Konstantin Vorontsov - BigARTM: Open Source Library for Regularized Multimodal Topic Modeling of Large Collections

BigARTM: Open Source Library for
Regularized Multimodal Topic Modeling
of Large Collections
Konstantin Vorontsov, Oleksandr Frei, Murat Apishev,
Peter Romov, Marina Dudarenko
Yandex • CC RAS • MIPT • HSE • MSU
Analysis of Images, Social Networks and Texts
Ekaterinburg • 9–11 April 2015

Contents
1 Theory
Probabilistic Topic Modeling
ARTM — Additive Regularization for Topic Modeling
Multimodal Probabilistic Topic Modeling
2 BigARTM implementation — http://bigartm.org
BigARTM: parallel architecture
BigARTM: time and memory performance
How to start using BigARTM
3 Experiments
ARTM for combining regularizers
Multi-ARTM for classiﬁcation
Multi-ARTM for multi-language TM

Theory
BigARTM implementation — http://bigartm.org
Experiments
What is “topic”?
Topic is a special terminology of a particular domain area.
Topic is a set of coherent terms (words or phrases)
that often occur together in documents.
Formally, topic is a probability distribution over terms:
p(w|t) is (unknown) frequency of word w in topic t.
Document semantics is a probability distribution over topics:
p(t|d) is (unknown) frequency of topic t in document d.
Each document d consists of terms w1, w2, . . . , wnd
:
p(w|d) is (known) frequency of term w in document d.
When writing term w in document d author thinks about topic t.
Topic model tries to uncover latent topics from a text collection.
Konstantin Vorontsov (voron@yandex-team.ru) BigARTM: Open Source Topic Modeling 3 / 38

Theory
Experiments
Goals and applications of Topic Modeling
Goals:
Uncover a hidden thematic structure of the text collection
Find a compressed semantic representation of each document
Applications:
Information retrieval for long-text queries
Semantic search in large scientiﬁc document collections
Revealing research trends and research fronts
Expert search
News aggregation
Recommender systems
Categorization, classiﬁcation, summarization, segmentation
of texts, images, video, signals, social media
and many others

Theory
Experiments
Probabilistic Topic Modeling: milestones and mainstream
1 PLSA — Probabilistic Latent Semantic Analysis (1999)
2 LDA — Latent Dirichlet Allocation (2003)
3 100s of PTMs based on Graphical Models & Bayesian Inference
David Blei. Probabilistic topic models // Communications of the ACM, 2012.
Vol. 55. No. 4. Pp. 77–84.

Theory
Experiments
Generative Probabilistic Topic Model (PTM)
Topic model explains terms w in documents d by topics t:
p(w|d) =
t
p(w|t)p(t|d)
Ра а а а -а а а
а . М а а а а а а а
а а а а
GC- GA- а а а а а . На
а а а , а а а а а а
а ( а , а а а ) а а а
а. М а а а а а а а а а . О
а а а а , а
а а . Е а а а
( а а а а а).
‫)݀|ݐ(݌‬
‫:)ݐ|ݓ(݌‬
‫ݐ‬ଵ, … , ‫ݐ‬௡೏
‫ݓ‬ଵ, … , ‫ݓ‬௡೏
:
0.018 а а а
0.013
0.011 а
… … … …
0.023
0.016
0.009
… … … …
0.014 а
0.009
0.006 а
… … … …
‫ݐ‬ଵ ‫ݐ‬ଶଵ ‫ݐ‬ଷଵଷ ‫ݐ‬ସଷଵ‫ݐ‬ହଷଵ

Theory
Experiments
PLSA: Probabilistic Latent Semantic Analysis [T. Hofmann 1999]
Given: D is a set (collection) of documents
W is a set (vocabulary) of terms
ndw = how many times term w appears in document d
Find: parameters φwt =p(w|t), θtd =p(t|d) of the topic model
p(w|d) =
t
φwtθtd .
The problem of log-likelihood maximization under
non-negativeness and normalization constraints:
d,w
ndw ln
t
φwtθtd → max
Φ,Θ
,
φwt 0,
w∈W
φwt = 1; θtd 0,
t∈T
θtd = 1.

