SAND2009-2391C1/20

ParaText™Leveraging Scalable Scientific Computing Capabilities for Large-Scale Text Analysis

and Visualization

Daniel M. Dunlavy, Timothy M. Shead, Patricia J. CrossnoSandia National Laboratories

SIAM Conference on Computational Science and EngineeringMarch 2, 2009

Sandia is a multiprogram laboratory operated by Sandia Corporation, a Lockheed Martin Company,for the United States Department of Energy’s National Nuclear Security Administration

under contract DE-AC04-94AL85000.

2/20

Motivation

Unstructured text

DatabaseDatabase

Data analyst

Processing and analysis Visualization

Gigabytes to terabytes

Few andoverworked

Scalable: ParaText™ Scalable: VTK / ParaView / Titan

3/20

Text Analysis Pipeline

Ingestion

Pre-processing

Transformation

Analysis

Post-processing

Archiving

File readers (ASCII, UTF-8, XML, PDF, ...)

Tokenization, stemming, part-of-speech taggingnamed entity extraction, sentence boundaries

Data modeling, dimensionality reduction, feature weighting, feature extraction/selection

Information retrieval, clustering, summarization,classification, pattern recognition, statistics

Visualization, filtering, summary statistics

Database, file, web site

4/20

Vector Space Model

• Vector Space Model for Text– Terms (features): – Documents (objects):– Term Document Matrix:– : measure of importance of term in document

• Term Examples

– Sentence: “Danny re-sent $1.”– Words: danny, sent, re [# chars?], $ [sym?], 1 [#?], re-sent [-?]– n-grams (3): dan, ann, nny, ny_, _re, re-, e-s, sen, ent, nt_, …– Named entities (people, orgs, money, etc.): danny, $1

• Document Examples– Documents, paragraphs, sentences, fixed-size chunks

[G. Salton, A. Wong, and C. S. Yang (1975), "A Vector Space Model for Automatic Indexing," Comm. ACM.]

5/20

Latent Semantic Analysis (LSA)

• SVD:

• Boolean query (query as new “doc”):

• Truncated SVD:

• LSA query:

term

s

documentsd1 d2 dn

t2

t1

tm

…d3 d4

.

.

. Truncated SVD

term

s

concepts documents

con

cep

ts

singular values

[Deerwester, S. C., et al. (1990), “Indexing by latent semantic analysis.” J. Am. Soc. Inform. Sci.]

6/20

ParaText™: Scalable Text Analysis

• ParaText™– Scalable client-server system

• Leverages existing capabilities• ParaView, VTK, Trilinos

– Latent semantic analysis• Term-document matrix analysis• Document similarities• Term (concept) similarities

– Algebraic Engine• General object-feature analysis

• General Use– OverView: open source information visualization– Timeline Treemap Browser: temporal information visualization– LSAView: text algorithm analysis– LDRDView: funding portfolio analysis– ParaSpace: simulation data analysis

7/20

Leveraging VTK / ParaView

• Useful capabilities already in VTK / ParaView– Pipeline architecture– Components for reading, processing, and visualizing (scientific) data sets– Distributed-memory parallel processing– Streaming data– ParaView client/server architecture

• Informatics capabilities recently added by Titan– Data structures: tables, trees, graphs– Data storage: delimited text files, common graph formats, SQL databases– Algorithms: graph, statistics, Matlab and R integration– Visualization: informatics-oriented paradigms

• Recent Titan features driven by ParaText™– Dense & sparse N-way arrays for algebraic / tensor analysis– Unicode text storage, manipulation, and rendering– Text ingestion, LSA, creating/storing algebraic models, performing

algebraic analyses

8/20

Leveraging Trilinos

• Useful capabilities already in Trilinos– Sparse matrix distributed data structure (Epetra)

– SVD (Anasazi)

– Incremental SVD (RBGen)

– Load balancing (Isorropia/Zoltan)

• Remaining issues for ParaText™– Data layout: different needs throughout pipeline

– Data storage: Epetra only provides doubles (Tpetra)

– Data storage: compressed row storage

– Target platforms: Linux focus (Autotools → CMake)

9/20

ParaText™ Server (PTS)

Artifact DB

PTS PTS PTS

Reader

P0 P1 Pk

Parser

Matrix

SVD

Reader

Parser

Matrix

SVD

Reader

Parser

Matrix

SVD

Fullyscalable pipeline

Matrices DB

HPC Resource (cluster, multicore server, etc.)

