Learning Semantic Context-sensitive Term Associations for Information Retrieval
Tamsin MaxwellSchool of Informatics, University of Edinburgh
Dawei SongSchool of Computing, The Robert Gordon University
Outline
Motivation Context-sensitive Information Inference and
Semantics Event Extraction Algorithm Application in Information Retrieval
Motivation
T1 = “ President Ronald Reagan ”
US former president, administration, budget, tax, etc.US former president, administration, budget, tax, etc.
T2 = “ President Reagan and Iran-Contra affair ”
Iran arms sales scandalIran arms sales scandal
“Reagan” in different contexts
T3 = “ Reagan and Nakasone ”
Japan trade warJapan trade war
Motivation
T2 = “President Reagan and Iran-Contra affair”
Iran arms sales scandalIran arms sales scandalInformation Inference
“Reagan” in context of “Iran contra” carries/implies the information of “arms sales scandal”
Context-sensitive Information Inference
Automatic derivation of implicit term associations from text Multi-dimensional representation of information Concept combination Information flow computation
Kemp oppose president reagan stock tax urges
kemp 3 5 4 2 1 6
oppose 6 5
president 5 6 4 4 2
reagan 6 5 4
stock 6
tax
urges 4 6 5 3 2
Multi-dimensional Representation of Information
Hyperspace Analogue to Language (HAL)
Reagan = < administration: 0.46, bill: 0.07, budget: 0.08, congress: 0.07, economic: 0.05, house: 0.09, officials: 0.05, president: 0.80, reagan: 0.09, senate: 0.05, tax: 0.06, trade: 0.09, veto: 0.08, white: 0.06, …>
Multi-dimensional Representation of Information
Collection: Reuters-21589
...presence (on) German soil. (The) Germans, given (as they are to) romanticism, pacifism...
4 45 6 5
6
weight window size: 6
weight = window_size – distance + 1
“…presence on German soil. The Germans, given as they are to romanticism, pacifism and self-absorption, aren't sure whether they will allow American nuclear weapons to remain in Germany much longer.” --WSJ 1990
Handling Complex Sentences
soil: 6, given: 6, German: 5, romanticism: 5, presence: 4, pacifism: 4
HAL vs Semantic HAL
Semantic HAL
allow: 6, weapons: 6, want: 6, missiles: 6, seem: 6, believe: 6, American: 5, nuclear: 4
allow Germans weapons American nuclear want not Germans missiles seem Germans believe Germans
“The Germans, given as they are to romanticism, pacifism and self-absorption, aren't sure whether they will allow American nuclear weapons to remain in Germany much longer.”
Combining Vectors in HAL Space
A more general and flexible way of deriving the meaning from any arbitrary composition of related terms, not being limited to syntactically valid phrases.
Information Flow
Combining Vectors in HAL Space
Concepts ordered by dominance values (based on IDF) Scaling the dimensions in the dominant concept higher Increase the weights of intersecting dimensions Vector addition Normalize the composition vector and set a threshold to cut
off lowly weighted dimensions For more than two concepts, this can be done recursively
Reagan = < administration: 0.46, bill: 0.07, budget: 0.08, congress: 0.07, economic: 0.05, house: 0.09, officials: 0.05, president: 0.80, reagan: 0.09, senate: 0.05, tax: 0.06, trade: 0.09, veto: 0.08, white: 0.06, …, >
Iran = < arms: 0.71, attack: 0.18, gulf: 0.21, iran: 0.33, iraq: 0.31, missiles: 0.11, offensive: 0.13, oil: 0.18, reagan: 0.10, sales: 0.20, scandal: 0.25, war: 0.20, … >
Reagan Iran= < administration: 0.11, affair: 0.06, arms: 0.72, attack: 0.08, contra: 0.14, deal: 0.08, diversion: 0.07, gulf: 0.11, house: 0.10, initiative: 0.06, iran: 0.22, november: 0.06, policy: 0.07, president: 0.26, profits: 0.08, reagan: 0.23, sales: 0.15, scandal: 0.31, secret: 0.06, senate: 0.06, war: 0.12 >
Combining Vectors in HAL Space
Combining Vectors in HAL Spacewith Semantics
Concepts can be ordered by semantic dominance (based on IDF)
weapons American nuclear Use modification dictionary in event parser Proceed as for normal HAL space
Pred=allow Arg0=they modArg0a=weapons modArg0b=American modArg0c=nuclear
