Argument Mining: the Bottleneck of Knowledge and Reasoning Patrick - - PowerPoint PPT Presentation

Argument Mining: the Bottleneck of Knowledge and Reasoning

Patrick Saint-Dizier

IRIT - CNRS, Toulouse, France. stdizier@irit.fr

April 14, 2016

The Relatedness Problem

Given a controversial issue: argument mining from texts ⇒ besides linguistic aspects, domain knowledge + inferences are often required:

◮ Issue: the situation of women has improved in India, ◮ Support: (a) we now see long lines of happy young girls with

school bags walking along the roads

◮ Then: (b) School buses must be provided so that schoolchildren

do not reach the school totally exhausted after a long early morning walk.

◮ (b) is an attack of (a) (these young girls may not be so happy) it

is not an attack of the issue: the facet that is concerned in the relation between (b) and (a) does not concern women’s conditions in particular.

⇒ knowledge and reasoning useful to establish relateness and polarity: the WHY, HOW and HOW MUCH.

Problem Analysis and Research questions

◮ Corpus construction: issues + arguments found in various texts ◮ How to tag arguments to characterize the need of knowledge

and reasoning ?

◮ How to categorize the knowledge involved ? ◮ How to pair NLP with KR for Argument mining ? ◮ Knowledge-driven argument mining: how to account for the

diversity of arguments w.r.t. an issue ?

◮ The Qualia of the Generative Lexicon: an useful lexical and

knowledge representation for argument mining?

◮ Case studies: knowledge vs. reasoning ?

Corpus Construction

Issue Corpus size

• nb. of arguments

(1) Ebola vaccination 16 texts, 50 is necessary 8300 words (2) Women’s condition has 9 texts, 24 improved in India 4600 words (3) The development of nu- 7 texts, 31 clear plants is necessary 5800 words (4) Organic agriculture 19 texts, 17 is the future 5800 words Total 51 texts, 122 24500 words

Arguments and argument compounds

⇒ Arguments seldom come in isolation. ⇒ They are often articulated within a context that indicates e.g.: circumstances, restrictions, illustrations, concessions, comparisons, purposes, and various forms of elaborations. ⇒ We call such a form an argument compound, where the argument is the kernel: —-> allows for a larger diversity of arguments.

Corpus Tagging

The following tags have been identified, but need to be further elaborated:

• 1. the text span involved that delimits the argument compound

and its kernel,

• 2. the polarity of the argument w.r.t. the issue: support, attack,

argumentative concession or contrast.

• 3. the conceptual relation(s) with the issue,
• 4. the knowledge involved, to identify the argument: list of the

main concepts used,

• 5. the a priori strength of the argument,
• 6. the discourse structures associated with the argument kernel.

Illustration for issue (1)

<argument nb= 11, polarity= attack with concession , relation To Issue= limited proofs of efficiency and safety of vaccination, concepts Involved= efficiency measures, safety measures, test and evaluation methods, strength= moderate > <concession> Even if the vaccine seems 100% efficient and without any side effects on the tested population, < /concession> <main arg> it is necessary to wait for more conclusive data before making large vaccination campaigns. < /main arg> <elaboration> The national authority of Guinea has approved the continuation of the tests on targeted populations. </elaboration> < /argument>.

Evidence for knowledge for argument mining

Need of knowledge: total nb of arguments / nb of those that require knowledge.

Issue need of knowledge total number of concepts nb of cases (rate) involved (estimate) (1) 44 (88%) 54 (2) 18 (75%) 23 (3) 18 (58%) 19 (4) 15 (88%) 25 Total 95 (78%) 121

Main concepts used in argument kernels and their expression in language (issue 1)

Supports: efficiency is very good, 100% protection; avoids or reduces dissemination of disease; limited side-effects, etc. Attacks: limited number of cases and deaths compared to other diseases; limited risks of contamination, ignorance of contamination forms; toxicity and high side-effects, etc. Concessions or Contrasts: some side-effects; high production and development costs; vaccine not yet available; ethical problems, etc. The above arguments are expressed in various ways:

• evaluative expressions: Vaccine development is very expensive,
• comparatives: number of sick people much smaller than for

Malaria.

