The RatioLog Project-Rational Extensions

of Logical Reasoning
Submitted By:
1.Shailendra Kumar and 2K20/A5/02
2.Siddharth Mehra and 2K20/A5/10
Submitted to:Miss Deepti Singh

LOGICAL REASONING,SLOT-4
DELHI TECHNOLOGICAL UNIVERSITY
BAWANA ROAD, NEW DELHI-110042
 ACKNWLEDGMENT

I would like to express my special thanks of gratitude to my teacher Miss

Deepti Singh who gave me this oppprtunity to do this wonderful project
on the topic The RatioLog Project-Rational Extensions of Logical Reasoning ,Which also
helped me in doing a lot of research and I came to know about so many
new things I am really thankful to them.

Secondly I would also like to thank my friends who helped me a lot in

finalizing this project within the limited time frame.
 ABSTRACT

Higher-level cognition includes logical reasoning and the ability of question

answering with common sense. Our RatioLog project addresses the problem of
rational reasoning in deep question answering by methods from automated
deduction and cognitive computing. In a first phase, we combine techniques from
information retrieval and machine learning to find appropriate answer
candidates from the huge amount of text in the German version of the free
encyclopedia “Wikipedia”. In a second phase, an automated theorem prover tries
to verify the answer candidates on the basis of their logical representations. In a
third phase — because the knowledge may be incomplete and inconsistent —,
we consider extensions of logical reasoning to improve the results. In this context,
we work toward application of techniques from human reasoning: We employ
defeasible reasoning to compare the answers w.r.t. specificity, deontic logic,
normative reasoning, and model construction. Moreover, we use integrated
case-based reasoning and machine learning techniques on the basis of the
semantic structure of the questions and answer candidates to learn giving the
right answers.
 CONTENT

1.Title
2.Acknwledgment
3.Abstract
4.Chapter 1: Rational Reasoning and Question Answering
1.1 : Deep Questioning Answering and the LogAnswers System
1.2 : The LogAnswer System and its Modules
5.Chapter 2: Searching for Good Answers
2.1: CBR Similarity Measures and Machine Learning
2.2: The Specificity Criterion
2.3: Making Use of Deontic Logic
6.Conclusion
7.Reference
 Chapter 1:- Rational Reasoning and query Answering
The improvement of formal logic played a large function in the area of automated reasoning,
which caused the development of the sector of artificial intelligence (AI). Programs of
automated deduction in mathematics had been investigated from the early years on.
Nowadays automated deduction techniques are correctly carried out in tough- and software
verification and many other regions.
In comparison to formal logical reasoning, but, human reasoning does not strictly comply with
the rules of classical good judgment, reasons can be incomplete expertise, wrong ideals, and
inconsistent norms. From the very beginning of AI studies, there was a strong emphasis on
incorporating mechanisms for rationality, such as abductive or defeasible reasoning. From
those efforts, as a part of the sphere of knowledge illustration, common-feel reasoning has
emerged as a branching subject with many packages in AI.
These days there may be a risk to enroll in computerized deduction and common-experience
reasoning inside the paradigm of cognitive computing, which allows the implementation of
rational reasoning. The general motivation for the development of cognitive systems is that
computer systems can resolve well defined mathematical issues with vast precision at a pace
moderately enough in exercise. It stays hard, however, to solve issues which are best vaguely
outlined. One important function of cognitive computing is that many extraordinary
information formats and many extraordinary information processing methods are utilized in a
mixed fashion. Additionally the quantity of expertise is big and, even worse, it is even increasing
steadily. For the logical reasoning, a comparable argument holds exceptional reasoning
mechanisms have to be hired and mixed, which include classical deduction (forward reasoning)
on the only hand, and abduction or different non-monotonic reasoning mechanisms on the
other.
Let us illustrate this with a well-known example from the literature:

The LogAnswers system uses information retrieval (IR), decision tree learning (DT), reasoning and natural
languages answer generation to compute answers.

1.Tom is an emu.
2.Emus are birds.
3.Birds normally fly.
4.Emus do not fly.