Database Servers

XMLHTTP

Master ParaText™

Server

ParaText™ Client

OverView

10/20

ParaText™ Scaling Examples

• Types of scaling– Strong: fixed data size, increasing # of processors– Weak: increasing data size, increasing # of processors– Fixed resource: increasing data size, fixed # of processors

• Data– Subset of Spock Challenge Data (~10 Gb HTML documents)– 64 Mb, 3318 documents, 281904 unique terms

• Hardware– 16 Dell 1850 Servers, Dual 3.6Ghz EM64T, 6GB RAM

11/20

Strong Scaling

12/20

Weak Scaling

13/20

Weak Scaling by Filter

14/20

Fixed Resource Scaling

15/20

<DOC><DOCNO><s docid="APW19990519.0113" num="1" stype="-1">APW19990519.0113</s></DOCNO><DATE_TIME><P><s docid="APW19990519.0113" num="2" stype="-1"> 1999-05-19 21:11:17</s> </P></DATE_TIME><BODY><CATEGORY><s docid="APW19990519.0113" num="3"stype="-1"> usa</s></CATEGORY><HEADLINE><P><s docid="APW19990519.0113" num="4" stype="0"> Pulses May Ease SchizophrenicVoices</s> </P></HEADLINE><TEXT><P><s docid="APW19990519.0113" num="5" stype="1"> WASHINGTON (AP)Schizophrenia patients whose medication couldn't stop the imaginary voices in theirheads gained some relief after researchers repeatedly sent a magnetic field into asmall area of their brains.</s> </P><P><s docid="APW19990519.0113" num="6"stype="1"> About half of 12 patients studied said their hallucinations becamemuch less severe after the treatment, which feels like ``having a woodpeckerknock on your head'' once a second for up to 16 minutes, said researcherDr.Ralph Hoffman.</s> <s docid="APW19990519.0113" num="7" stype="1">The voices stopped completely in three of these patients.</s> </P><P><sdocid="APW19990519.0113" num="8" stype="1"> The effect lasted for up toa few days for most participants, and one man reported that it lasted seven weeksafterbeing treated daily for four days.</s> </P><P><sdocid="APW19990519.0113" num="9" stype="1"> Hoffman stressed that thestudy is only preliminary and can't prove that the treatment would be useful.</s> <sdocid="APW19990519.0113" num="10" stype="1"> ``We need to do much moreresearch on this,'' he said in an interview.</s> </P><P><sdocid="APW19990519.0113" num="11" stype="1"> Hoffman, deputy medicaldirector of the Yale Psychiatric Institute, is scheduled to present the workThursday at the annual meeting of the American Psychiatric Association.</s> </P><P><s docid="APW19990519.0113" num="12" stype="1"> Not all people withschizophrenia hear voices, and of those who do, Hoffman estimated that maybe 25percent can't control them with medications even when other disease symptomsabate.</s> <s docid="APW19990519.0113" num="13" stype="1"> So the workcould pay off for ``a small but very ill group of patients,'' he said.</s> </P><P><sdocid="APW19990519.0113" num="14" stype="1"> The treatment is calledtranscranial magnetic stimulation, or TMS.</s> <s docid="APW19990519.0113"num="15" stype="1"> While past research indicates it mightbe helpful in lifting depression, it hasn't been studied much in schizophrenia.</s> </P><P><s docid="APW19990519.0113" num="16" stype="1"> In TMS, anelectromagnetic coil is placed on the scalp and current is turned on and off to create apulsing magnetic field that reaches into a small area of the brain.</s> <sdocid="APW19990519.0113" num="17" stype="1"> The goal is to make brain cellsunderneath the coil fire messages to adjoining cells.</s> </P><P><sdocid="APW19990519.0113" num="18" stype="1"> The procedure is muchdifferent from electroconvulsive therapy, called ECT, which applies pulses ofelectricity rather than a magnetic field to the brain.</s> <sdocid="APW19990519.0113"num="19" stype="1"> Unlike TMS, ECT creates a briefseizure and is performed under general anesthesia.</s> <sdocid="APW19990519.0113" num="20" stype="1"> ECT is used most often fortreatingsevere depression.</s> </P><P><s docid="APW19990519.0113" num="21"stype="1"> In TMS, the magnetic pulses are thought to calm the affected partof the brain if they're given as slowly as once per second, Hoffman said.</s> <s docid="APW19990519.0113" num="22" stype="1"> He and colleagues targeted an area involved in understanding speech, above and behind the left ear, on the theory that hallucinated voices come from overactivity there.</s> </P><P><s docid="APW19990519.0113" num="23" stype="1"> The treatment can make scalp muscles muscle contract, leading tothe woodpecker feeling, he said, but patients could tolerate it.</s> <s docid="APW19990519.0113" num="24" stype="1">Headachewas the most common side effect, and there was no sign that the treatment affected the ability to understandspeech, he said.</s> </P><P><s docid="APW19990519.0113" num="25" stype="1"> To make sure the study resultsdidn'treflect just the psychological boost of getting a treatment, researchers gave sham and real treatments to each studyparticipant and studied the difference in how patients responded.</s> <s docid="APW19990519.0113" num="26" starting with

16/20

LSAView: Algorithm Analysis/ Development

• LSAView– Analysis and exploration of

impact of informatics algorithms on end-user visual analysis of data

– Aids in discovery process of optimal algorithm parameters for given data and tasks