• facts related to properties of the main concept(s) of the issue:

Vaccine is not yet available. There is no risk of dissemination.

• facts related to the consequences, purposes, uses or goals of

the issue: vaccine prevents bio-terrorism.

From Concepts to Knowledge Representation

The terms used in argument kernels concern: purposes, properties, parts, creation and development, etc. of the head terms of the issue or of derived concepts. These are relatively well defined in the Generative Lexicon. From issue (1): Vaccine(X):

            

CONSTITUTIVE:

• ACTIVE PRINCIPLE, ADJUVANT
• ,

TELIC:

• MAIN: PROTECT FROM(X,Y,D), AVOID(X,DISSEMINATION(D)),

MEANS: INJECT(Z,X,Y)

• ,

FORMAL:

• MEDICINE, ARTEFACT
• ,

AGENTIVE :

• DEVELOP(T,X), TEST(T,X), SELL(T,X)
• 

           

Modeling the Diversity / the Generative expansion of arguments

◮ Arguments attack or support specific facets of the concepts of

the controversial issue (called root concepts). (protect from(X,Y, (infect(E1,ebola, Y) ⇒ get sick(E2,Y) ⇒ ♦ die(E3,Y))) ∧ avoid(X,dissemination(ebola)).

◮ Arguments may also attack or support concepts derived from

these initial concepts (related to functions, parts, etc.).

◮ For example, they may attack properties or purposes of the

adjuvant or of the protocols used to test the vaccine. Arguments must however remain functionally close to the root. ⇒ Develop a network of Concepts and their Qualias derived from those involved in the controversial issue, with a limited depth. ⇒ Qualias structure the concepts in terms of parts, functions, etc.

Generative expansion of arguments: a simple example

• Consider: constitutive(vaccine(X)) = {active principle, adjuvant}.

Network of Concepts and Qualias:

• ‘active principle’: terminal concept in the network, associated with

its lexicalizations (e.g. active principle, vaccine),

• ‘Adjuvant’: non-terminal concept, included with its Qualia into the

network: Adjuvant(Y,X1):   

FORMAL :

• MEDICINE, CHEMICALS
• ,

TELIC:

• DILUTE(Y,X1), ALLOW(INJECT(X1,P))
• 

 

• Then the non-terminal concepts (e.g. medicine, dilute(Y,X1),

inject(X1,P) introduce new Qualias in the network. Natural language terms are associated to these concepts, e.g.: medicine, chemicals, inject, injection, dilute, dilution.

Case study 1

Utterance A1: The adjuvant of the Ebola vaccine is toxic Utterance A1 matches with the language pattern: [np, ‘is’, evaluative expression] A1 negatively evaluates the adjuvant (lexical feature of the adjective ‘toxic’), but it does not explicitly say anything about the vaccine. Then, given: Ebola:

    

FORMAL:

• DISEASE
• ,

TELIC:

• INFECT(E1,EBOLA, P) ⇒ GET SICK(E2,P) ⇒ ♦ DIE(E3,P)

∧ E1 E2 E3.

• 

   

The constitutive role of vaccine(X) says that the adjuvant is part of the

• vaccine. The Qualia of ‘adjuvant’ indicates that the active principle X1

is mixed by dilution with the adjuvant Y :

◮ Adjuvant Qualia: dilute(Y,X1) Y and X1 are mixed together to

form a single entity, the vaccine X.

◮ upwards inheritance of a property in a part-of relation: if a

(major) constitutive part K1 of an object K has a property P , then (probably) the entire object K has P: has property(K1,P) ∧ part of(K1,K) ⇒ has property(K,P).