The query is: Can emus fly or now not? ahead reasoning lets in us to infer that emus are birds
and consequently can commonly fly. that is in struggle, however, with the stern historical past
understanding that emus do no longer fly. The battle can be solved through assuming certain
understanding as default or defeasible, which best holds generally. for this reason we might
also conclude here that emus and consequently Tom does not fly. we can come returned to
this example later.
Rational reasoning need to be able to cope with incomplete in addition to conflicting (or even
inconsistent) information. Moreover, big expertise bases with inconsistent contents have to be
treated. therefore, it appears to be an awesome concept to integrate and thus decorate
rational reasoning by means of information retrieval strategies, e.g. techniques from system
mastering. This holds in particular for the domain of deep query answering, in which
communique with styles of human reasoning is appropriate.
1.2 Deep Question Answering and the LogAnswer System
generally, query answering systems, which includes software packages which includes okay
Google or Apple’s Siri speak with the person in natural language. They receive well formulated
questions and go back concise solutions. these automatically generated answers are usually
not extracted at once from the web, but, similarly, the system operates on an extensive
(background) understanding base, which has been derived from textual sources in advance.
LogAnswer [8, 9] is an open-domain question answering machine, available through an
internet interface (www.loganswer.de) similar to that of a search engine. The know-how used
to reply the question is won from 29.1 million naturallanguage sentences of a image of the
German Wikipedia. Moreover, a heritage understanding including 12,000 logical facts and rules
is used. The LogAnswer gadget become advanced within the DFG-funded LogAnswer task, a
cooperation between the agencies on sensible information and verbal exchange systems on
the FernUniversitat Hagen and the AI studies group on the college of Koblenz-Landau. The
mission aimed at the improvement of efficient and strong techniques for common sense-
primarily based question answering. The consumer enters a question and LogAnswer affords
the five excellent solutions from a photo of the German “Wikipedia”, highlighted in the context
of the relevant textual sources.
Maximum question answering systems depend upon shallow linguistic methods for answer
derivation, and there's best little attempt to consist of semantics and logical reasoning. This
may also make it not possible for the system to find any answers: A superficial phrase matching
set of rules is sure to fail if the textual resources use synonyms of the words inside the question.
therefore, the LogAnswer system fashions a few form of history expertise, and combines
cognitive components of linguistic analysis, inclusive of semantic nets in a logical
representation, with gadget studying strategies for determining the most appropriate solution
candidate.
Contrary to different structures, LogAnswer uses an automatic theorem prover to compute the
replies, namely Hyper, an implementation of the hypertableaux calculus, extended with
equality amongst others. It has validated its strength specially for reasoning troubles with a
massive number of inappropriate axioms, as they're function for the placing of query
answering. The logical reasoning is done on the premise of a logical representation of the
semantics of the whole text contained within the Wikipedia photo. this is computed
beforehand with a device advanced by means of computational linguists which employs the
MultiNet graph formalism (Multilayered prolonged Semantic Networks).
Given that techniques from natural-language processing are regularly faced with wrong textual
data, they strive toward robustness and velocity, however frequently lack the potential to carry
out more complex inferences. by using assessment, a theorem prover makes use of a sound
calculus to derive particular proofs of a higher complexity; even minor flaws or omissions
within the statistics, but, cause a failure of the whole derivation manner. therefore, extra
techniques from gadget learning, defeasible and normative reasoning and so forth. must be
carried out to enhance the first-rate of the solutions — as executed inside the RatioLog Project.

Techniques used in the different modules of the Loganswer system.

