Domain Cartridge: Unsupervised Framework for Shallow Domain Ontology Construction from Corpus

Domain Cartridge: UnsupervisedFramework for Shallow Domain Ontology

Construction from Corpus

Subhabrata MukherjeeJitendra Ajmera, Sachindra Joshi

Max Planck Institute for InformaticsIBM India Research Lab

CIKM 2014

October 9, 2014

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Motivation: Domain Term Discovery

Usefulness for Parsing. Consider the examples:

I “install sprint navigation”I “problem with touch screen of the mobile”

ESG Parser generates noisy or incomplete parse without thesmartphone domain knowledge that “sprint navigation, touchscreen” are multi-word concepts

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Motivation: Domain Term Discovery

Usefulness for Parsing. Consider the examples:

I “install sprint navigation”I “problem with touch screen of the mobile”

ESG Parser generates noisy or incomplete parse without thesmartphone domain knowledge that “sprint navigation, touchscreen” are multi-word concepts

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Motivation: Domain Relation Discovery

I Interactive dialogue systemsI For user query “battery of my device depletes fast", the

knowledge ‘battery’ is a Feature-Of ‘device’ enables the systemto clarify about the type of device

I Query expansionI E.g. Consider Synonyms along with the original query, ‘battery’

is a Feature-Of ‘phone’ as well as ‘tablet’ ‘device’

I Query re-formulationI For user query “screen freezes E5150", the knowledge ‘E5150’

is a Type-Of ‘Error’ results in query re-formulation “screenfreezes error E5150"

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Motivation: Domain Relation Discovery

I Interactive dialogue systemsI For user query “battery of my device depletes fast", the

knowledge ‘battery’ is a Feature-Of ‘device’ enables the systemto clarify about the type of device

I Query expansionI E.g. Consider Synonyms along with the original query, ‘battery’

is a Feature-Of ‘phone’ as well as ‘tablet’ ‘device’

I Query re-formulationI For user query “screen freezes E5150", the knowledge ‘E5150’

is a Type-Of ‘Error’ results in query re-formulation “screenfreezes error E5150"

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Motivation: Domain Relation Discovery

I Interactive dialogue systemsI For user query “battery of my device depletes fast", the

knowledge ‘battery’ is a Feature-Of ‘device’ enables the systemto clarify about the type of device

I Query expansionI E.g. Consider Synonyms along with the original query, ‘battery’

is a Feature-Of ‘phone’ as well as ‘tablet’ ‘device’

I Query re-formulationI For user query “screen freezes E5150", the knowledge ‘E5150’

is a Type-Of ‘Error’ results in query re-formulation “screenfreezes error E5150"

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Unsupervised Framework

I Typically for a domain, a lot of knowledge articles, manuals,tutorials etc. are available in a variety of formats

I Most of these documents have less hyperlink and table(info-box as in Wikipedia) information

I Challenge is to learn a shallow ontology from raw unannotatedplain text

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Unsupervised Framework

I Typically for a domain, a lot of knowledge articles, manuals,tutorials etc. are available in a variety of formats

I Most of these documents have less hyperlink and table(info-box as in Wikipedia) information

I Challenge is to learn a shallow ontology from raw unannotatedplain text

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Unsupervised Framework

I Typically for a domain, a lot of knowledge articles, manuals,tutorials etc. are available in a variety of formats

I Most of these documents have less hyperlink and table(info-box as in Wikipedia) information

I Challenge is to learn a shallow ontology from raw unannotatedplain text

Domain Cartridge as a Graph

install insert

device

handset

android

blackberry

Operatingsystem

samsung

sim

card

SamsungGalaxyvictory

Samsungarray

Domainterm

Domainprocess

Domain Cartridge as a Graph

install insert

device

handset

android

blackberry

Operatingsystem

samsung

sim

card

SamsungGalaxyvictory

Samsungarray

Domainterm

Domainprocess

Synonyms

Domain Cartridge as a Graph

install insert

device

handset

android

blackberry

Operatingsystem

samsung

sim

card

SamsungGalaxyvictory

Samsungarray

Domainterm

Domainprocess

Feature-Of

Domain Cartridge as a Graph

install insert

device

handset

android

blackberry

Operatingsystem

samsung

sim

card

SamsungGalaxyvictory

Samsungarray

Domainterm

Domainprocess

Type-Of

Domain Cartridge as a Graph

install insert

device

handset

android

blackberry

Operatingsystem

samsung

sim

card

SamsungGalaxyvictory

Samsungarray

Domainterm

Domainprocess

Action-On

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Roadmap

I Unsupervised framework for shallow domain ontologyconstruction:

