Bayro-Corrochano E., Scheuermann G. Geometric Algebra Computing: in Engineering and Computer Science

Подождите немного. Документ загружается.

398 E.M.S. Hitzer et al.

Table 7 Table of triclinic and monoclinic layer groups. The pure translators T

are omitted

Layer Intern. 3D space Intern. Geom. Geom. Layer group

group # notat. group # 3D SGN 3D SGN notat. generators

Triclinic/oblique

p11P 1 P

1 p

12P

1 P 22 p22 a ∧b ∧c

Monoclinic/oblique

p112 3 P 2 P

2 p

2 a ∧b

p11m 6 Pm P1 p..1 c

p11a 7 Pc P

1 p..

1 cT

p112/m 10 P 2/m P

22 p

22 a ∧b, c

p112/a 13 P 2/c P

2 p

a ∧b, cT

Monoclinic/rectangular

p211 3 P 2 P

2 p.

2 b ∧c

11 4 P 2

(b ∧c)T

c211 5 C2 A

2 c.

2 b ∧c, T

1/2

a+b

pm11 6 Pm P1 p1 a

pb11 7 Pc P

1 p

1 aT

cm11 8 Cm A1 c1 a, T

1/2

a+b

p2/m11 10 P 2/m P2

2 p2

2 a, b ∧c

/m11 11 P 2

/m P2

a, (b ∧c)T

p2/b11 13 P 2/c P

2 p

2 aT

, b ∧c

/b11 14 P 2

/c P

, (b ∧c)T

c2/m11 12 C2/m A2

2 c2

2 a, b ∧c, T

1/2

a+b

Fig. 11 How a future

subperiodic space group

viewer software might depict

rod groups 13: p

2 and 14:

2, and the layer group

11: p1,basedon[32, 46]

ure 11 shows how the rod groups 13: p

2 and 14: p

2, and the layer group 11:

p1 might be visualized in the future, based on [32, 46].

Acknowledgements E. Hitzer wishes to thank God: It (Jerusalem) shone with the glory of God,

and its brilliance was like that of a very precious jewel, like a jasper, clear as crystal [48].

Space Group Visualization and Subperiodic Groups in GA 399

He thanks his family, his students M. Sakai, K. Yamamoto, T. Ishii, his colleagues G. Sommer,

D. Hestenes, J. Holt, H. Li, T. Matsumoto, D. Litvin, H. Wondratschek, M. Aroyo, H. Fuess,

N. Ashcroft, M. Tribelski, A. Hayashi, N. Onoda, Y. Koga, and the organizers of AGACSE 2008.

References

1. Ahlfors, L.: Clifford numbers and Moebius transformations in R

. In: Chisholm, J., Com-

mon, A. (eds.) Clifford Algebras and Their Applications in Mathematical Physics. Reidel,

Dordrecht (1986)

2. Angles, P.: Construction de revêtements du groupe conforme d’un espace vectoriel muni d’une

métrique de type (p, q). Ann. IHP Sect. A 33(1), 33–51 (1980)

3. Angles, P.: Conformal Groups in Geometry and Spin Structures. Progress in Mathematical

Physics. Birkhäuser, Basel (2007)

4. Aragón-González, G., Aragón, J.L., Rodríguez-Andrade, M.A., Verde-Star, L.: Reﬂections,

rotations, and Pythagorean numbers. Adv. Appl. Clifford Algebr. Online First (2008)

5. Bonola, R.: La Geometria Non-euclidia: Exposizione Storico-Critico del Suo Sviluppo.

Zanichelli, Bologna (1906). Translated by H.S. Carslaw, Non-Euclidean Geometry. Dover,

New York (1955). References to Friedrich Ludwig Wachter’s horosphere on pp. 62, 63 and

88. Download version: http://www.archive.org/details/noneuclideangeom00bonorich

6. Cartan, E.: La Théorie des Spineurs I, II. Actualités Scientiﬁques et Industrielles, vol. 643.

Hermann, Paris (1938)

7. Cox, H.: Application of Grassmann’s Ausdehnungslehre to properties of circles. Q. J. Pure

Appl. Math. 25, 1–71 (1891)