For this, the reasoning in classical common sense is prolonged by way of various forms of non-
monotonic factors, which include defeasible argumentation. By those extensions, the open-
domain query answering device LogAnswer is became a system for rational question
answering, which offers a testbed for the assessment of rational reasoning.
1.2 The LogAnswer System and its Modules
when processing a query, the LogAnswer machine performs numerous special steps. parent 1
presents information on those steps. in the beginning, information retrieval is used to clear out
textual content passages suitable for the given query from the textual representation of the
Wikipedia. Then selection tree gaining knowledge of chooses a fixed of solution candidates
from these textual content passages. Inside the next step, the Hyper theorem prover is used
to test if these text passages provide a solution to the question. For each solution candidate, a
primary-order good judgment illustration of both the query and the solution candidate is
combined with a big background knowledge. those proofs provide the answer to the question
by using variable assignments. The proofs for the solution applicants are then ranked the usage
of decision tree learning. For the five high-quality answers, text passages offering the solution
are highlighted and presented to the consumer.
Inside the LogAnswer system, diverse strategies work interlocked. See determine 2 for an
overview of the special techniques collectively with the modules in which they're used.
Extraction of textual content passages for a positive query is done in the candidate choice
module. in this module, both statistics retrieval and choice tree gaining knowledge of paintings
hand in hand to discover a listing of solution candidates for the current query. For each answer
candidate, the reasoning module is invoked. This module consists of the Hyper theorem
prover, which is used to test if the solution candidate provides a solution for the query.
considering the fact that Hyper is capable of cope with first-order good judgment with equality
and understanding bases given in description logic, it's miles possible to incorporate historical
past know-how given in diverse languages. An exciting extension of traditional heritage
knowledge is the usage of a understanding base containing normative statements formalized
in deontic common sense. these normative statements enable the gadget to purpose in a
rational manner. on account that deontic good judgment may be translated into description
logics, Hyper may be used to cause on such knowledge bases. Reasoning in defeasible common
sense is any other technique contained in the reasoning module of the LogAnswer gadget.
With the help of defeasible common sense reasoning, specific proofs produced by means of
Hyper are in comparison. The proofs determined with the aid of Hyper provide solutions to the
given question through variable assignments. Evaluating the proofs for one of a kind solution
candidates consequently is used to decide the first-class answer. therefore defeasible good
judgment is contained inside the answer validation module as well. In addition to that, the
solution validation module incorporates decision tree mastering to rank exclusive proofs found
by way of Hyper and casebased reasoning. Information on the usage of case-based totally
reasoning and reasoning in defeasible common sense can be observed in the section 2.

Chapter 2:-Searching for Good Answers

As depicted earlier than, the reasoning issue of the LogAnswer gadget delivers proofs, which
represent the feasible answers to the given query. The proofs are ranked by means of choice
trees which don't forget several attributes of the reasoning procedure collectively with the
attribute from the preceding information retrieval step.
Similarly to this rating we experiment with exclusive different techniques to enhance the
assessment of answers. Those are case-primarily based reasoning (CBR) (section 2.1),
defeasible reasoning (segment 2.2), and normative (deontic) reasoning (section 2.3). To carry
out systematic and giant checks with LogAnswer, we used the CLEF database, strictly speaking,
its question answering part. CLEF stands for move-language evaluation discussion board. It is
an international campaign providing language statistics in distinctive languages, e.g. from
newspaper articles. Its workshop and competition collection incorporates a song on query
answering. We used information from CLEF2007 and CLEF-2008 .
2.1 CBR Similarly Measures and Machine Learning
Answer validation can be more suitable by the usage of enjoy knowledge in form of instances
in a case base. The resulting machine module is designed as a studying machine and primarily
based on a devoted CBR manipulate shape. opposite to not unusual techniques in herbal-
language processing, but, we do no longer observe the textual approach, where experiences
are to be had in unstructured or semi-based textual content form, but use a dependent
approach along the lines of [4]. this is viable due to the fact the knowledge supply is to be had
no longer best in textual however additionally in a logical layout. The semantics of the natural-
language textual content is given basically with the aid of first-order predicate good judgment
formulae represented through the MultiNet graphs [15]. Our basis is a manually executed type
for every pair of query (from the CLEF 2007 and 2008 records) and solution candidate (from
the LogAnswer gadget) whether or not the answer candidate is a superb one for the question.
so as to evaluate and to define a similarity measure of the MultiNet graphs, we have evolved
a new graph similarity measure [14, 22] which improves other current measures, e.g. [4, 6].
We measured the CBR device class accuracy via strolling checks with a case base from the CLEF
2007 and 2008 facts. for instance, in a user interaction simulation (see parent 3), we examined
the improvement of the outcomes for a growing understanding base. We simulated customers
that deliver reliable remarks to new questions for which the LogAnswer system gives solutions
applicants. The check putting was to guess the type of questions and solution applicants the
system does not have inside the knowledge base. The consequences display the growth of the
category accuracy with a developing variety of accurate instances in the case base.

The x-axis is the wide variety of cases inside the case base. The y-axis is the classification accuracy in percentage,
for proper and wrong solution applicants, as well as the overall type accuracy for the person interaction
simulation.