I Domain Term Discovery (DTD)I Improvement of Parser performance by DTDI Domain Relation Discovery (DRD)

I Use-Case: Improvement of an in-house Question-Answeringsystem

I Experiments: Manual Evaluation, Comparison with BabelNet,WordNet, Yago

I Conclusions

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Roadmap

I Unsupervised framework for shallow domain ontologyconstruction:

I Domain Term Discovery (DTD)I Improvement of Parser performance by DTDI Domain Relation Discovery (DRD)

I Use-Case: Improvement of an in-house Question-Answeringsystem

I Experiments: Manual Evaluation, Comparison with BabelNet,WordNet, Yago

I Conclusions

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Roadmap

I Unsupervised framework for shallow domain ontologyconstruction:

I Domain Term Discovery (DTD)I Improvement of Parser performance by DTDI Domain Relation Discovery (DRD)

I Use-Case: Improvement of an in-house Question-Answeringsystem

I Experiments: Manual Evaluation, Comparison with BabelNet,WordNet, Yago

I Conclusions

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Roadmap

I Unsupervised framework for shallow domain ontologyconstruction:

I Domain Term Discovery (DTD)I Improvement of Parser performance by DTDI Domain Relation Discovery (DRD)

I Use-Case: Improvement of an in-house Question-Answeringsystem

I Experiments: Manual Evaluation, Comparison with BabelNet,WordNet, Yago

I Conclusions

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Corpus: Knowledge articles, manuals, tutorials etc.

Domain Cartridge: Framework

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Parsing

Domain Cartridge: Framework

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Parsing

English Slot Grammar (ESG) parser is used

50 - 100 times faster than the Charniak parser

Shallow semantic relationship (SSR) annotation done over ESGparser output to generate normalized parser relations

E.g., “Samsung has a battery" and “Samsung’s battery died" bothgenerate the same relation ‘nnMod:samsung_battery’

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Domain Cartridge: Framework

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Lucene Index

For efficient retrieval of relations, documents, positional (offset)information, handling proximity based queries or answer scorers

E.g: (Doc. Freq. of) All relations containing “install” or“rel:svo:quickbook_update_file”

TokenStream modified to produce contiguous representation ofall types of concepts (ngrams, relations, words etc.)

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Domain Cartridge: Framework

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Domain Term DiscoveryESG parser maintains a domain term lexicon of multi-wordconcepts. E.g. “touch screen, sprint navigation”

Noun Phrase Chunking done on document titles to extractfrequently occuring concepts as domain wordsI Enrich lexicon and bootstrap parserI Parser generates refined output

High precision but low recall — as titles are precise, clean butshort

To extract more fine-grained domain terms HITS is used onparser output

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Domain Term DiscoveryESG parser maintains a domain term lexicon of multi-wordconcepts. E.g. “touch screen, sprint navigation”

Noun Phrase Chunking done on document titles to extractfrequently occuring concepts as domain wordsI Enrich lexicon and bootstrap parserI Parser generates refined output

High precision but low recall — as titles are precise, clean butshort

To extract more fine-grained domain terms HITS is used onparser output

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Domain Term DiscoveryESG parser maintains a domain term lexicon of multi-wordconcepts. E.g. “touch screen, sprint navigation”

Noun Phrase Chunking done on document titles to extractfrequently occuring concepts as domain wordsI Enrich lexicon and bootstrap parserI Parser generates refined output

High precision but low recall — as titles are precise, clean butshort

To extract more fine-grained domain terms HITS is used onparser output

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

HITS

I Any Shallow Semantic Relation (SSR) from ESG parser can bevisualized as a hub generating domain terms

I Any domain term can be visualized as an authority influencedby incoming features from hubs