8. Coxeter, H.S.M.: Discrete groups generated by reﬂections. Ann. Math. 35, 588–621 (1934)

9. Coxeter, H.S.M.: Non-Euclidean Geometry, Toronto, 1942, 3rd edn. Oxford University Press,

Cambridge (1957). References to Friedrich Ludwig Wachter’s horosphere on pp. 7 and 220

10. Coxeter, H.S.M., Moser, W.O.J.: Generators and Relations for Discrete Groups, 4th edn.

Springer, Berlin (1980)

11. Crumeyrolle, A.: Chap. 12 of Orthogonal and Symplectic Clifford Algebras. Kluwer, Dor-

drecht (1990)

12. Dieudonné, J.: Sur les Groupes Classiques. Actualités Scientiﬁques et Industrielles, vol. 1040.

Hermann, Paris (1948)

13. Doran, C., Lasenby, A.: Geometric Algebra for Physicists. Cambridge University Press, Cam-

bridge (2003)

14. Dorst, L., Fontijne, D., Mann, S.: Geometric Algebra for Computer Science: An Object-

oriented Approach to Geometry. Morgan Kaufmann Series in Computer Graphics. Morgan

Kaufmann, San Mateo (2007)

15. Dress, A.W.M., Havel, T.F.: Distance geometry and geometric algebra. Found. Phys. 23(10),

1357–1374 (1993)

16. Engel, P., Matsumoto, T., Steinmann, G., Wondratschek, H.: The non-characteristic orbits of

the space groups. Z. Kristallogr., Supplement Issue No. 1. Raumgruppen-Projektionen. In Ab-

handlungen der Mathematisch-Physikalischen Klasse der Sächsischen Akademie (1984)

17. Forder, H.G.: The Calculus of Extension. Cambridge University Press, Cambridge (1941)

18. Gallier, J.: Geometric Methods and Applications. For Computer Science and Engineering.

Springer, New York (2001). Chap. 7

19. Gutierrez, J.D.M.: Operaciones de simitria mediante algebra geometrica aplicadas a grupos

cristalograﬁcos. Thesis, UNAM, Mexico (1996)

20. Hahn, T. (ed.) Space-group Symmetry, 5th edn. International Tables of Crystallography,

vol. A. Springer, Dordrecht (2005). Online: it.iucr.org

21. Hestenes, D.: Old wine in new bottles. In: Bayro-Corrochano, E., Sobczyk, G. (eds.) Geomet-

ric Algebra: A Geometric Approach to Computer Vision, Quantum and Neural Computing,

Robotics, and Engineering, pp. 498–520. Birkhäuser, Basel (2001)

400 E.M.S. Hitzer et al.

22. Hestenes, D.: Point groups and space groups in geometric algebra. In: Dorst, L., et al. (ed.) Ap-

plications of Geometric Algebra in Computer Science and Engineering, pp. 3–34. Birkhäuser,

Basel (2002)

23. Hestenes, D.: Spacetime physics with geometric algebra. Am. J. Phys. 71(7), 691–714 (2003)

24. Hestenes, D., Holt, J.: The crystallographic space groups in geometric algebra. J. Math. Phys.

48, 023514 (2007)

25. Hestenes, D., Li, H., Rockwood, A.: New algebraic tools for classical geometry. In: Som-

mer, G. (ed.) Geometric Computing with Clifford Algebras, pp. 4–26. Springer, Berlin (2001)

26. Hitzer, E.: Euclidean Geometric Objects in the Clifford Geometric Algebra of {Origin,

3-Space, Inﬁnity}. Bull. Belg. Math. Soc. Simon Stevin 11(5), 653–662 (2004)

27. Hitzer, E.: Conic sections and meet intersections in geometric algebra, computer algebra and

geometric algebra with applications. Lecture Notes in Computer Science, vol. 3519, pp. 350–

362. Springer, Berlin (2005). GIAE 2004, Revised Selected Papers

28. Hitzer, E., Perwass, C.: Crystal cells in geometric algebra. In: Proc. Int. Symp. on Adv. Mech.

Eng. (ISAME), Fukui, pp. 290–295 (2004)

29. Hitzer, E., Perwass, C.: Full geometric description of all symmetry elements of crystal space

groups by the suitable choice of only three vectors for each bravais cell or crystal family.

