
Domain complexity
Modeling accuracy loss for cross-domain SC
Model construction and validation
Graph-based algorithms
Correlation for different domain similarity measures
Table: Correlation with accuracy drop
measure R (freq) R (filtr.,freq) R (filtr.,TFIDF) R (filtr.,IDF)
cosine -0.790 -0.840 -0.836 -0.863
Jaccard -0.869 -0.879 -0.879 -0.879
χ2 0.855 0.869 0.876 0.879
DKL 0.734 0.827 0.676 0.796
DJS 0.829 0.833 0.804 0.876
Natalia Ponomareva Cross-domain Sentiment Classification
74. Background
Domain similarity
Preliminary experiments
Domain complexity
Modeling accuracy loss for cross-domain SC
Model construction and validation
Graph-based algorithms
Domain similarity: χ2
inv
The boundary between similar and distinct domains approximately
corresponds to χ2 = 1.7.
inv
Natalia Ponomareva Cross-domain Sentiment Classification
75. Background
Domain similarity
Preliminary experiments
Domain complexity
Modeling accuracy loss for cross-domain SC
Model construction and validation
Graph-based algorithms
Domain complexity
Similarity between domains is mostly controlled by frequent
words, but the shape of the corpus distribution is also
influenced by rare words representing its tail.
It was shown that richer domains with more rare words are
more complex for SC.
We also observed that the accuracy loss is higher in
cross-domain settings when source domain is more complex
than the target one.
Natalia Ponomareva Cross-domain Sentiment Classification
76. Background
Domain similarity
Preliminary experiments
Domain complexity
Modeling accuracy loss for cross-domain SC
Model construction and validation
Graph-based algorithms
Measures of domain complexity
We propose several measures to approximate domain complexity:
percentage of rare words;
word richness (proportion of vocabulary size in a corpus size);
relative entropy.
Correlation of domain complexity measures with in-domain
accuracy:
% of rare words word richness rel.entropy
-0.904 -0.846 0.793
Natalia Ponomareva Cross-domain Sentiment Classification
77. Background
Domain similarity
Preliminary experiments
Domain complexity
Modeling accuracy loss for cross-domain SC
Model construction and validation
Graph-based algorithms
Measures of domain complexity
We propose several measures to approximate domain complexity:
percentage of rare words;
word richness (proportion of vocabulary size in a corpus size);
relative entropy.
Correlation of domain complexity measures with in-domain
accuracy:
% of rare words word richness rel.entropy
-0.904 -0.846 0.793
Natalia Ponomareva Cross-domain Sentiment Classification
78. Background
Domain similarity
Preliminary experiments
Domain complexity
Modeling accuracy loss for cross-domain SC
Model construction and validation
Graph-based algorithms
Domain complexity
corpus accuracy % of rare words word richness rel.entropy
BO 0.786 64.77 0.064 9.23
DV 0.796 64.16 0.061 8.02
MU 0.774 67.16 0.063 8.98
EL 0.812 61.71 0.049 12.66
KI 0.829 61.49 0.053 14.44
TO 0.816 63.37 0.053 15.27
HE 0.808 61.83 0.056 15.82
Natalia Ponomareva Cross-domain Sentiment Classification
79. Background
Domain similarity
Preliminary experiments
Domain complexity
Modeling accuracy loss for cross-domain SC
Model construction and validation
Graph-based algorithms
Domain complexity
corpus accuracy % of rare words word richness rel.entropy
BO 0.786 64.77 0.064 9.23
DV 0.796 64.16 0.061 8.02
MU 0.774 67.16 0.063 8.98
EL 0.812 61.71 0.049 12.66
KI 0.829 61.49 0.053 14.44
TO 0.816 63.37 0.053 15.27
HE 0.808 61.83 0.056 15.82
Natalia Ponomareva Cross-domain Sentiment Classification
80. Background
Domain similarity
Preliminary experiments
Domain complexity
Modeling accuracy loss for cross-domain SC
Model construction and validation
Graph-based algorithms
Modeling accuracy loss
To model the performance drop we assume a linear dependency on
domain similarity and complexity variance and propose the
following linear regression model:
F (sij , ∆cij ) = β0 + β1 sij + β2 ∆cij , (1)
where
sij – domain similarity (or distance) between target domain i and
source domain j
∆cij = ci − cj , – difference between domain complexities.