I Link strength between the hub and authority is taken as thenumber of mutual participating SSR in the Lucene Index

I Good authorities are incorporated in Parser Domain TermLexicon

I Parser is re-run, refined relations generated and previous stepsiterated till convergence

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

HITS

I Any Shallow Semantic Relation (SSR) from ESG parser can bevisualized as a hub generating domain terms

I Any domain term can be visualized as an authority influencedby incoming features from hubs

I Link strength between the hub and authority is taken as thenumber of mutual participating SSR in the Lucene Index

I Good authorities are incorporated in Parser Domain TermLexicon

I Parser is re-run, refined relations generated and previous stepsiterated till convergence

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

HITS

I Any Shallow Semantic Relation (SSR) from ESG parser can bevisualized as a hub generating domain terms

I Any domain term can be visualized as an authority influencedby incoming features from hubs

I Link strength between the hub and authority is taken as thenumber of mutual participating SSR in the Lucene Index

I Good authorities are incorporated in Parser Domain TermLexicon

I Parser is re-run, refined relations generated and previous stepsiterated till convergence

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

HITS

I Any Shallow Semantic Relation (SSR) from ESG parser can bevisualized as a hub generating domain terms

I Any domain term can be visualized as an authority influencedby incoming features from hubs

I Link strength between the hub and authority is taken as thenumber of mutual participating SSR in the Lucene Index

I Good authorities are incorporated in Parser Domain TermLexicon

I Parser is re-run, refined relations generated and previous stepsiterated till convergence

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

HITS

I Any Shallow Semantic Relation (SSR) from ESG parser can bevisualized as a hub generating domain terms

I Any domain term can be visualized as an authority influencedby incoming features from hubs

I Link strength between the hub and authority is taken as thenumber of mutual participating SSR in the Lucene Index

I Good authorities are incorporated in Parser Domain TermLexicon

I Parser is re-run, refined relations generated and previous stepsiterated till convergence

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Feedback

Domain Cartridge: Framework

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Parser Performance Improvement

Number of incomplete parses went down by 73% afterincorporating domain terms in the parser lexicon

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Domain Terms

samsung blackberry software-version application htc-evo wi-fi memory-card bluetooth motorola kyocera browser voicemailmicrosoft-exchange lg-optimus samsung-m400 samsung-galaxy-victory software-updates samsung-array text-messaging wallpa-per synchronization iphone touch-screen blackberry-bold

Table: Snapshot of multi-word domain terms extracted by NP Chunking.

wi-fi optimus-g set-up novatel-wireless e-mail sierra-wirelessapple-id google-maps play-music mobile-network 10-digitinternet-explorer slacker-radio caller-id google-search address-book my-computer software-update blackberry-id as-well-aswindows-update terms-of-service drop-down pro-700 add-onscp-2700 mac-os device-manager voice-mail non-camera

Table: Snapshot of multi-word domain terms extracted by HITS (notfound by NP Chunking).

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Domain Cartridge: Framework

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Random Index (RI)

Random Indexing is used for computing word-relatedness anddimensionality reduction

RI considers “term X term” co-occurrence, as opposed to“term X document” matrix — allowing for incremental learning ofcontext information, scaling up with the corpus size

We adopt the notion of Relational Distributional Similarity. Twoterms are considered similar if they appear in a similar contextparticipating in similar Shallow Semantic Relations

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Random Index (RI)

Random Indexing is used for computing word-relatedness anddimensionality reduction

RI considers “term X term” co-occurrence, as opposed to“term X document” matrix — allowing for incremental learning ofcontext information, scaling up with the corpus size

We adopt the notion of Relational Distributional Similarity. Twoterms are considered similar if they appear in a similar contextparticipating in similar Shallow Semantic Relations

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Random Index

Instead of taking the raw neighborhood of N terms, we use onlythose neighboring terms that share a syntactic dependency(given by ESG parser) with the target term

Lucene Index is queried for term and relation co-occurrencestatistics while constructing high dimensional random indexvector for each term

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Domain Cartridge: Framework

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Synonym DiscoveryRandom Index is used to get top N similar terms for a given termbased on Relational Distributional Similarity