◮ since Y and X1 are parts of X, then since Y is toxic for humans, it

follows that X is also toxic for humans. Therefore, A1 attacks the controversial issue. This statement may also be interpreted as a contrast to the controversial issue: ’the vaccine is necessary BUT it is toxic’. (Winterstein 2012)

A2: Seven persons died during the Ebola vaccine tests

◮ In the GL structure of vaccine(X), the ’test’ activity is

related to the agentive role.

◮ Axiomatization of the GL structure:

by definition, the agentive role is pre-telic: it occurs before the functions or the roles given in the telic role and their related properties are active: ∀ P(E) ∈ agentive-role, ∀ Q(E1) ∈ telic-role, E E1 ∧ ¬(P ⇒ Q).

◮ From that point of view, A2 is about tests, it does not say

anything about the vaccine roles, functions and consequences once it has been fully tested and approved.

◮ Argument2 is irrelevant or neutral w.r.t. the controversial

issue.

A3: No one is infected by Ebola in Europe

∀ Z, human(Z), in(Z,europe) ⇒ ¬ infect(E,ebola,Z) . (1) contradiction with telic role(epidemic(ebola)): infect(E1,ebola,Z). (2) Therefore, from the telic role of epidemic, there is no dissemination at the moment and therefore, from the telic of vaccine, no need of vaccine. (3) Therefore, Argument A3 is a partial attack of the controversial issue, with several restrictions: it is valid only in Europe and it relates a fact that occurs at the present time (compositionality rule again), (4) A3 may be analyzed as a contrast: vaccine is indeed necessary as a general principle, BUT since there are no cases in Europe at the moment it may not be necessary in Europe.

A4: The vaccine is not always efficient: 3 vaccined peopled died in Monrovia

The higher-order adjective ’efficient’ modifies the predicates in the telic role of vaccine: protect from(X,Y,Ebola), avoid(X, dissemination(X,Ebola)). (1) A4 says: ∃ Y, patient(Y), (inject(E1,X,Ebola vaccine,Y) ∧ get sick(E2,Y) ∧ ♦ die(E3,Y)) ⇒ ¬(protect from(X,Y,Ebola). (2) A4 contradicts the telic of vaccine(X). (3) Since the protection of the vaccine is not always guaranteed, statement 4 is an argument that weakly attacks the controversial issue, it is interpreted as a concession : Vaccine protects the population HOWEVER there are cases where it does not work. A concession basically supports the controversial issue, but adds some restrictions.

More complex cases

• Personal decision: everyone is free to have the vaccine

injected or not (therapeutic freedom principle),

• Temporal dependency of the statement, implicit conclusions:

there are now less cases in Africa,

• Debatable character of information sources on which the

statement is based: statistics are false, efficiency is evaluated by producer.

Required knowledge and inference: a Synthesis

◮ lexical knowledge: semantic features for lexical items, in

particular polarity, e.g. for verbs (avoid), intensifiers (for adverbs), scales, etc.,

◮ domain knowledge: encoded via the formalism of the

Generative Lexicon, including event structures and causal chains, via a network of Qualia structures,

◮ reasoning shemes:

• 1. inferences related to the semantics of the Qualia roles
• 2. inferences related to lexical semantics structures,
• 3. inferences related to general purpose or domain knowledge
• 4. inferences dedicated to argumentation, that allow to

compute relations and their strength between the controversial issue and the argument at stake. These are specific compositionality rules.

The next steps:

Next steps to further identify the WHY, HOW, HOW MUCH and the facets of an issue which are supported or attacked:

◮ categorize the types of knowledge (lexical, domain, ontology)

and inferences which are needed

◮ develop principles and methods to define GL entries ◮ investigate scalability (possibly about 50 GL entries are needed

per topic)

◮ identify processing strategies, implement and test on relatively

large sets of texts

◮ pair with ’standard’ argument mining systems or research ◮ develop more cases and specific evaluation methods. ◮ evaluation, testing: construction of clusters of closely related

arguments (argument synthesis).