The unknown coefficients βi are solutions of the following system
of linear equations:
β0 + β1 sij + β2 ∆cij = ∆aij , (2)
where ∆aij is the accuracy drop when adapting the classifier from
domain i to domain j.
Natalia Ponomareva Cross-domain Sentiment Classification
81. Background
Domain similarity
Preliminary experiments
Domain complexity
Modeling accuracy loss for cross-domain SC
Model construction and validation
Graph-based algorithms
Model evaluation
The evaluation of the constructed regression model includes
following steps:
Global test (or F-test) to verify statistical significance of
regression model with respect to all its predictors.
Test on individual variables (or t-test) to reveal regressors that
do not bring a significant impact into the model.
Leave-one-out-cross validation for the data set of 42 examples.
Natalia Ponomareva Cross-domain Sentiment Classification
82. Background
Domain similarity
Preliminary experiments
Domain complexity
Modeling accuracy loss for cross-domain SC
Model construction and validation
Graph-based algorithms
Global test
The null hypothesis for global test states that there is no
correlation between regressors and the response variable.
Our purpose is to demonstrate that this hypothesis must be
rejected with a high level of confidence.
In other words, we have to show that coefficient of
determination R 2 is high enough to consider its value
significantly different from zero.
R2 R F-value p-value
0.873 0.935 134.60 << 0.0001
Natalia Ponomareva Cross-domain Sentiment Classification
83. Background
Domain similarity
Preliminary experiments
Domain complexity
Modeling accuracy loss for cross-domain SC
Model construction and validation
Graph-based algorithms
Test on individual coefficients
β0 β1 β2
value -8.67 27.71 -0.55
standard error 1.08 1.77 0.11
t-value -8.00 15.67 -4.86
p-value << 0.0001 << 0.0001 << 0.0001
All coefficients are justified to be statistically significant with
the confidence level higher than 99.9%.
Natalia Ponomareva Cross-domain Sentiment Classification
84. Background
Domain similarity
Preliminary experiments
Domain complexity
Modeling accuracy loss for cross-domain SC
Model construction and validation
Graph-based algorithms
Leave-one-out cross-validation results
accuracy drop standard error standard deviation max error, 95%
all data 1.566 1.091 3.404
< 5% 1.465 1.133 3.373
> 5%, < 10% 1.646 1.173 3.622
> 10% 1.556 1.166 3.519
Natalia Ponomareva Cross-domain Sentiment Classification
85. Background
Domain similarity
Preliminary experiments
Domain complexity
Modeling accuracy loss for cross-domain SC
Model construction and validation
Graph-based algorithms
Leave-one-out cross-validation results
accuracy drop standard error standard deviation max error, 95%
all data 1.566 1.091 3.404
< 5% 1.465 1.133 3.373
> 5%, < 10% 1.646 1.173 3.622
> 10% 1.556 1.166 3.519
We are able to predict accuracy loss with standard error of 1.5%
and maximum error not exceeding 3.4%.
Natalia Ponomareva Cross-domain Sentiment Classification
86. Background
Domain similarity
Preliminary experiments
Domain complexity
Modeling accuracy loss for cross-domain SC
Model construction and validation
Graph-based algorithms
Leave-one-out cross-validation results
accuracy drop standard error standard deviation max error, 95%
all data 1.566 1.091 3.404
< 5% 1.465 1.133 3.373
> 5%, < 10% 1.646 1.173 3.622
> 10% 1.556 1.166 3.519
We are able to predict accuracy loss with standard error of 1.5%
and maximum error not exceeding 3.4%.