HITS is used to get the dominant domain terms and domain(SSR) relations

Sim(wi ,wj) =

∑p Ili=lj ,ki=kj (fwki

,p, fwkj,p′)∑

p∑

r Ili=lr ,ki=kr (fwki,p, fwkr ,p′)

Numerator counts the number of times any word appears in bothfeature vectors with similar dominant SSR relations

Denominator counts the number of times it appears in the featurevector of any other word with that SSR relation

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Synonym DiscoveryRandom Index is used to get top N similar terms for a given termbased on Relational Distributional Similarity

HITS is used to get the dominant domain terms and domain(SSR) relations

Sim(wi ,wj) =

∑p Ili=lj ,ki=kj (fwki

,p, fwkj,p′)∑

p∑

r Ili=lr ,ki=kr (fwki,p, fwkr ,p′)

Numerator counts the number of times any word appears in bothfeature vectors with similar dominant SSR relations

Denominator counts the number of times it appears in the featurevector of any other word with that SSR relation

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Synonym Discovery

High recall but low precision. E.g. ‘folder’ and ‘tab’ have a highrelational distributional similarity due to overlapping contextwords like “open, close, minimize, shortcut, multiple"

ESG parser attributes like noun, animated, physical object,device are used to increase precision

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Domain Cartridge: Framework

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Action-On Discovery

Action-On represents any activity or ‘method’ on a domain term

E.g. “rel:svo:tap_add_account, rel:svo:phone_access_internet,rel:svo:mobile_sync_phone, rel:svo:account_use_phone etc."

HITS index is traversed to extract all the dominant dm (action onentities), and verb-object from svo (subject-verb-object) SSR

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Action-On Discovery

Action-On represents any activity or ‘method’ on a domain term

E.g. “rel:svo:tap_add_account, rel:svo:phone_access_internet,rel:svo:mobile_sync_phone, rel:svo:account_use_phone etc."

HITS index is traversed to extract all the dominant dm (action onentities), and verb-object from svo (subject-verb-object) SSR

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Type-Of DiscoveryType-Of relations depict Is-A hierarchy i.e. a parent-child relation

E.g. “rel:svo:devices_include_HTC, rel:npo:applications_such-as_WhatsApp, rel:npo:features_like_call,rel:npo:contact_such-as_address".

svo like “include” and npo “like, such-as, as” etc. are mostinformative

Dominant domain terms from HITS, bearing Hearst patterns like“or”, “especially”

E.g. service_or_process, prev_or_next,messages_especially_sms_and_ mms

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Type-Of DiscoveryType-Of relations depict Is-A hierarchy i.e. a parent-child relation

E.g. “rel:svo:devices_include_HTC, rel:npo:applications_such-as_WhatsApp, rel:npo:features_like_call,rel:npo:contact_such-as_address".

svo like “include” and npo “like, such-as, as” etc. are mostinformative

Dominant domain terms from HITS, bearing Hearst patterns like“or”, “especially”

E.g. service_or_process, prev_or_next,messages_especially_sms_and_ mms

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Feature-Of Discovery

Feature-Of relations depict components or functionalities of adomain term

nnMod depicting noun-noun modifications, and subject-object(so) extracted from svo SSR used to discover Feature-Of fromHITS index. E.g.

“rel:nnMod:network_life, rel:nnMod:account_settings,rel:nnMod:iPhone_battery etc."

“rel:svo:motorola-photon-4g_run_device-software, rel:svo:router_decrease_signal-strength,rel:svo:find-a-store_open_locator etc."

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Feature-Of Discovery

Feature-Of relations depict components or functionalities of adomain term

nnMod depicting noun-noun modifications, and subject-object(so) extracted from svo SSR used to discover Feature-Of fromHITS index. E.g.

“rel:nnMod:network_life, rel:nnMod:account_settings,rel:nnMod:iPhone_battery etc."

“rel:svo:motorola-photon-4g_run_device-software, rel:svo:router_decrease_signal-strength,rel:svo:find-a-store_open_locator etc."

