1 Argument-based Critics and Recommenders: A qualitative perspective on user support systems Universitat de Girona (Febrer 2006) Carlos Iván Chesñevar [email protected]Artificial Intelligence Research Group Department of Computer Science - University of Lleida Lleida, Catalunya (Spain)
70
Embed
Argument-based Critics and Recommenderseia.udg.es/arl/Agentsoftware/PDF_ULl_2006.pdf1 Argument-based Critics and Recommenders: A qualitative perspective on user support systems Universitat
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Artificial Intelligence Research GroupDepartment of Computer Science - University of Lleida
Lleida, Catalunya (Spain)
2
Where I am from…
Univ. Nacional del Sur Bahía Blanca (Argentina)
University of Lleida Lleida, Catalonia (Spain)
3
Research linesResearch lines
•• Theory of Argumentation in a CS settingTheory of Argumentation in a CS setting
•• Argumentation: Applications and extensionsArgumentation: Applications and extensions
I have been working in I have been working in defeasible defeasible argumentationargumentation since 1992. since 1992.
Main research lines:
4
Applications & extensions Applications & extensions •• Recommender Systems and argumentationRecommender Systems and argumentation
•• Semantic Web Applications of ArgumentationSemantic Web Applications of Argumentation
• Learning and argumentationLearning and argumentation
•• Decision Theoretic extensions of DeLPDecision Theoretic extensions of DeLP
•• Integrating DeLP and Neural Networks Integrating DeLP and Neural Networks
•• Implementation Issues in Argument SystemsImplementation Issues in Argument Systems
5
About this talk• Argumentation systems provide a sound setting to formalize qualitative reasoning, with a growing number of real-world applications.• Recommender Systems (or suggesters) have evolved as sophisticated tools to assist user in computer-mediated tasks (mainly web-oriented).• Proposal: to integrate RS technologies with an argumentation-based framework called DeLP, based on extended logic programming.
6
Outline
•(2) Recommender Systems (RS)•(1) Argumentation & DeLP
•(3) Argument-Based RS
•(4) A case study: • ArgueNet: An Argument-Based Search Engine• Argumentation for Natural Lang. Assessment
•(5) Conclusions. Ongoing work.
7
Defeasible Argumentation: a very brief overview
• When a rule supporting a conclusion may be defeated by new information, we say that such a rule is defeasible.• When we chain defeasible reasons to reach a conclusion, we say that we have an argumentinstead of a proof (in a logical sense).• Arguments may compete, rebutting (attacking) each other, so that the process of argumentation is a natural result of a search for arguments.• Argumentation systems formalize the above notions in different ways.
8
Logic Programming & Defeasible Argumentation
• Recently several approaches based on combining logic programming & defeasible argumentation have been proposed.• Defeasible Logic Programming (DeLP) is one of such approaches.• DeLP has been particularly promising for modelling problems in the context of developing real-world applications.
9
Logic Programming
IntelligentAgents
Multiagent Systems
DefeasibleArgumentation
DeLP: related areas
DeLP
10
DeLP: SyntaxA fact is a ground literal. E.g.: innocent(mark)
A strict rule is denoted Lo ← L1 ,L2 , ... ,Ln
and stands for sound information
A defeasible rule is denoted Lo % L1 ,L2 , ... ,Ln
and stands for tentative information
~guilty(mark) ← innocent(mark)mammal(X) ← dog(X)
flies % bird~good_weather % low_pressure
11
Defeasible Logic Program A Defeasible Logic Program (delp) is a set of facts, strict rules and defeasible rules denoted P= (Π , ∆),where- Π is a set of facts and strict rules.- ∆ is a set of defeasible rules.
Defeasible DerivationA defeasible derivation for L from P= (Π , ∆) is a finite sequence of ground literals L1 ,L2 , ... , Lnsuch that for each literal Li- Li is a fact, or- there exists a rule in P with head Li and bodyB1,B2,.. Bk and each literal of the body is in L1,L2,..,Li-1
A set S ⊆ (Π,∆) is contradictory if there exists adefeasible derivation for a pair of complementary literalsp and ~p from S.Obs: for any delp P= (Π , ∆), the set Π is required to be not contradictory.