dominates dominates
HAL-based “information flow”
)degree( iff ,,1 jin ccjii
scandal iran reagan,
Barwise & Seligman (1997)
Information described by tokens i1…,in carries information described by j
..with respect to a given collection
iff concepts are included
Event Extraction Algorithm
Preprocessing Combined syntactic-semantic parsing
Semantic role labeling Dependency parsing
Trace the dependency tree from predicates and arguments to identify event structure
Event or modifier pruning
Semantic Role Labeling
SRL for Event Representation
Not all predicates indicate events Events are interpreted using dependencies
Semantic-Syntactic Parsing
Event Extraction
replied defendant permit not replied defendant enjoy lands
The defendant replied that no City permit was necessary as defendant lands enjoy interjurisdictional immunity…
Application in Information Retrieval
IR can be viewed as a reasoning process to capture the information transformation Query Expansion: QQ’
The use of information flow to derive an improved query
space program |-
program:1.00 space:1.00 nasa:0.97 U.S.:0.96 agency:0.95 shuttle:0.95 national:0.95 soviet:0.95 aeronautics:0.87 satellite:0.87 scientists:0.83 flights:0.78 pentagon:0.78
Information Flow for Query Expansion
Q as initial query submitted to a search system Apply information flow computation to a number (e.g., 30) of pseudo-
relevant documents A number of top ranked information flows derived from Q and their
associated weights form an expanded query Submit the expanded query back to the retrieval system and evaluate
the average precision of the newly retrieved documents
Aspect Hidden Markov Model
Qj
di di+1 di+2
Qj+1
w
P(Qj)
P(w|di,Qj)
P(di|Qj)
... ...
... ...
P(w|Q)
)|(*)|(),|(
)(*)|(*),|()|(;
dwPQwPdQwP
QPQdPdQwPQwP
jj
jjDdQQ
j
i
QqqqQkjjjj },...,,{
21
Information flow
Importance of Qj in Q
Q = {space program} {{space}, {program}, {space program}}
Huang, Q., and Song, D. (2008) A Latent Variable Model for Query Expansion Using the Hidden Markov Model. ACM 17th Conference on Information and Knowledge Management (CIKM 2008), poster, pp. 1417-1418.
Evaluation
Baseline Relevance Model
InformationFlow
AHMM
AP89Topics 1-50
0.1991 0.2270(+14%)
0.2677(+34.5%)
0.2778(+39.3%)
AP88-89Topics 101-150
0.2338 0.3069(+31.3%)
0.3193(+36.6%)
0.3259(+39.4%)
AP88-89Topics 151-200
0.3135 0.3471(+10.7%)
0.3965(+26.5%)
0.4081(+30.2%)
Food for Thought
Can incorporation of semantic word dependencies consistently enhance IR precision/performance?
Can they be incorporated into existing IR systems?
References
Dawei Song and Peter Bruza (2001), Discovering Information Flow Using a High Dimensional Conceptual Space. SIGIR 2001: 327-333.
Dawei Song and Peter Bruza (2003), Towards Context Sensitive Information Inference. JASIST 54(4): 321-334.
K. Tamsin Maxwell, Jon Oberlander and Victor Lavrenko (2008). Evaluation of Semantic Events for Legal Case Retrieval. ESAIR 2008: 39-41.
Huang and Dawei Song (2008), A Latent Variable Model for Query Expansion Using the Hidden Markov Model. ACM 17th Conference on Information and Knowledge Management (CIKM 2008), poster, pp. 1417-1418.
Questions?
Thank you!
Baseline Relevance Model
AHMM
AP88-90 (730MB)Topics 151-200
0.2077 0.2639(+27.1%)
0.2806(+35.1%)
ROBUST (1.9GB)Topics 601-700
0.2920 0.3143(+7.1%)
0.3660(+25.3%)
WT10G (10.9GB)Topics 501-550
0.2032 0.2134(+5%)
0.2370(+16.6%)
Aspect Hidden Markov Model Evaluation
Query: “What is the liability of the United States under the Federal Tort Claims Act for injuries sustained by employees of an independent contractor working under contract with an agency of the United States government?”
Document: “The DEFENDANT replied that no City permit was necessary because DEFENDANT lands enjoy interjurisdictional immunity as public property within the meaning of STATUTE of the Constitution Act , 1867 , or because the management of those lands is vital to the DEFENDANT ‘s federal under taking pursuant to the federal STATUTE jurisdiction over navigation and shipping .”
Sample Legal Query