We in addition incorporated case-based totally reasoning into the already existing solution
choice strategies in LogAnswer. For this, the effects of the CBR level have been become
numeric features. A rating model determined via a supervised gaining knowledge of-to-rank
technique blended those CBR-based functions with different solution choice features
determined through shallow linguistic processing and logical solution validation The final
device getting to know ranker is an ensemble of ten rank-optimizing selection timber, acquired
by way of stratified bagging, whose man or woman chance estimates are blended by using
averaging. while education the device mastering ranker on a case base optimized for best
treatment of correct answer applicants, we get the best average bring about our checks, with
an average reciprocal rank (MRR) of 0.seventy four and a accurate pinnacle-ranked answer
chosen in sixty one% of the instances. it's far instructive to take into account the usage of CBR
capabilities in the system mastering ranker, by means of analyzing all branching situations
inside the generated timber and counting the frequency of incidence of each characteristic in
one of these branching circumstance, when you consider that 10 bags of 10 selection trees
have been generated within the 10 go-validation runs, there may be a total of a hundred trees
to base consequences on [14, 22]. In general, 42.5% of all break up situations in the learned
bushes involve one of the CBR attributes. This similarly demonstrates the robust effect of CBR
consequences on solution re-ranking.
2.2 The Specifically Criterion
Extra particular answer candidates are to be desired to less unique ones, and we are able to
examine them in line with their specificity as follows. To reap what argumentation theories
name an argument, we form a pair of an answer candidate and its derivation. The derivation
may be based on positiveconditional policies, generated from Hyper’s verifications and
capturing the net page of the solution candidate and the linguistic know-how definitely carried
out. Now we discover ourselves within the putting of defeasible reasoning and might type the
arguments in keeping with their specificity.
In defeasible reasoning, sure knowledge is believed to be defeasible. Strict information, but,
is exact through contingent information (e.g., in the emu example from segment 1, “Tom is an
emu”) and general regulations conserving in all feasible worlds without exception (e.g. “emus
do now not fly”). Strict understanding is constantly preferred to knowledge relying additionally
on defeasible rules (e.g. “Birds typically fly”).
Already in 1985, David Poole had the idea to select greater precise arguments in case of
conflicting consequences as follows [19]: For any derivation of a given end result, represented
as a tree, remember the units of all leaves that make a contribution to the programs of
defeasible guidelines. An activation set is a hard and fast of literals from which all literals
labeling such a set of leaves is derivable. Thereby, an activation set is sufficient to set off the
defeasible parts of a derivation inside the feel of a presupposition, with out the usage of any
extra contingent facts.
One argument is now extra precise than some other one if all its activation sets are activation
sets of the opposite one. this means that each activation set of the greater particular argument
(seen as the conjunction of its literals) ought to be extra unique than an activation set of the
other one. note that the which means of the latter utilization of word “unique” is simply the
traditional not unusual-feel idea of specificity, in keeping with which a criterion (here:
conjunction of literals) is extra unique than every other one if it includes the opposite one.
We discovered several weaknesses of Poole’s relation, inclusive of its non-transitivity:
opposite to what's glaringly supposed in [19] and “proved” in [20], Poole’s relation isn't a quasi-
ordering and can't generate an ordering. We have been able to remedy all of the observed
weaknesses by means of defining a quasi-ordering [23, 24] (i.e. a reflexive and transitive binary
relation), which may be visible as a correction of Poole’s relation, preserving and clarifying
Poole’s authentic intuition.
The intractability of Poole’s relation, known at least in view that 2003 [21], turned into
attenuated by using our quasi-ordering after which triumph over via restricting the policies to
times that had been absolutely used in the proofs found by using Hyper, and by using treating
the last variables (if any) as constants. With these restrictions, the intractability did now not
show up anymore in any of the loads of examples we tested with our PROLOG implementation.
Walking this implementation thru the entire CLEF2008 database, almost all suggested answer
answers became out to be incomparable w.r.t. specificity, despite the fact that our
quasiordering can compare greater arguments in exercise than Poole’s authentic relation. One
hassle right here is that we ought to classify the rules of the CLEF examples as being either
trendy or defeasible, however there may be no apparent way to classify them. any other hassle
with the knowledge encoded within the MultiNet formalism is that it first and major encodes
best linguistic expertise, e.g. who's the agent of a given sentence. simplest little history
knowledge is to be had, consisting of on ontology. All statistics from the net pages, however,
are represented through literals.
To employ extra (defeasible) history expertise we investigated other examples, together with
the emu instance from section 1. right here, the formalization in first-order logic of the herbal-
language knowledge on people may be done with the Boxer device [5, 7], which is devoted to
largescale language processing programs. these examples may be successfully dealt with with
the specificity criterion and also with deontic logic (see subsequent section).
2.3 Making Use of Deontic Logic
Normative statements like “you ought not scouse borrow” are omnipresent in our normal
lifestyles and human beings are used to do motive with respect to them. considering the fact
that norms may be beneficial to model rationality, they represent an important aspect for
commonplace-experience reasoning. that is why normative reasoning is investigated in the
RatioLog project [10]. preferred deontic common sense (SDL) [11] is a good judgment which is
very appropriate for the formalization of information about norms. SDL corresponds to the
modal logic okay collectively with a seriality axiom. In SDL the modal operator ✷ is interpreted
as “it's far compulsory that” and the ✸ operator as “it's far accredited that”. as an example a
norm like “you ought no longer steal” may be intuitively formalized as ✷¬steal. From a version
theoretic point of view, the seriality axiom contained in SDL ensures that, whenever it's far
obligatory that something holds, there may be continually an excellent international pleasant
the obligation.
in the RatioLog challenge, we test with SDL with the aid of including normative statements into
the historical past expertise. The emu instance from segment 1 contains the normative
assertion
Birds generally fly.

