A Domain Based Approach to Information Retrieval in Digital Libraries - Rotella, Ferilli, Leuzzi

Embed Size (px)

Citation preview

Universit degli studi di Bari Aldo Moro
Dipartimento di Informatica

A Domain Based Approach to Information Retrieval in Digital LibrariesF. Rotella, S. Ferilli, F. [email protected], {fabio.leuzzi, rotella.fulvio}@gmail.com

8th Italian Research Conference on Digital LibrariesBari, Italy, February 9-10, 2012

L.A.C.A.M. http://lacam.di.uniba.it:8000

Overview

Introduction & Objectives

Keyword Extraction

Word Sense Disambiguation

Synset Clustering

A Multistrategy Similarity Measure

Document Partitioning

User Query Processing

A Preliminary Evaluation

Conclusions & Future Works

A Domain Based Approach to Information Retrieval in Digital Libraries - F. Rotella, S. Ferilli, F. Leuzzi

Some repositories leave the responsibility of quality to the authors.+Anybody can produce and distribute documents.=Possible low average quality of the repository contents.

Users are often overwhelmed by documents that only apparently are suitable for satisfying their information needs.

Introduction

A Domain Based Approach to Information Retrieval in Digital Libraries - F. Rotella, S. Ferilli, F. Leuzzi

Introduction

Possible way out: Information Retrieval systems

Numerical/statistical manipulation of (key)words has been widely explored in the literatureStill unable to fully solve the problem

Achieving better retrieval performance requires to go beyond simple lexical interpretation of the user queriesPass through an understanding of their semantic content and aims

Ontological taxonomyWordNet

WordNet Domains

A Domain Based Approach to Information Retrieval in Digital Libraries - F. Rotella, S. Ferilli, F. Leuzzi

Objectives

Improving fruition of a DL

Use of advanced techniques for document retrieval

Try to overcome the ambiguity of natural language

Inspired by the typical behavior of humans:take into account the possible meanings of words

select the most appropriate one according to the context of the discourse

A Domain Based Approach to Information Retrieval in Digital Libraries - F. Rotella, S. Ferilli, F. Leuzzi

Keyword Extraction

Each document in the digital library is progressively split into paragraphs, sentences, and single wordsIntegrated in the DOMINUS framework

Obtained the syntactic structure of sentences, and the lemmasIntegrated in the Stanford Parser

Classical VSM TF*IDF weighting

Two filters:Only nouns consideredThe representation of adverbs, verbs and adjectives in WordNet is different

Only the top 10% keywords for each documentTo be noise-tolerant

To limit the possibility of including non-discriminative and very general words in the representation of a document

A Domain Based Approach to Information Retrieval in Digital Libraries - F. Rotella, S. Ferilli, F. Leuzzi

Word Sense Disambiguation
Domain Driven

One Domain per Discourse assumption: many uses of a word in a coherent portion of text tend to share the same domain.

Prevalent domain individuationExtraction of allsynsets for each termExtraction of alldomains for each synsetChoice of prevalent domain synset

A Domain Based Approach to Information Retrieval in Digital Libraries - F. Rotella, S. Ferilli, F. Leuzzi

Synset Clustering

Pairwise complete link agglomerative strategy

A Domain Based Approach to Information Retrieval in Digital Libraries - F. Rotella, S. Ferilli, F. Leuzzi

Each synset generates a singleton cluster

For each pair of clustersIf the complete link property holdsMerge the involved clusters

A Multistrategy
Similarity Measure

Cooperating Techniques for Extracting Conceptual Taxonomies from Text - S. Ferilli, F. Leuzzi, F. Rotella

3 components are summed and normalized, in ]0,1[depth (ancestors)

breadth (direct neighbors)

breadth (inverse neighbors)

WordNet relationship are considered

A Multistrategy Similarity Measure
Cosidered Relationship

member meronimy: the latter synset is a member meronym of the former;substance meronimy: the latter synset is a substance meronym of the former;part meronimy: the latter synset is a part meronym of the former;similarity: the latter synset is similar in meaning to the former;antonym: specifies antonymous word;attribute: defines the attribute relation between noun and adjective synsetpairs in which the adjective is a value of the noun;additional information: additional information about the first word can beobtained by seeing the second word;part of speech based: specifies two different relations based on the parts ofspeech involved;participle: the adjective first word is a participle of the verb second word;hyperonymy: the latter synset is a hypernym of the former.

A Domain Based Approach to Information Retrieval in Digital Libraries - F. Rotella, S. Ferilli, F. Leuzzi

Document Partitioning

SynsetWord structure:Original word

TF*IDF weight

Synset

The Pairwise Clustering step returned a set of synset clusters

For each document in the collectionEach of its SynsetWord votes with its TF*IDF weight

The first three clusters are chosen from the ranked listThey represent the intensional description of the document

A Domain Based Approach to Information Retrieval in Digital Libraries - F. Rotella, S. Ferilli, F. Leuzzi

Users Query Elaboration
Overview

Same grammatical preprocessing as in the previous phase

Query usually very shortNo keyword extraction: all nouns retained for the next operations

WSD Domain Driven unreliableFor each word, all corresponding synsets in WordNet are kept

A single lexical query yields many semantic queriesAll possible combinations of synsets

A Domain Based Approach to Information Retrieval in Digital Libraries - F. Rotella, S. Ferilli, F. Leuzzi

Users Query Elaboration
A Brute Force WSD

For each combination:

a similarity evaluated against each cluster that has at least one associated document

using the same similarity function as for clustering

Twofold objective:

finding the combination of synsets that represents the best word sense disambiguation

obtaining the most similar cluster to the involved words

A Domain Based Approach to Information Retrieval in Digital Libraries - F. Rotella, S. Ferilli, F. Leuzzi

Users Query Elaboration
Query Results

The best combination is used to obtain the list of clusters ranked by descending relevance, that can be used as an answer to the user search.