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Smartphone Ontology Snapshot

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Domain Term Evaluation

5000 articles, tutorials and manuals from the smartphone domain

We used the Back-of-the-Book Index (BOI) of manuals, to createground truth for domain term discovery

Baselines:I WordNet (G. A. Miller. Wordnet: A lexical database for english. COMMUNICATIONS OF THE ACM, 38,

1995.)

I BabelNet (R. Navigli and S. P. Ponzetto. BabelNet: Building a very large multilingual semantic network. ACL

’10.)

I Yago (F. M. Suchanek, G. Kasneci, and G. Weikum. Yago: a core of semantic knowledge. WWW ’07.)

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Domain Term Evaluation

Method RecallWordNet 22.62%NP Chunking on Titles 32.45%HITS 40.87%Yago 43.77%BabelNet 53.74%

Table: Domain term evaluation.

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Recall of a Question-Answering System

Recall@N With DomainTerm Lexicon

Without domainterm lexicon

recall@1 0.40 0.33recall@2 0.49 0.45

Table: Performance of a QA system with and without domain termlexicon.

Incorporation of domain terms in parser lexicon improves QAsystem performance

1D. Gondek et al. A framework for merging and ranking of answers inDeepQA. IBM Journal of Research and Development, 56(3), 2012.

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Domain Relation Evaluation

2000 word pairs (500 for each of four categories) are manuallyannotated by two annotators

High Kappa score of 0.92 for Synonyms and 0.70 for Type-Of dueto the large number of word pairs, 84% and 62%, the annotatorsagree are not Synonyms and Type-Of respectively

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

ComparisonI Relations in BabelNet come either from (unlabeled) Wikipedia

hyperlinks or WordNetI Yago uses features like “wikipedia:redirect” links to detect

synonymous contextsI WordNet “Hyponyms, “Hypernyms” are similar to Type-Of, and

“Meronyms, Holonyms” are subset of Feature-OfI BabelNet, WordNet, Yago do not have any category like

Action-OnI Yago has Type-Of categories derived from WordNet hyponymy

and hypernymy, and Wikipedia categories

System Type-Of Feature-Of Action-OnBabelNet, WordNet 19.27% - -Yago 25.12% - -Domain Cartridge 77% 85.7% 68%

Table: Recall comparison of systems for 3 relations.

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

ComparisonI Relations in BabelNet come either from (unlabeled) Wikipedia

hyperlinks or WordNetI Yago uses features like “wikipedia:redirect” links to detect

synonymous contextsI WordNet “Hyponyms, “Hypernyms” are similar to Type-Of, and

“Meronyms, Holonyms” are subset of Feature-OfI BabelNet, WordNet, Yago do not have any category like

Action-OnI Yago has Type-Of categories derived from WordNet hyponymy

and hypernymy, and Wikipedia categories

System Type-Of Feature-Of Action-OnBabelNet, WordNet 19.27% - -Yago 25.12% - -Domain Cartridge 77% 85.7% 68%

Table: Recall comparison of systems for 3 relations.

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

ComparisonI Relations in BabelNet come either from (unlabeled) Wikipedia

hyperlinks or WordNetI Yago uses features like “wikipedia:redirect” links to detect

synonymous contextsI WordNet “Hyponyms, “Hypernyms” are similar to Type-Of, and

“Meronyms, Holonyms” are subset of Feature-OfI BabelNet, WordNet, Yago do not have any category like

Action-OnI Yago has Type-Of categories derived from WordNet hyponymy

and hypernymy, and Wikipedia categories

System Type-Of Feature-Of Action-OnBabelNet, WordNet 19.27% - -Yago 25.12% - -Domain Cartridge 77% 85.7% 68%

Table: Recall comparison of systems for 3 relations.

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

ComparisonI Relations in BabelNet come either from (unlabeled) Wikipedia

hyperlinks or WordNetI Yago uses features like “wikipedia:redirect” links to detect

synonymous contextsI WordNet “Hyponyms, “Hypernyms” are similar to Type-Of, and

“Meronyms, Holonyms” are subset of Feature-OfI BabelNet, WordNet, Yago do not have any category like

Action-OnI Yago has Type-Of categories derived from WordNet hyponymy

and hypernymy, and Wikipedia categories

System Type-Of Feature-Of Action-OnBabelNet, WordNet 19.27% - -Yago 25.12% - -Domain Cartridge 77% 85.7% 68%

Table: Recall comparison of systems for 3 relations.