which can be modeled the usage of SDL as

bird → ✷Flies

and is introduced to the history understanding. further to normative statements, the heritage
understanding furthermore contains assertions not containing any modal operators, e.g. some
thing just like the statement that each one emus are birds. Formulae representing contingent
data, just like the announcement

Tom is an emu.

within the emu example, are blended with the historical past expertise containing records
about norms. The Hyper theorem prover [3] can be used to research the resulting know-how
base. for instance it's far feasible to ask the prover if the determined global with the emu Tom
fulfills the norm that birds usually are able to fly.
in the RatioLog mission both defeasible good judgment and deontic logic are used. There are
similarities among defeasible common sense and deontic common sense. as an instance in
defeasible logic there are rules which can be taken into consideration to be no longer strict but
defeasible. these defeasible guidelines are just like normative statements, in view that norms
simplest describe how the world ought to be and now not the way it in reality is. this is why we
also are investigating the connection between those two logics within the RatioLog Project.

CONCLUSIONS
Deep question answering does not only require pattern matching and indexing techniques, but also
rational reasoning. This has been investigated within the RatioLog project as demonstrated in this
article. Techniques from machine learning with similarity measures and case-based reasoning,
defeasible reasoning with (a revision of) the specificity criterion, and normative reasoning with deontic
logic help to select good answer candidates. If the background knowledge, however, mainly encodes
linguistic knowledge — without general common-sense world knowledge — then the effect on finding
good answer candidates is low. Therefore, future work will concentrate on employing even more
background world knowledge (e.g. from ontology databases), so that rational reasoning can be
exploited more effectively when applied to this concrete knowledge
 REFERENCES
1. Baumgartner, P., Furbach, U., Niemela, I.: Hyper tableaux. In: ¨ J.J. Alferes, L.M. Pereira, E. Orlowska
(eds.) Proceedings of 5th European Workshop on Logics in AI – JELIA’96, LNCS 1126, pp. 1–17. Springer
(1996)

2. Beckert, B., Hahnle, R.: Reasoning and verification: State of the ¨ art and current trends. IEEE
Intelligent Systems 29(1), 20–29 (2014)

3. Bender, M., Pelzer, B., Schon, C.: System description: EKRHyper 1.4 – extensions for unique names
and description logic. In: M.P. Bonacina (ed.) CADE-24, LNCS, vol. 7898, pp. 126–134. Springer (2013)

4. Bergmann, R.: Experience Management – Foundations, Development Methodology and Internet-

Based Applications. Springer, Berlin, Heidelberg, New York (2002)

5. Bos, J.: Towards wide-coverage semantic interpretation. In: Proceedings of Sixth International
Workshop on Computational Semantics IWCS-6, pp. 42–53. Tilburg, Netherlands (2005)

6. Bunke, H., Messmer, B.T.: Similarity measures for structured representations. In: Topics in Case-
Based Reasoning: First European Workshop EWCBR-93 (1993)

7. Curran, J.R., Clark, S., Bos, J.: Linguistically motivated largescale NLP with C&C and Boxer. In:
Proceedings of the ACL 2007 Demo and Poster Sessions, pp. 33–36. Prague, Czech Republic (2007)