Lower values are being noticed for domains which are more
similar.
Natalia Ponomareva Cross-domain Sentiment Classification
87. Background
Domain similarity
Preliminary experiments
Domain complexity
Modeling accuracy loss for cross-domain SC
Model construction and validation
Graph-based algorithms
Leave-one-out cross-validation results
accuracy drop standard error standard deviation max error, 95%
all data 1.566 1.091 3.404
< 5% 1.465 1.133 3.373
> 5%, < 10% 1.646 1.173 3.622
> 10% 1.556 1.166 3.519
We are able to predict accuracy loss with standard error of 1.5%
and maximum error not exceeding 3.4%.
Lower values are being noticed for domains which are more
similar.
This is a strength of the model as our main purpose is to identify
the closest domains.
Natalia Ponomareva Cross-domain Sentiment Classification
88. Background
Domain similarity
Preliminary experiments
Domain complexity
Modeling accuracy loss for cross-domain SC
Model construction and validation
Graph-based algorithms
Comparing actual and predicted drop
Natalia Ponomareva Cross-domain Sentiment Classification
89. Background
Domain similarity
Preliminary experiments
Domain complexity
Modeling accuracy loss for cross-domain SC
Model construction and validation
Graph-based algorithms
Comparing actual and predicted drop
Natalia Ponomareva Cross-domain Sentiment Classification
90. Background
Comparison
Preliminary experiments
Document similarity
Modeling accuracy loss for cross-domain SC
Strategy for choosing the best parameters
Graph-based algorithms
Graph-based algorithms: OPTIM
Goldberg et al., 2006:
The algorithm is based on the
assumption that the rating function is
smooth with respect to the graph.
Rating difference between the closest
nodes is minimised.
Difference between initial rating and
the final value is also minimised.
The result is a solution of an
optimisation problem.
Natalia Ponomareva Cross-domain Sentiment Classification
91. Background
Comparison
Preliminary experiments
Document similarity
Modeling accuracy loss for cross-domain SC
Strategy for choosing the best parameters
Graph-based algorithms
Graph-based algorithms: RANK
Wu et al., 2009:
On each iteration of the algorithm
sentiment scores of unlabeled
documents are updated on the basis of
the weighted sum of sentiment scores
of the nearest labeled neighbours and
the nearest unlabeled neighbours.
The process stops when convergence
is achieved.
Natalia Ponomareva Cross-domain Sentiment Classification
92. Background
Comparison
Preliminary experiments
Document similarity
Modeling accuracy loss for cross-domain SC
Strategy for choosing the best parameters
Graph-based algorithms
Comparison
OPTIM algorithm RANK algorithm
(Goldberg et al., 2006) (Wu et al., 2009)
Natalia Ponomareva Cross-domain Sentiment Classification
93. Background
Comparison
Preliminary experiments
Document similarity
Modeling accuracy loss for cross-domain SC
Strategy for choosing the best parameters
Graph-based algorithms
Comparison
Initial setting of RANK does not allow in-domain and
out-domain neighbours to be different: easy to change!
The condition of smoothness of sentiment function over the
nodes is satisfied for both algorithms.
Unlike RANK, OPTIM requires the closeness of initial
sentiment values and output ones for unlabeled nodes.
The last condition makes the OPTIM solution more stable.
Natalia Ponomareva Cross-domain Sentiment Classification
94. Background
Comparison
Preliminary experiments
Document similarity
Modeling accuracy loss for cross-domain SC
Strategy for choosing the best parameters
Graph-based algorithms
Comparison
Initial setting of RANK does not allow in-domain and
out-domain neighbours to be different: easy to change!
The condition of smoothness of sentiment function over the
nodes is satisfied for both algorithms.
Unlike RANK, OPTIM requires the closeness of initial
sentiment values and output ones for unlabeled nodes.
The last condition makes the OPTIM solution more stable.
What about the measure of similarity between graph nodes?