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

ComparisonI Relations in BabelNet come either from (unlabeled) Wikipedia

hyperlinks or WordNetI Yago uses features like “wikipedia:redirect” links to detect

synonymous contextsI WordNet “Hyponyms, “Hypernyms” are similar to Type-Of, and

“Meronyms, Holonyms” are subset of Feature-OfI BabelNet, WordNet, Yago do not have any category like

Action-OnI Yago has Type-Of categories derived from WordNet hyponymy

and hypernymy, and Wikipedia categories

System Type-Of Feature-Of Action-OnBabelNet, WordNet 19.27% - -Yago 25.12% - -Domain Cartridge 77% 85.7% 68%

Table: Recall comparison of systems for 3 relations.

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Distributional Similarity Comparison

System Precision Recall F-ScoreYago 38% 32% 34.37%BabelNet, WordNet 83% 31% 45.14%Domain Cartridge (DC) 58% 41% 47.60%DC + WordNet 62% 40% 49.00%DC + ESG Parser Features 65% 39% 49.14%

Table: Precision-Recall comparison of Domain Cartridge(random-indexing, HITS and sim. eqn.) with other systems.

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

Comparison with Distributional SimilarityMeasures in WordNet

WordNet F-ScoreLCH 0.22RES 0.31JCN 0.42PATH 0.42LIN 0.43WUP 0.43LESK 0.45Domain Cartridge 0.49

Table: F-Score comparison of WordNet similarity measures withDomain Cartridge.

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

ConclusionsI Unsupervised framework for shallow domain ontology

construction, without using manually annotated resources

I Multi-words form an important component of Domain TermDiscovery

I Incorporation of domain terms in parser lexicon results in 73%reduction in incomplete parses, improving performance of anin-house QA system by upto 7%

I Synonym discovery approach, using Relational DistributionalSimilarity, RI, HITS etc., performs better than other existingapproaches

I Future Work: Measure customization time required for domainadaptation of the QA system

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

ConclusionsI Unsupervised framework for shallow domain ontology

construction, without using manually annotated resources

I Multi-words form an important component of Domain TermDiscovery

I Incorporation of domain terms in parser lexicon results in 73%reduction in incomplete parses, improving performance of anin-house QA system by upto 7%

I Synonym discovery approach, using Relational DistributionalSimilarity, RI, HITS etc., performs better than other existingapproaches

I Future Work: Measure customization time required for domainadaptation of the QA system

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

ConclusionsI Unsupervised framework for shallow domain ontology

construction, without using manually annotated resources

I Multi-words form an important component of Domain TermDiscovery

I Incorporation of domain terms in parser lexicon results in 73%reduction in incomplete parses, improving performance of anin-house QA system by upto 7%

I Synonym discovery approach, using Relational DistributionalSimilarity, RI, HITS etc., performs better than other existingapproaches

I Future Work: Measure customization time required for domainadaptation of the QA system

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

ConclusionsI Unsupervised framework for shallow domain ontology

construction, without using manually annotated resources

I Multi-words form an important component of Domain TermDiscovery

I Incorporation of domain terms in parser lexicon results in 73%reduction in incomplete parses, improving performance of anin-house QA system by upto 7%

I Synonym discovery approach, using Relational DistributionalSimilarity, RI, HITS etc., performs better than other existingapproaches

I Future Work: Measure customization time required for domainadaptation of the QA system

Domain Adaptation for IE and IR Domain Term Discovery Domain Relation Discovery Experiments

ConclusionsI Unsupervised framework for shallow domain ontology

construction, without using manually annotated resources

I Multi-words form an important component of Domain TermDiscovery

I Incorporation of domain terms in parser lexicon results in 73%reduction in incomplete parses, improving performance of anin-house QA system by upto 7%

I Synonym discovery approach, using Relational DistributionalSimilarity, RI, HITS etc., performs better than other existingapproaches

I Future Work: Measure customization time required for domainadaptation of the QA system