Theory
Experiments
Topic Modeling is an ill-posed inverse problem
Topic Modeling is the problem of stochastic matrix factorization:
p(w|d) =
t∈T
φwtθtd .
In matrix notation P
W ×D
= Φ
W ×T
· Θ
T×D
, where
P = p(w|d) W ×D
is known term–document matrix,
Φ = φwt W ×T
is unknown term–topic matrix, φwt =p(w|t),
Θ = θtd T×D
is unknown topic–document matrix, θtd =p(t|d).
Matrix factorization is not unique, the solution is not stable:
ΦΘ = (ΦS)(S−1
Θ) = Φ′
Θ′
for all S such that Φ′ = ΦS, Θ′ = S−1Θ are stochastic.
Then, regularization is needed to ﬁnd appropriate solution.

Theory
Experiments
ARTM: Additive Regularization of Topic Model
Additional regularization criteria Ri (Φ, Θ) → max, i = 1, . . . , n.
The problem of regularized log-likelihood maximization under
non-negativeness and normalization constraints:
d,w
ndw ln
t∈T
φwtθtd
log-likelihood L (Φ,Θ)
+
n
i=1
τi Ri (Φ, Θ)
R(Φ,Θ)
→ max
Φ,Θ
,
φwt 0;
w∈W
φwt = 1; θtd 0;
t∈T
θtd = 1
where τi > 0 are regularization coeﬃcients.
Vorontsov K. V., Potapenko A. A. Tutorial on Probabilistic Topic Modeling:
Additive Regularization for Stochastic Matrix Factorization // AIST’2014,
Springer CCIS, 2014. Vol. 436. pp. 29–46.

Theory
Experiments
ARTM: available regularizers
topic smoothing (equivalent to LDA)
topic sparsing
topic decorrelation
topic selection via entropy sparsing
topic coherence maximization
supervised learning for classiﬁcation and regression
semi-supervised learning
using documents citation and links
modeling temporal topic dynamics
using vocabularies in multilingual topic models
and many others
Vorontsov K. V., Potapenko A. A. Additive Regularization of Topic Models //
Machine Learning. Special Issue “Data Analysis and Intelligent Optimization
with Applications”. Springer, 2014.

Theory
Experiments
Given a text document collection Probabilistic Topic Model ﬁnds:
p(t|d) — topic distribution for each document d,
p(w|t) — term distribution for each topic t.
Topics of documents
Words and keyphrases of topics
doc1:
doc2:
doc3:
doc4:
...
Text documents
Topic
Modeling
D
o
c
u
m
e
n
t
s
T
o
p
i
c
s

Theory
Experiments
Multimodal Topic Model ﬁnds topical distribution for terms p(w|t),
authors p(a|t), time p(y|t),
Topics of documents
doc1:
doc2:
doc3:
doc4:
...
Text documents
Topic
Modeling
D
o
c
u
m
e
n
t
s
T
o
p
i
c
s
Metadata:
Authors
Data Time
Conference
Organization
URL
etc.

Theory
Experiments
authors p(a|t), time p(y|t), objects on images p(o|t),
Topics of documents
doc1:
doc2:
doc3:
doc4:
...
Text documents
Topic
Modeling
D
o
c
u
m
e
n
t
s
T
o
p
i
c
s
Metadata:
Authors
Data Time
Conference
Organization
URL
etc.
Images

Theory
Experiments
linked documents p(d′|t),
Topics of documents
doc1:
doc2:
doc3:
doc4:
...
Text documents
Topic
Modeling
D
o
c
u
m
e
n
t
s
T
o
p
i
c
s
Metadata:
Authors
Data Time
Conference
Organization
URL
etc.
Images Links

Theory
Experiments
linked documents p(d′|t), advertising banners p(b|t),
Topics of documents
doc1:
doc2:
doc3:
doc4:
...
Text documents
Topic
Modeling
D
o
c
u
m
e
n
t
s
T
o
p
i
c
s
Metadata:
Authors
Data Time
Conference
Organization
URL
etc.
Ads Images Links

Theory
Experiments
linked documents p(d′|t), advertising banners p(b|t), users p(u|t),
Topics of documents
doc1:
doc2:
doc3:
doc4:
...
Text documents
Topic
Modeling
D
o
c
u
m
e
n
t
s
T
o
p
i
c
s
Metadata:
Authors
Data Time
Conference
Organization
URL
etc.
Ads Images Links
Users