Natalia Ponomareva Cross-domain Sentiment Classification
95. Background
Comparison
Preliminary experiments
Document similarity
Modeling accuracy loss for cross-domain SC
Strategy for choosing the best parameters
Graph-based algorithms
Document representation
We consider 2 types of document representation:
feature-based
sentiment units-based
Natalia Ponomareva Cross-domain Sentiment Classification
96. Background
Comparison
Preliminary experiments
Document similarity
Modeling accuracy loss for cross-domain SC
Strategy for choosing the best parameters
Graph-based algorithms
Document representation
We consider 2 types of document representation:
feature-based, that involves weighted document features.
sentiment units-based
Natalia Ponomareva Cross-domain Sentiment Classification
97. Background
Comparison
Preliminary experiments
Document similarity
Modeling accuracy loss for cross-domain SC
Strategy for choosing the best parameters
Graph-based algorithms
Document representation
We consider 2 types of document representation:
feature-based, that involves weighted document features.
Features are filtered by POS: adjectives, verbs and adverbs.
Features are weighted using either tfidf or idf.
sentiment units-based
Natalia Ponomareva Cross-domain Sentiment Classification
98. Background
Comparison
Preliminary experiments
Document similarity
Modeling accuracy loss for cross-domain SC
Strategy for choosing the best parameters
Graph-based algorithms
Document representation
We consider 2 types of document representation:
feature-based, that involves weighted document features.
Features are filtered by POS: adjectives, verbs and adverbs.
Features are weighted using either tfidf or idf.
sentiment units-based, that is based upon the percentage of
positive and negative units in a document.
Natalia Ponomareva Cross-domain Sentiment Classification
99. Background
Comparison
Preliminary experiments
Document similarity
Modeling accuracy loss for cross-domain SC
Strategy for choosing the best parameters
Graph-based algorithms
Document representation
We consider 2 types of document representation:
feature-based, that involves weighted document features.
Features are filtered by POS: adjectives, verbs and adverbs.
Features are weighted using either tfidf or idf.
sentiment units-based, that is based upon the percentage of
positive and negative units in a document.
Units can be either sentences or words.
PSP states for positive sentences percentage, PWP - for
positive words percentage.
Lexical approach was exploited to calculate semantic
orientation of sentiment units with the use of SentiWordNet
and SOCAL dictionary.
SO of sentences are averaged by a number of its positive and
negative words.
Natalia Ponomareva Cross-domain Sentiment Classification
100. Background
Comparison
Preliminary experiments
Document similarity
Modeling accuracy loss for cross-domain SC
Strategy for choosing the best parameters
Graph-based algorithms
Results
Correlation between document’s ratings and document features/units:
domain idf tfidf PSP SWN PSP SOCAL PWP SWN PWP SOCAL
BO 0.387 0.377 0.034 0.206 0.067 0.252
DV 0.376 0.368 0.064 0.251 0.098 0.316
EL 0.433 0.389 0.048 0.182 0.043 0.196
KI 0.444 0.416 0.068 0.238 0.076 0.230
Natalia Ponomareva Cross-domain Sentiment Classification
101. Background
Comparison
Preliminary experiments
Document similarity
Modeling accuracy loss for cross-domain SC
Strategy for choosing the best parameters
Graph-based algorithms
Results
Correlation between document’s ratings and document features/units:
domain idf tfidf PSP SWN PSP SOCAL PWP SWN PWP SOCAL
BO 0.387 0.377 0.034 0.206 0.067 0.252
DV 0.376 0.368 0.064 0.251 0.098 0.316
EL 0.433 0.389 0.048 0.182 0.043 0.196
KI 0.444 0.416 0.068 0.238 0.076 0.230
Feature-based document representation with idf-weights better
correlates with document rating than any other representation.