But..there is also a derivationfor “~buy_stock(acme)”:
good_price(acme)In_fusion(acme,steel)
risky_company(acme)~buy_stock(acme)
15
Argument•An argument for a ground literal h from P= (Π , ∆) is a pair <A,h> where A is a set of defeasible rules such that:(1) there exists a defeasible derivation for h from A∪Π.(2) A∪Π is not a contradictory set.(3) A is a minimal set satisfying (1) & (2)
Example: < A ,~buy_stock(acme)>A={ ~buy_stock(acme) % good_price(acme),
<B,p> is a counter-argument for <A,h> ifthere exists a subargument <S,q> of <A,h>such that {p,q} ∪ Π is a contradictory set.Note: the subargument can be <A,h> itself.
h
A
q
S
p
B
h
A
p
B
18
Example of counterarguments
~risky_company(acme)
in_fusion(acme,steel) strong(steel)
in_fusion(acme,steel) strong(steel)
~buy_stock(acme)
good_price(acme) risky_company(acme)
good_price(acme) in_fusion(acme,steel)
in_fusion(acme,steel)
19
Comparison of arguments
Arguments and counterarguments arearguments in conflict.
An argument can be deemed better than another by using a preference criterionamong arguments: “≤” ⊆ Args x Args
DeLP relies on specificity as a syntax-based preference criterion.
20
Comparison of argument (1)
• Specificity criterion: prefers arguments(a) with more information or (b) more direct
Preference criterion among arguments is modular (ie, specificity could be changed for other alternative criterion).
>spec
22
Defeater<B,p> is a defeater for <A,h> iff <B,p> is a counter-argument for <A,h> attacking the subargument <S,q>and:
a) <B,p> is strictly better than <S,q> wrt the preference criterion (proper defeat), or
b) <B,p> is not comparable with <S,q> wrt the preference criterion (blocking defeat)
h
A
q
S
p
B
23
DeLP with Default Negation
h % not pp % cc
h{ h% not p }
•DeLP can be extended to include default negated literals (not p) as new potential points of attack.
p{ p % c }
Blocking defeater
Discussion: “Relating Defeasible and Normal Logic Programming through Transformation Properties” (C.Chesñevar, J.Dix, F.Stolzenburg, G.Simari). Theoretical Computer Science, Vol.290, Issue 1, pp. 499-529, 2003.
24
This results in tree-like structure, in which every path can be thought of as an exchange of arguments between two parties (Pro and Con).
Arguments, defeaters, and defeaters for defeaters, ….
A
h As defeaters are arguments, they may on its turn be defeated by other arguments…
25
A0
h0
Argumentation lineAn argumentation line is a sequence
[<A0,h0>, <A1,h1>, <A2,h2>, <A3,h3>, <A4,h4>,...] where each argument (except for the first) is a defeater for the previous argument in the sequence.
h1
A1
h2
A2
h3
A3
h4
A4
Argumentation lines should be acceptable by verifying some additional constraints (e.g. no cycles) in order to avoid fallacious argumentation.
26
U
UD
D
D
U
UD
A
h
An argument <A,h> is warrantedif the root of the associated tree T<A,h> is labelled as U.
In order to determine whether an argument <A,h> is ultimately acceptable, a dialectical tree T<A,h> rootedin <A,h> is built.
Leaves are U-nodes.
Inner node is U iff everychildren node is a D-node.
Inner node is D iff at least one children node is a U-node.
Dialectical Tree
Arguments are undefeated or defeated nodes (U- or D-nodes).
27
How DeLP works
DeLPInterpreter
Defeasible rules
AbstractMachine
?- buy_stock(acme)
• YES (there exists a warrantedargument <A,h> )
• NO (there exists a warrantedargument for <A,~h>)
• UNDECIDED (none of theabove cases hold).