Theory
Experiments
linked documents p(d′|t), advertising banners p(b|t), users p(u|t),
and binds all these modalities into a single topic model.
Topics of documents
doc1:
doc2:
doc3:
doc4:
...
Text documents
Topic
Modeling
D
o
c
u
m
e
n
t
s
T
o
p
i
c
s
Metadata:
Authors
Data Time
Conference
Organization
URL
etc.
Ads Images Links
Users

Theory
Experiments
Multi-ARTM: combining multimodality with regularization
M is the set of modalities
W m is a vocabulary of tokens of m-th modality, m ∈ M
W = W 1 ⊔ · · · ⊔ W M is a joint vocabulary of all modalities
The problem of multimodal regularized log-likelihood
maximization under non-negativeness and normalization constraints:
m∈M
λm
d∈D w∈W m
ndw ln
t∈T
φwtθtd
modality log-likelihood Lm(Φ,Θ)
+
n
i=1
τi Ri (Φ, Θ)
R(Φ,Θ)
→ max
Φ,Θ
,
φwt 0,
w∈W m
φwt = 1, m ∈ M; θtd 0,
t∈T
θtd = 1.
where λm > 0, τi > 0 are regularization coeﬃcients.

Theory
Experiments
Multi-ARTM: multimodal regularized EM-algorithm
EM-algorithm is a simple-iteration method for a system of equations
Theorem. The local maximum (Φ, Θ) satisﬁes the following
system of equations with auxiliary variables ptdw = p(t|d, w):
ptdw = norm
t∈T
φwtθtd ;
φwt = norm
w∈W m
nwt + φwt
∂R
∂φwt
; nwt =
d∈D
λm(w)ndw ptdw ;
θtd = norm
t∈T
ntd + θtd
∂R
∂θtd
; ntd =
w∈d
λm(w)ndw ptdw ;
where norm
t∈T
xt = max{xt ,0}
s∈T
max{xs ,0} is nonnegative normalization;
m(w) is the modality of the term w, so that w ∈ W m(w).

Theory
Experiments
Fast online EM-algorithm for Multi-ARTM
Input: collection D split into batches Db, b = 1, . . . , B;
Output: matrix Φ;
1 initialize φwt for all w ∈ W , t ∈ T;
2 nwt := 0, ñwt := 0 for all w ∈ W , t ∈ T;
3 for all batches Db, b = 1, . . . , B
4 iterate each document d ∈ Db at a constant matrix Φ:
(ñwt) := (ñwt) + ProcessBatch (Db, Φ);
5 if (synchronize) then
6 nwt := nwt + ñdw for all w ∈ W , t ∈ T;
7 φwt := norm
w∈W m
nwt + φwt
∂R
∂φwt
for all w ∈W m, m∈M, t ∈T;
8 ñwt := 0 for all w ∈ W , t ∈ T;

Theory
Experiments
Fast online EM-algorithm for Multi-ARTM
ProcessBatch iterates documents d ∈ Db at a constant matrix Φ.
matrix (ñwt) := ProcessBatch (set of documents Db, matrix Φ)
1 ñwt := 0 for all w ∈ W , t ∈ T;
2 for all d ∈ Db
3 initialize θtd := 1
|T| for all t ∈ T;
4 repeat
5 ptdw := norm
t∈T
φwtθtd for all w ∈ d, t ∈ T;
6 ntd :=
w∈d
λm(w)ndw ptdw for all t ∈ T;
7 θtd := norm
t∈T
ntd + θtd
∂R
∂θtd
for all t ∈ T;
8 until θd converges;
9 ñwt := ñwt + λm(w)ndw ptdw for all w ∈ d, t ∈ T;

Theory
Experiments
ARTM approach: benefits and restrictions
Benefits
Single EM-algorithm for many models and their combinations
PLSA, LDA, and 100s of PTMs are covered by ARTM
No complicated inference and graphical models
ARTM reduces barriers to entry into PTM research field
ARTM encourages any combinations of regularizers
Multi-ARTM encourages any combinations of modalities
Multi-ARTM is implemented in BigARTM open-source project
Under development (not really restrictions):
3-matrix factorization P = ΦΨΘ, e.g. Author-Topic Model
Further generalization of hypergraph-based Multi-ARTM
Adaptive optimization of regularization coefficients

Theory
Experiments
The BigARTM project: main features
Parallel online Multi-ARTM framework
Open-source http://bigartm.org
Distributed storage of collection is possible
Built-in regularizers:
smoothing, sparsing, decorrelation, semi-supervised learning,
and many others coming soon
Built-in quality measures:
perplexity, sparsity, kernel contrast and purity,
and many others coming soon
Many types of PTMs can be implemented via Multi-ARTM:
multilanguage, temporal, hierarchical, multigram,
and many others