The results are then displayed to the user, in particular are displayed the first n sets of document such that n is the minimum value that shows at least 10 results.

A Domain Based Approach to Information Retrieval in Digital Libraries - F. Rotella, S. Ferilli, F. Leuzzi

A Preliminary Evaluation
The Quality of Clusters

A Domain Based Approach to Information Retrieval in Digital Libraries - F. Rotella, S. Ferilli, F. Leuzzi

86 documents, 4 topics: 27 general science and physics; 21 music; 15 politics; 23 religion.

Query: Reincarnation and eternal lifeBest combination:synset: 106191212; lemmas: reincarnation; gloss: the Hindu or Buddhist doctrine that a person may be reborn successively into one of five classes of living beings (god or human or animal or hungry ghost or denizen of Hell) depending on the persons own actions;

synset: 100006269; lemmas: life; gloss: living things collectively.

Most similar cluster:synset: 106191212; lemmas: reincarnation; gloss: the Hindu or Buddhist doctrine that a person may be reborn successively into one of five classes of living beings (god or human or animal or hungry ghost or denizen of Hell) depending on the persons own actions;

synset: 105943300; lemmas: doctrine, philosophical system, philosophy and school of thought; gloss: a belief (or system of beliefs) accepted as authoritative by some group or school;

synset: 105941423; lemmas: belief; gloss: any cognitive content held as true.

Query: Ornaments and melodiesBest combination:synset: 103169390; lemmas: decoration, ornament and ornamentation; gloss: something used to beautify;

synset: 107028373; lemmas: air, line, melodic line, melodic phrase, melody, strain and tune; gloss: a succession of notes forming a distinctive sequence.

Most similar cluster:synset: 107025900; lemmas: classical, classical music and serious music; gloss: traditional genre of music conforming to an established form and appealing to critical interest and developed musical taste;

synset: 107033753; lemmas: mass; gloss: a musical setting for a Mass;

synset: 107026352; lemmas: opera; gloss: a drama set to music, consists of singing with orchestral accompaniment and an orchestral overture and interludes;

synset: 107071942; lemmas: genre, music genre, musical genre and musical style; gloss: an expressive style of music;

synset: 107064715; lemmas: rock, rock n roll, rock and roll, rock music, rocknroll and rock-and-roll; gloss: a genre of popular music originating in the 1950s, a blend of black rhythm-and-blues with white country-and-western.

A Domain Based Approach to Information Retrieval in Digital Libraries - F. Rotella, S. Ferilli, F. Leuzzi

A Preliminary Evaluation
The Quality of Clusters

A Domain Based Approach to Information Retrieval in Digital Libraries - F. Rotella, S. Ferilli, F. Leuzzi

#QueryOutcomesPrecisionRecall

1Ornaments and melodies[1 to 9] music[10 to 11] religion0.82 (1.0)0.43 (9/21)

2Reincarnation and eternal life[1 to 9] religion[10] science0.9 (1.0)0.39 (9/23)

3Traditions and folks[1 to 4] music[5 to 6] religion[7 to 10] music0.8 (1.0)0.38 (8/21)

4Limits of theory of relativity[1 to 2] science[3] politics[4 to 5] religion[6 to 15] science0.80.44 (12/27)

5Capitalism vs communism[1 to 3] politics[4] science[5 to 6] religion[7 to 11] politics[12] science[13] music0.61 (0.77)0.53 (8/15)

6Markets and new economy[1] politics[2] music[3] science[4 to 8] politics[9 to 10] religion0.6 (0.7)0.4 (6/15)

7Relationship between democracy and parliament[1 to 3] politics[4] science[5 to 6] politics[7 to 10] religion0.5 (0.6)0.33 (5/15)

A Preliminary Evaluation
Synthesis of Outcomes

Conclusions

Proposed an approach to extract information from digital libraries Go beyond simple lexical matching, toward the semantic content underlying queries

The approach consists of:An off-line preprocessing on the entire corpus Find sets of synset as intensional descriptions for the documents

An on-line phase on the queriesFind the most suitable sense, evaluating all possible combinations of synset against each intensional descriptions of the documentsIn order to propose as result the most relevant ones

Preliminary experiments show that this approach can be viable.

A Domain Based Approach to Information Retrieval in Digital Libraries - F. Rotella, S. Ferilli, F. Leuzzi

Future Works

Substitution of the ODD assumption with a more elaborated strategy for WSD

Avoiding the pre-processing stepTo handle cases when new documents are progressively included in the collection

Including adverbs, verbs and adjectivesTo improve the quality of the semantic representatives of the documents

To explore other approaches to choose better intensional descriptions of each document

A Domain Based Approach to Information Retrieval in Digital Libraries - F. Rotella, S. Ferilli, F. Leuzzi

Muokkaa otsikon tekstimuotoa napsauttamalla

Muokkaa jsennyksen tekstimuotoa napsauttamallaToinen jsennystasoKolmas jsennystasoNeljs jsennystasoViides jsennystasoKuudes jsennystasoSeitsems jsennystasoKahdeksas jsennystasoYhdekss jsennystaso