Possible Answers to Query h
User Query
Strict rules Facts
DeLP Program P
Obs: for any DeLP program P, it holds that h and ~h cannot be both warranted.
28
DeLP Implementation
• A special abstract machine (similar to Warren’s abstract machine for Prolog) was developed to efficiently solve queries in DeLP.• A Java-based Integrated Development Environment (IDE) for DeLP (based on this abstract machine) is available.• A web-based querying service for solving DeLP queries is currently under development, available at:
http://lidia.cs.uns.edu.ar/DeLP
29
DeLP Applications• Several DeLP-based applications have been developed.• E.g.: pattern recognition, web-based forms, recommender systems, etc.• Current research focuses on:Integrating DeLP and Knowledge Mangamentunder the JITIK Platform (in collaboration with TEC Monterrey, Mexico).Integrating DeLP and recommendation technologies (in collaboration with Ana Maguitman, Indiana Univ., USA & Univ. Nac. Del Sur, Argentina).
30
Outline
•(2) Recommender Systems (RS)•(1) Argumentation & DeLP
•(3) Argument-Based RS
•(4) A case study: • ArgueNet: An Argument-Based Search Engine• Argumentation for Natural Lang. Assessment
•(5) Conclusions. Ongoing work.
31
• Recommender Systems address the problem of information overload by providing guidelines or hints.
The Problem: Information Overload
32
Recommender Systems (RS)• RS are programs that create a model of the
user’s preferences or the user’s tasks to help identify worthwhile stuff (news, web pages, books, etc.)
• Find what users want.• Know what users like.• Gain trustworthiness from users.
Goals
33
Limitations of Traditional Views• Mostly unable to perform qualitative inference on
the recommendations.• Mostly unable to deal with the defeasible nature of
user’s preferences.• Unable to provide explanations: trustworthiness
issues!
Integrate recommender system technologies with DeLPOur Proposal
Remark: part of this research work is being jointly developed with Ana Maguitman (Indiana University, USA), who has been working with web-based suggesters, recommender systems and automated concept map generation (supported by a NASA Project).
34
Traditional Approaches to RS• Collaborative Filtering Recommenders: infer
preferences of individual users based on behavior of multiple users.
• Content-Based Recommenders: infer preferences of individual users based on what the user liked in the past.
• Hybrid Recommenders: combine both.
35
Hybrid RS: outline
36
Outline
•(2) Recommender Systems (RS)•(1) Argumentation & DeLP
•(3) Argument-Based RS
•(4) A case study: • ArgueNet: An Argument-Based Search Engine• Argumentation for Natural Lang. Assessment
•(5) Conclusions. Ongoing work.
37
Argument-based RS
Proposal: Model the users’ preference criteria in terms of a DeLP program built on top of a content-based search engine.
Users’ preference criteria are generally:•Incomplete.•Potentially Inconsistent.
As a basis for manipulating knowledge, we rely on techniques forconverting XML code into first-order formulas.A. Hunter, R. Summerton: Fusion Rules for Context-Dependent Aggregation of Structured News Reports. Journal of Applied Non-Classical Logics 14(3): 329-366 (2004)
38
Encoding Users’ Preferences
DeLP Program
P
Puser : preferences and behavior of active user
Ppool: preferences and behavior of pool of users
Pdomain: domain background knowledge
39
Argument-Based RS Architecture
40
Prioritizing Recommendations• Recommendations can be prioritized
according to their epistemic status:• Sw: warranted results: those results si for which
there exists at least one warranted argument supporting “rel(si)” wrt P’
• Sd: defeated results. those results si for which there exists at least one warranted argument supporting “~rel(si)” wrt P’
• Su: undecided results: results which are neither warranted nor defeated.