Theory
Experiments
The BigARTM project: parallel architecture
Concurrent processing of batches
Simple single-threaded code for ProcessBatch
User controls when to update the model in online algorithm
Deterministic (reproducible) results from run to run

Theory
Experiments
BigARTM vs Gensim vs Vowpal Wabbit
3.7M articles from Wikipedia, 100K unique words
procs train inference perplexity
BigARTM 1 35 min 72 sec 4000
Gensim.LdaModel 1 369 min 395 sec 4161
VowpalWabbit.LDA 1 73 min 120 sec 4108
BigARTM 4 9 min 20 sec 4061
Gensim.LdaMulticore 4 60 min 222 sec 4111
BigARTM 8 4.5 min 14 sec 4304
Gensim.LdaMulticore 8 57 min 224 sec 4455
procs = number of parallel threads
inference = time to infer θd for 100K held-out documents
perplexity is calculated on held-out documents.

Theory
Experiments
Running BigARTM in Parallel
3.7M articles from Wikipedia, 100K unique words
Amazon EC2 c3.8xlarge (16 physical cores + hyperthreading)
No extra memory cost for adding more threads

Theory
Experiments
1 Download links, tutorials, documentation:
http://bigartm.org
2 Linux: compile and start examples
Windows: start examples

Theory
Experiments
1 Download links, tutorials, documentation:
2 Linux: compile and start examples
Windows: start examples
BigARTM community:
1 Post questions in BigARTM discussion group:
https://groups.google.com/group/bigartm-users
2 Report bugs in BigARTM issue tracker:
https://github.com/bigartm/bigartm/issues
3 Contribute to BigARTM project via pull requests:
https://github.com/bigartm/bigartm/pulls

Theory
Experiments
License and programming environment
Freely available for commercial usage (BSD 3-Clause license)
Cross-platform — Windows, Linux, Mac OS X (32 bit, 64 bit)
Simple command-line API — available now
Rich programming API in C++ and Python — available now
Rich programming API in C# and Java — coming soon

Theory
Experiments
Combining Regularizers: experiment on 3.7M Wikipedia collection
Additive combination of 5 regularizers:
smoothing background (common lexis) topics B in Φ and Θ
sparsing domain-speciﬁc topics S = TB in Φ and Θ
decorrelation of topics in Φ
ΦW ×T ΘT×D

Theory
Experiments
Combining Regularizers: experiment on 3.7M Wikipedia collection
Additive combination of 5 regularizers:
smoothing background (common lexis) topics B in Φ and Θ
sparsing domain-speciﬁc topics S = TB in Φ and Θ
decorrelation of topics in Φ
R(Φ, Θ) = + β1
t∈B w∈W
βw ln φwt + α1
d∈D t∈B
αt ln θtd
− β0
t∈S w∈W
βw ln φwt − α0
d∈D t∈S
αt ln θtd
− γ
t∈T s∈Tt w∈W
φwtφws
where β0, α0, β1, α1, γ are regularization coeﬃcients.

Theory
Experiments
Combining Regularizers: LDA vs ARTM models
P10k , P100k — hold-out perplexity (10K, 100K documents)
SΦ, SΘ — sparsity of Φ and Θ matrices (in %)
Ks , Kp, Kc — average topic kernel size, purity and contrast
Model P10k P100k SΦ SΘ Ks Kp Kc
LDA 3436 3801 0.0 0.0 873 0.533 0.507
ARTM 3577 3947 96.3 80.9 1079 0.785 0.731
Convergence of LDA (thin lines) and ARTM (bold lines)
1 · 106
2 · 106
3 · 106
0.34
0.52
0.69
0.87
1.04
1.22
·104
Perplexity
0
20
40
60
80
100
Sparsity
Perplexity Phi Theta
1 · 106
2 · 106
3 · 106
0
0.25
0.5
0.75
1
1.25
·103
Kernelsize
0
0.2
0.4
0.6
0.8
1
Purityandcontrast
Size Purity Contrast

Theory
Experiments
EUR-Lex corpus
19 800 documents about European Union law
Two modalities: 21K words, 3 250 categories (class labels)
EUR-Lex is a “power-law dataset” with unbalanced classes:
Left: # unique labels with a given # documents per label
Right: # documents with a given # labels
Rubin T. N., Chambers A., Smyth P., Steyvers M. Statistical topic models for
multi-label document classiﬁcation // Machine Learning, 2012, 88(1-2).
Pp. 157–208.

