Kernel Canonical
Correlation Analysis
Cross-language information
retrieval
Blaz Fortuna
JSI, Slovenija
Input
Two different views of
the same data:
 Text documents
written in different
languages
 Images with
attached text
…
Goal
Find pairs of features
from both views with
highest correlations
Example: words that
co-appear in document
and its translation
car, vehicle, …
Auto,
Fahrzeug, …
meat, chicken,
beef, pork, …
Fleisch,
Hahnchen,
Rindfleisch,
Schweinerne, …
Theory behind CCA


Documents are presented with pairs of
vectors – one for each view
Result of CCA are basis vectors for each
view such that the correlation between
the projections of the variables onto
these basis vectors are mutually
maximized
Kernelisation of CCA



Method can be rewritten so feature
vectors only appear inside inner-product
We can use Kernel for calculating innerproduct
Input documents don not need to be
vectors (eg. text documents together
with string kernel)
Cross-Language Text Mining



KCCA constructs language independent
representation for text documents
Good part: documents from different
languages can be compared using this
representation
Bad part: paired dataset is needed for
training (can be avoided using machine
translation tools)
KCCA and LSI

LSI discovers statistically most significant cooccurrences of terms in documents


When word appears in a document, what other
words usually also appear?
KCCA matches terms from the first language
with terms from the second based on
co-occurrences

When word appears in a document, does it also
appear in its translation?
Text document retrieval



Query databases with multilingual
documents
Documents from database and query
are transformed into language
independent representation
Nearest neighbour
Experiments




36th Canadian Parliament proceedings corpus
Part of documents used for training
For testing 5 most relevant keywords were extracted
from a document and used as queries
English query, French documents
100
200
300
400
500
LSI
30/67
38/75
42/79
45/81
49/84
KCCA
68/94
75/96
78/97
79/98
81/98
retrieval accuracy (top-ranked/top-ten-ranked) [%]
Text categorization



Categorize multilingual documents
All documents are transformed into
language independent representation
Classifier is trained on transformed
labelled documents
Experiments




NTCIR-3 patent retrieval test collection
Japanese – English
SVM trained on English documents
Tested both on the Japanese and English
Eng-train
Eng-test
Jp-train
Jp-test
50
100
150
Full
87.6
85.1
87.4
77.2
93.9
87.4
92.9
77.7
95.8
87.0
95.4
77.3
97.1
87.9
96.8
78.4
Average precision [%]
Image-Text Retrieval



Retrieval of images based on a text
query
No labels associated with images
Paired dataset:


Image retrieved from internet
Text on web page where image appeared
Experiments





Querying database with images with text queries
Images were split into three clusters
10 or 30 images that best match query are retrieved
In first test success is when images are of same label
In second test success is when images that actually
matched query is retrieved
10
30
10
30
30 dim
85%
91%
17%
60%
150 dim
83%
91%
32%
69%
Images retrieved for the text query:
”height: 6-11 weight: 235 lbs position: forward born:
september 18, 1968, split, croatia college: none”
”at phoenix sky harbor on july 6, 1997. 757-2s7, n907wa phoenix
suns taxis past n902aw teamwork america west america west 7572s7, n907wa phoenix suns taxis past n901aw arizona at phoenix
sky harbor on july 6, 1997.”
Feature work



Use of machine translation for making
paired dataset
Experiments with SVEZ-IJS EnglishSlovene ACQUIS Corpus
Sparse version of KCCA
Descargar

Kernel Canonical Correlation analysis