Natalia Ponomareva Cross-domain Sentiment Classification
102. Background
Comparison
Preliminary experiments
Document similarity
Modeling accuracy loss for cross-domain SC
Strategy for choosing the best parameters
Graph-based algorithms
Results
Correlation between document’s ratings and document features/units:
domain idf tfidf PSP SWN PSP SOCAL PWP SWN PWP SOCAL
BO 0.387 0.377 0.034 0.206 0.067 0.252
DV 0.376 0.368 0.064 0.251 0.098 0.316
EL 0.433 0.389 0.048 0.182 0.043 0.196
KI 0.444 0.416 0.068 0.238 0.076 0.230
Feature-based document representation with idf-weights better
correlates with document rating than any other representation.
SentiWordNet does not provide good results for this task, probably
due to high level of noise which comes from its automatic
construction.
Natalia Ponomareva Cross-domain Sentiment Classification
103. Background
Comparison
Preliminary experiments
Document similarity
Modeling accuracy loss for cross-domain SC
Strategy for choosing the best parameters
Graph-based algorithms
Results
Correlation between document’s ratings and document features/units:
domain idf tfidf PSP SWN PSP SOCAL PWP SWN PWP SOCAL
BO 0.387 0.377 0.034 0.206 0.067 0.252
DV 0.376 0.368 0.064 0.251 0.098 0.316
EL 0.433 0.389 0.048 0.182 0.043 0.196
KI 0.444 0.416 0.068 0.238 0.076 0.230
Feature-based document representation with idf-weights better
correlates with document rating than any other representation.
SentiWordNet does not provide good results for this task, probably
due to high level of noise which comes from its automatic
construction.
Document similarity is calculated using cosine measure.
Natalia Ponomareva Cross-domain Sentiment Classification
104. Background
Comparison
Preliminary experiments
Document similarity
Modeling accuracy loss for cross-domain SC
Strategy for choosing the best parameters
Graph-based algorithms
Best accuracy improvement achieved by the algorithms
We tested the performance of each algorithm for several
values of their parameters.
The best accuracy improvement that was given by each
algorithm:
OPTIM RANK
Natalia Ponomareva Cross-domain Sentiment Classification
105. Background
Comparison
Preliminary experiments
Document similarity
Modeling accuracy loss for cross-domain SC
Strategy for choosing the best parameters
Graph-based algorithms
General observations
We selected and examined only those results that were inside
the confidence interval of the best accuracy for α = 0.01.
RANK: tends to depend a lot on values of its parameters and
the most unstable results are obtained when source and target
domains are different.
RANK: A great improvement is achieved when adapting the
classifier from more complex to more simple domains.
OPTIM: Stable, but results are modest.
Natalia Ponomareva Cross-domain Sentiment Classification
106. Background
Comparison
Preliminary experiments
Document similarity
Modeling accuracy loss for cross-domain SC
Strategy for choosing the best parameters
Graph-based algorithms
Analysis of RANK behaviour
Natalia Ponomareva Cross-domain Sentiment Classification
107. Background
Comparison
Preliminary experiments
Document similarity
Modeling accuracy loss for cross-domain SC
Strategy for choosing the best parameters
Graph-based algorithms
Analysis of RANK behaviour
Within clusters of similar domains the majority of good
answers have γ ≥ 0.9.
Natalia Ponomareva Cross-domain Sentiment Classification
108. Background
Comparison
Preliminary experiments
Document similarity
Modeling accuracy loss for cross-domain SC
Strategy for choosing the best parameters
Graph-based algorithms
Analysis of RANK behaviour
Within clusters of similar domains the majority of good
answers have γ ≥ 0.9.
This demonstrates that information provided by labeled data
is more valuable.
Natalia Ponomareva Cross-domain Sentiment Classification
109. Background
Comparison
Preliminary experiments
Document similarity
Modeling accuracy loss for cross-domain SC
Strategy for choosing the best parameters
Graph-based algorithms
Analysis of RANK behaviour
Within clusters of similar domains the majority of good
answers have γ ≥ 0.9.
This demonstrates that information provided by labeled data
is more valuable.
For non-similar domains, when source domain is more complex
than the target one, best results are achieved with smaller γ
close to 0.5.
