Ontology-based Faceted Search Engine for Halakhic text

Meni Adler, Yoav Goldberg, Michael Elhadad

Ben Gurion University

ISCOL 2010

Motivation

� Query שבת� Full text

� Google (many documents…)

� ת"פרויקט השו (~100K documents)

� מפעל המילון ההיסטו רי (~20K documents)

� מאגר ספ רות הקודש (~10K documents)

� מכון ממ רא (~10K documents)

� Scanned (OCR full text)� אוצר החכמה (~42K books)

� מאגר ספ רים סרוקים (~1.5K books)

� Hebrew Books (~42K books)

Motivation

� בר אילן, ת"פרויקט השו

� Morphology based

� Indexing of all possible analyses for each token

� Query expansion

� Mixture of lexemes

� שבת (37689 results)

� )צמח(ֶׁשֶבת , ) מראה מקום–שם מסכת (ַׁשָּבת, ֶׁשֶבת, ָׁשַבת, ַׁשָּבת

� )תחילית(בת , )שם עצם( ַּבת

� ָׁשב

Motivation

� האקדמיה ללשון, מפעל המילון ההיסטורי

� Lexeme-based search

� Covers Hebrew literature 300 B.C. - 1400 A.C.

� 10,287 results for the specific ַׁשָּבת query

� Mixture of ‘topics’, unclassified data

� The indexing is based on the manual tagged lemma (and part-of-speech) of each token

Selected results for query ַשָּבת

How can we improve search in a specific

domain

� Morphology

� Next talk…

� Semantics

� Topic models

� Ontology

� Combination

� Model

� Resource construction

Faceted Search

� Faceted Search

� Implementation of Exploratory Search

� Unfamiliar domain, uncertain goals

� Navigation through well defined facets

� Query � list of facets � selection � list of facets

� selection … � relevant documents

� Topic Modeling

Latent Dirichlet Allocation (LDA)

� Fully unsupervised method

� Describe topics as distribution of words

� Benefits

� Can recognize topics in documents

� Can cluster documents by topic

� Recently used for text summarization

LDA – Generative Model

α

zm,n

β φk

θm

wm,n

K topics

M documents� �

��

k in [1,K]

m in [1,M]

n in [1,Nm]

Implementation

� Inference/unsupervised learning by Gibbs sampling

� Dataset� ם"לרמבמשנה תורה

� 1000 documents (chapters)

� 128 topics http://www.cs.bgu.ac.il/~adlerm/rambam/128/topics/

� How should it be evaluated?

� Can it be improved?

Selected Topics

Lexeme-based LDA

� Hebrew� Rich Morphology

� Affixation

� Inflection

� Morphological analysis helps� Named-entity recognition

� Noun phrase chunking

� Parsing

� CFG

� Dependency trees

� Text Summarization

Lexeme-based LDAα

Lm,nβ φk

θm

wm,n

K topics

M documents

R lexemes

� �

�

�

zm,n

δ

�

Ψl

��

n in [1,Nm]

m in [1,M]

k in [1,K]

l in [1,R]

Selected Topics

Top results for query שור

Can we make it better?

� Ontology� Concepts

� Relations

� Research question� Can ontology improve topic modeling?

� Ontology construction vs. documents labeling

� Combination of topic model with existing knowledge� Dataset preparation

� Ontology construction

� Design of an algorithm that learns a topic model compatible with an ontology

The Dataset

� Responsa literature

� Medieval era

� Book � answers � paragraphs

� Germany, France, Italy

� 151 books

� Spain, North Africa

� 100 books

Halakhic Ontology - Concepts

� A hard decision…

� The juridical Halakhic index (the institute for

research in Jewish law)

� 120 entries

� Subcategories

Methodology

� Given: entry, subcategory

� Define the entry as a concept

� Generalize the subcategory into a concept

� הטלת מס על מציאה � ?הטלה , מציאה, מס

� Composed concept

� חוב-פירעון, חוב, פירעון

� Define relevant relation for the 2 concepts

� Currently: ~1000 concepts, ~2000 relations

Halakhic Ontology – A sample

Halakhic Ontology - Relations

� נכסים המשועבדים לאישה נשואה

� בעלות אלמנה על נכ סי מלוג

� אישה-סוגשל-אלמנה

� Semantic

� אישה-שיעבוד-נכס

� אלמנה-שייכות- נכסי מלוג

� Syntactic

� אישה -יחס-נכס

� אלמנה-יחס- נכסי מלוג

Halakhic Ontology - Relations

� gerund direct-object noun

� gerund indirect-object noun

� noun relation noun

� noun object gerund

� noun role noun

� noun/gerund conjunction noun/gerund

� noun attribute adjective

� noun typeof noun

Ontology-based LDA

� How to combine the ontology into the model?

� Use the ontology for constructing distributions for each document (instead of Dirichlet distribution)

� Ontology relations as covariance matrix

� Multivariate normal distribution

Correlated Topic Model

Σ

zm,n

β φk

θm

wm,n

K topics

M documentsNm words in document NΣ covariance matrix

for all documents m in [1,M] do

generate parameter vector α~N(µ,Σ)

transform α into θm: θm,t=exp(αt)/sumt’exp(αt’)

sample document length Nm~Poiss(ξ)

for all words n in [1,Nm] in document m do

sample topic index zm,n~Mult(θm)

sample term for word wm,n~Mult(φzm,n)

Model: [Blei, Lafferty, 2006]

Gibbs Sampling: [Mimno Wallach, McCallum, 2008]

�

��

k in [1,K]

m in [1,M]

n in [1,Nm]

µ�

Testset

� Selected index entries

� , שטרות, פיקדון, עד ות, ערבות, נישואין, מוחז קות, גזילה

שכירות, שבועה

� Each subcategory induces a concept

intersection query

� Currently� 43 concept intersection queries

� 1848 results

Summary

� Combination of topic modeling and

knowledge on the domain� Token-based

� Lexeme-based

� Ontology-based

� Ontology Construction� Concept, Relations

� Dataset

� Testset

Future Work

� Complete the ontology

� Evaluation� Complete queries feeding

� Compare F-measure of the various topic models

� Ontology-based Topic Modeling� Richer representation

� Relation labels

� Relation frequencies

� Transitive relations

� Matching of manual topics

Acknowledgements

� Ministry of Science and Technology

� Deutsche Telekom Laboratories

תודה