• Features – Side-by-side comparison of visualizations for two sets of parameters– Small multiple view for analyzing 2+ parameter sets simultaneously– Linked document, graph, matrix, and tree data views– Interactive, zoomable, hierarchical matrix and matrix-difference views– Statistical inference tests used to highlight novel parameter impact

– Used in developing and understanding ParaText™ algorithms

17/20

LSAView

18/20

20 40 60 800

0.5

1

1.5

2

20 40 60 800

0.5

1

1.5

2

20 40 60 800

0.5

1

1.5

2

20 40 60 800

0.5

1

1.5

2

k

LSAView Impact

• Document similarities:

• Inner product view:

• Scaled inner product view:

What is the best weighting of singular values for document similarity graph generation?

original scaling no scaling inverse sqrt inverse

[Leisure Studies of America Data: 97 documents, 335 terms]

19/20

Summary and Future Work

• ParaText™– Leverages VTK / ParaView / Trilinos (scientific computing)

– Scalable infoviz / text analysis pipeline

– General data object-feature analysis capability

• More work ahead …– Scaling experiments: bigger data, layout constraints, etc.

– Extensions: load balancing, incremental SVD, etc.

– Other applications: clustering, summarization, classification, etc.

– Relevance feedback: automatically incorporating user feedback to improve analysis (e.g, priors, metric learning)

20/20

Thank You

ParaText™: Leveraging Scalable Scientific Computing Capabilities for Large-Scale

Text Analysis and Visualization

Danny Dunlavy

dmdunla@sandia.gov

http://www.cs.sandia.gov/~dmdunla

21/20

Backup Slides

22/20

Feature Weighting

Term Document Matrix Scaling:

23/20

d1 : Hurricane. A hurricane is a catastrophe.

d2 : An example of a catastrophe is a hurricane.

d3 : An earthquake is bad.

d4 : Earthquake. An earthquake is a catastrophe.

d1 : Hurricane. A hurricane is a catastrophe.

d2 : An example of a catastrophe is a hurricane.

d3 : An earthquake is bad.

d4 : Earthquake. An earthquake is a catastrophe.

1011catastrophe

2100earthquake

0012hurricane

d4d3d2d1

0catastrophe

0earthquake

1hurricane

qA

.30.15.60.59catastrophe

.92.96.02-.03earthquake

.11-.11.78.78hurricane

d4d3d2d1

A2

00.71.89qTA .11–.78.78qTA2

Removestopwords

normalization only rank-2 approximation

captures link to doc 4

LSA Example 1

.450.71.45catastrophe

.89100earthquake

00.71.89hurricane

d4d3d2d1

A

24/20

LSA Example 2

∆policy∆planning∆politics∆tomlinson∆1986oSport in Society: policy, Politics and Culture, ed A. Tomlinson (1990) oPolicy and Politics in Sport, PE and Leisure eds S. Fleming, M. Talbot and A. Tomlinson (1995) oPolicy and Planning (II), ed J. Wilkinson (1986) oPolicy and Planning (I), ed J. Wilkinson (1986) oLeisure: Politics, Planning and People, ed A. Tomlinson (1985)

∆parker∆lifestyles∆1989∆partoWork, Leisure and Lifestyles (Part 2), ed S. R. Parker (1989) oWork, Leisure and Lifestyles (Part 1), ed S. R. Parker (1989)

[Leisure Studies of America Data: 97 documents, 335 terms]

25/20

Modeling Text

• Model document meaning as a linear equation– Meaning (document) = j meaning (termj)

– Induces a high dimensional vector space model

• Create term-document (occurrence) matrix– Examples of “terms”:

• Words

• Word stems, lemmas

• Entities (person, organization, location, etc.)

• N-grams (characters or words)

term

s

documentsd1 d2 dn

t2

t1

tm

…d3d4

.

.

.

26/20

Latent Semantic Analysis

• Conceptual searching– rank(k) : more exact data similarities

– rank(k) : more conceptual data similarities

– Compute larger rank, then use smaller rank

terms

k = 6

terms

k = 24

concepts concepts

increasing k

more conceptual more exact

27/20

ParaText™ Operations

• Document parsing, matrix creation and weighting• SVD:• Truncated:• Query scores (query as new “doc”):• LSA Ranking:• Document similarities:• Term Similarities:

term

s

concepts documents

con

cep

ts

singular values

T

DT

28/20

Document Similarity Graphs

• Document similarity matrix

• Document similarity graph– Each document (or term, entity, etc.) is a vertex

– Each row defines an edge

do

cum

ents

concepts documents

con

cep

ts

singular values

threshold

sparse coordinate format

29/20

Graph Similarities

• Statistics on edges– One graph: one-sample t statistic

– Two graphs: two-sample t statistic

Edges from graph 1 Edges from graph 2