Natalia Ponomareva Cross-domain Sentiment Classification
110. Background
Comparison
Preliminary experiments
Document similarity
Modeling accuracy loss for cross-domain SC
Strategy for choosing the best parameters
Graph-based algorithms
Analysis of RANK behaviour
Within clusters of similar domains the majority of good
answers have γ ≥ 0.9.
This demonstrates that information provided by labeled data
is more valuable.
For non-similar domains, when source domain is more complex
than the target one, best results are achieved with smaller γ
close to 0.5.
This means that the algorithm benefits much from unlabeled
data.
Natalia Ponomareva Cross-domain Sentiment Classification
111. Background
Comparison
Preliminary experiments
Document similarity
Modeling accuracy loss for cross-domain SC
Strategy for choosing the best parameters
Graph-based algorithms
Analysis of RANK behaviour
Natalia Ponomareva Cross-domain Sentiment Classification
112. Background
Comparison
Preliminary experiments
Document similarity
Modeling accuracy loss for cross-domain SC
Strategy for choosing the best parameters
Graph-based algorithms
Analysis of RANK behaviour
For non-similar domains, when target one is more complex
than the source one, γ tends to increase to 0.7
Natalia Ponomareva Cross-domain Sentiment Classification
113. Background
Comparison
Preliminary experiments
Document similarity
Modeling accuracy loss for cross-domain SC
Strategy for choosing the best parameters
Graph-based algorithms
Analysis of RANK behaviour
For non-similar domains, when target one is more complex
than the source one, γ tends to increase to 0.7
That gives preference to more simple labeled data.
Natalia Ponomareva Cross-domain Sentiment Classification
114. Background
Comparison
Preliminary experiments
Document similarity
Modeling accuracy loss for cross-domain SC
Strategy for choosing the best parameters
Graph-based algorithms
Analysis of RANK behaviour
For non-similar domains, when target one is more complex
than the source one, γ tends to increase to 0.7
That gives preference to more simple labeled data.
Number of labeled and unlabeled neighbours is not equal,
there is a clear tendency to prefer results with smaller number
of unlabeled and higher number of labeled examples.
Natalia Ponomareva Cross-domain Sentiment Classification
115. Background
Comparison
Preliminary experiments
Document similarity
Modeling accuracy loss for cross-domain SC
Strategy for choosing the best parameters
Graph-based algorithms
Analysis of RANK behaviour
For non-similar domains, when target one is more complex
than the source one, γ tends to increase to 0.7
That gives preference to more simple labeled data.
Number of labeled and unlabeled neighbours is not equal,
there is a clear tendency to prefer results with smaller number
of unlabeled and higher number of labeled examples.
Proportion of 50 against 150 seems to be an ideal, covering
most of the cases.
Natalia Ponomareva Cross-domain Sentiment Classification
116. Background
Comparison
Preliminary experiments
Document similarity
Modeling accuracy loss for cross-domain SC
Strategy for choosing the best parameters
Graph-based algorithms
RANK best RANK
Natalia Ponomareva Cross-domain Sentiment Classification
117. Background
Comparison
Preliminary experiments
Document similarity
Modeling accuracy loss for cross-domain SC
Strategy for choosing the best parameters
Graph-based algorithms
OPTIM best RANK
Natalia Ponomareva Cross-domain Sentiment Classification
118. Background
Comparison
Preliminary experiments
Document similarity
Modeling accuracy loss for cross-domain SC
Strategy for choosing the best parameters
Graph-based algorithms
Conclusions and future work
Our strategy seems reasonable, the RANK performance is still
higher than the OPTIM performance.
In the future we aim to apply the gradient descent method to
refine parameters values.
Natalia Ponomareva Cross-domain Sentiment Classification
119. Background
Comparison
Preliminary experiments
Document similarity
Modeling accuracy loss for cross-domain SC
Strategy for choosing the best parameters
Graph-based algorithms
Thank you for your
attention!
Natalia Ponomareva Cross-domain Sentiment Classification