Theory
Experiments
Regularizers:
Uniform smoothing for Θ
Uniform smoothing for word–topic matrix Φ1
Label regularization for class–topic matrix Φ2:
R(Φ2
) = τ
c∈W 2
ˆpc ln p(c) → max,
where
p(c) =
t∈T
φctp(t) is the model distribution of class c,
p(t) = nt
n can be easily estimated along EM iterations,
ˆpc is the empirical frequency of class c in the training data.

Theory
Experiments
The comparative study of models on EUR-Lex classification task
DLDA (Dependency LDA) [Rubin 2012] is a nearest analog
of Multi-ARTM for classification among Bayesian Topic Models
Quality measures [Rubin 2012]:
AUC-PR (%, ⇑) — Area under precision-recall curve
AUC (%, ⇑) — Area under ROC curve
OneErr (%, ⇓) — One error (most ranked label is not relevant)
IsErr (%, ⇓) — Is error (no perfect classification)
Results:
|T|opt AUC-PR AUC OneErr IsErr
Multi-ARTM 10 000 51.3 98.0 29.1 95.5
DLDA [Rubin 2012] 200 49.2 98.2 32.0 97.2
SVM 43.5 97.5 31.6 98.1

Theory
Experiments
Multi-language ARTM
We consider languages as modalities in Multi-ARTM.
Collection of 216 175 Russian–English Wikipedia articles pairs.
Top 10 words with p(w|t) probabilities (in %):
Topic 68 Topic 79
research 4.56 институт 6.03 goals 4.48 матч 6.02
technology 3.14 университет 3.35 league 3.99 игрок 5.56
engineering 2.63 программа 3.17 club 3.76 сборная 4.51
institute 2.37 учебный 2.75 season 3.49 фк 3.25
science 1.97 технический 2.70 scored 2.72 против 3.20
program 1.60 технология 2.30 cup 2.57 клуб 3.14
education 1.44 научный 1.76 goal 2.48 футболист 2.67
campus 1.43 исследование 1.67 apps 1.74 гол 2.65
management 1.38 наука 1.64 debut 1.69 забивать 2.53
programs 1.36 образование 1.47 match 1.67 команда 2.14

Theory
Experiments
Multi-language ARTM
Collection of 216 175 Russian–English Wikipedia articles pairs.
Top 10 words with p(w|t) probabilities (in %):
Topic 88 Topic 251
opera 7.36 опера 7.82 windows 8.00 windows 6.05
conductor 1.69 оперный 3.13 microsoft 4.03 microsoft 3.76
orchestra 1.14 дирижер 2.82 server 2.93 версия 1.86
wagner 0.97 певец 1.65 software 1.38 приложение 1.86
soprano 0.78 певица 1.51 user 1.03 сервер 1.63
performance 0.78 театр 1.14 security 0.92 server 1.54
mozart 0.74 партия 1.05 mitchell 0.82 программный 1.08
sang 0.70 сопрано 0.97 oracle 0.82 пользователь 1.04
singing 0.69 вагнер 0.90 enterprise 0.78 обеспечение 1.02
operas 0.68 оркестр 0.82 users 0.78 система 0.96
All |T| = 400 topics were reviewed by an independent assessor,
and he successfully interpreted 396 topics.

Conclusions
ARTM (Additive Regularization for Topic Modeling)
is a general framework, which makes topic models
easy to design, to infer, to explain, and to combine.
Multi-ARTM is a further generalization of ARTM
for multimodal topic modeling
BigARTM is an open source project for parallel online
topic modeling of large text collections.
Join BigARTM community!

Konstantin Vorontsov - BigARTM: Open Source Library for Regularized Multimodal Topic Modeling of Large Collections

More Related Content

What's hot (20)

Viewers also liked (6)

Similar to Konstantin Vorontsov - BigARTM: Open Source Library for Regularized Multimodal Topic Modeling of Large Collections (20)

More from AIST (20)

Recently uploaded (20)

Konstantin Vorontsov - BigARTM: Open Source Library for Regularized Multimodal Topic Modeling of Large Collections