41
Distinguished predicate name to evaluate relevance of results
Algorithm RecommendOnQueryInput: Query Q, DeLP program P = Puser ∪ Ppool ∪ PdomainOutput: List LnewBeginLet L=[s1, s2, … sk] be the output of solving Q wrt content-
based search engine SEPsearch = {facts encoding info(s1), info(s2), … info(sk) }P’ = P ∪ PsearchInitialize Sw, Su, Sd as empty setsFOR EVERY si in L
DOSolve query “rel(si)” using DeLP program P’IF rel(si) is warranted wrt P’ THEN add si to Sw
ELSE IF ~rel(si) is warranted wrt P’ THEN add si to Sd
•(2) Recommender Systems (RS)•(1) Argumentation & DeLP
•(3) Argument-Based RS
•(4) A case study: • ArgueNet: An Argument-Based Search Engine• Argumentation for Natural Lang. Assessment
•(5) Conclusions. Ongoing work."Argument-Based Critics and Recommenders: A Qualitative Perspective on User Support Systems" (C. Chesñevar, A. Maguitman, G. Simari). In Data & Knowledge Engineering (to appear), 2006
43
Argument-Based Search Engine
44
Case-Study: Solving Web Search Queries• Consider a journalist who wants to search for
news articles about recent outbreaks of bird flu.
Outbreaks of bird flu
?
news bird flu
Too many results!Proposal: rank search results according to their epistemic status wrt a DeLP program representing user’s knowledge
S4: Is this Article Relevant?author(s4, bob_beak)address(s4,“mynews.com/..”)date(s4, 20031003)
rel(s4)
∅
S4Status of s4 Warranted as relevant
S1
S2
S3
S4
S4
S3
S1
S2
After analysis based on implicit knowledge
Warranted as rel.
Warranted as rel.
Undecided
Warranted as non-rel.
55
Outline
•(2) Recommender Systems (RS)•(1) Argumentation & DeLP
•(3) Argument-Based RS
•(4) A case study: • ArgueNet: An Argument-Based Search Engine• Argumentation for Natural Lang. Assessment
•(5) Conclusions. Ongoing work.
56
Using The Web as a Corpus• A huge amount of sentences in natural languages are available as Web documents.
• Pattern matching capabilities of current search engines allow to use the Web as a linguistic corpus, reflecting the current status of a living language (e.g. English, Spanish, etc.).
• We developed an argumentative approach to provide proactive assistance for language usage assessment on the basis of the Web Corpus
"An Argument-based Decision Support System for Assessing Natural Language Usage on the Basis of the Web Corpus" (C. Chesñevar, M. Sabaté, A. Maguitman). In Intl. Journal of Intelligent Systems, Wiley (in press), 2006.
57
Assessing Natural Language Usage• Usage indices provide statistical measures which are relevant to assess natural language usage.
Situations like these can be detected via usage indices. Usage Indices can be implemented as built-in Prolog predicates, and computed from hit-counts associated with Google queries.
E.g: in English, “warm” and “hot” have similar meanings…However, “warm regards” is pragmatically valid in English whereas “hot regards” is not !
E.g.: in English, “associated with” is correct whereas “associated to” is incorrect (common error for Spanish speakers).
58
Usage indices: examplesConsider the strings “rearing children”, “parents”, and “of twins”, and domains “.uk” and “.babycentre.co.uk”. Then it holds that:
• Constrained usage UC(s,D) = how many times a given string s appears in set D of web domains.• Ratio usage UR(s,D1,D2) = ratio between frequency of string s in D1 and frequency of string s in D2.• Prefix usage UP(s1,s2,D) = likelihood of finding a string s1immediately preceding another string s2 in a page from some domain in D.
59
Application Area
Consider an American journalist who writes articles in Spanish about Latinamerican issues intended for audiences both in Spain andArgentina....Spanish is not her mother tongue… she is in doubt about the usage of certain expressions which could lead to misunderstandings.
I’d like to write something about the “corralito”economic crisis in Argentina in 2001…
DeLP program encoding defeasible knowledge about language usage
60
Defeasible rules for language usageacc(S) % common_spanish(S)