ISAME, Busan, pp. 19–25 (2005)

30. Hitzer, E., Perwass, C.: Crystal cell and space lattice symmetries in Clifford geometric alge-

bra. In: Simos, T.E., et al. (ed.) Int. Conf. on Numerical Analysis and Applied Mathematics,

Rhodes, pp. 937–941. Wiley, VCH, New York, Weinheim (2005)

31. Hitzer, E., Perwass, C.: Space group visualizer for monoclinic space groups. Bull. Soc. Sci.

Form 21, 38–39 (2006)

32. Hitzer, E., Perwass, C.: The space group visualizer. In: Proceedings of ISAMPE, pp. 172–181

(2006)

33. Hitzer, E., Perwass, C.: Three vector generation of crystal space groups in geometric algebra.

Bull.Soc.Sci.Form21, 55–56 (2006)

34. Hitzer, E., Perwass, C.: Interactive 3D space group visualization with CLUCalc and the Clif-

ford geometric algebra description of space groups. Proc. of ICCA8, Las Campinas, Brazil

(2008, accepted)

35. Hitzer, E., Tachibana, K., Buchholz, S., Yu, I.: Carrier method for the general evaluation and

control of pose, molecular conformation, tracking, and the like. Adv. Appl. Clifford Algebr.

Online First (2009)

36. Kopcik, V.A.: Subnikovskie Gruppy: Spravochnik po Simmetrii i Fizicheskim Svoistvam

Kristallicheskikh Struktur. Izd. Moskovskogo Univ., Moskva (1966). 722 p. (in Russian)

37. Kopsky, V., Litvin, D.B. (eds.): Subperiodic Groups, 1st edn. International Tables for Crystal-

lography, vol. E. Kluwer, Dordrecht (2002). Online: it.iucr.org

38. Li, H.: Invariant Algebras and Geometric Reasoning. World Scientiﬁc, Singapore (2008)

39. Lie, S.: On a Class of Geometric Transformations. University of Oslo Press, Oslo (1872)

40. Litvin, D.B.: Tables of properties of magnetic subperiodic groups. Acta Crystallogr. Sect. A,

Found. Crystallogr. A 61, 382–385 (2005)

41. Lounesto, P.: Clifford Algebras and Spinors, 2nd edn. Cambridge University Press, Cambridge

(2001)

42. Lounesto, P., Latvamaa, E.: Conformal transformations and Clifford algebras. Proc. Am.

Math. Soc. 79(4), 533–538 (1980)

43. Maks, J.: Modulo (1, 1) periodicity of Clifford algebras and generalized (anti-) Moebius trans-

formations. Thesis, Delft University of Technology, The Netherlands (1989)

44. Perwass, C.: CLUCalc—interactive visualization. www.clucalc.info

45. Perwass, C., Hitzer, E.: Interactive visualization of full geometric description of crystal space

groups. ISAME, Busan, pp. 276–282 (2005)

46. Perwass, C., Hitzer, E.: Space group visualizer. www.spacegroup.info (2005)

47. Souvignier, B.: Space groups. Syllabus for MaThCryst summer school, 15–20 July 2007, Ha-

vanna, Cuba

48. The Holy Bible, New International Version. International Bible Society, Colorado Springs

(1984)

Geometric Algebra Model of Distributed

Representations

Agnieszka Patyk

Abstract Formalism based on GA is an alternative to distributed representation

models developed so far: Smolensky’s tensor product, Holographic Reduced Rep-

resentations (HRR), and Binary Spatter Code (BSC). Convolutions are replaced by

geometric products interpretable in terms of geometry, which seems to be the most

natural language for visualization of higher concepts. This paper recalls the main

ideas behind the GA model and investigates recognition test results using both inner

product and a clipped version of matrix representation. The inﬂuence of accidental

blade equality on recognition is also studied. Finally, the efﬁciency of the GA model

is compared to that of previously developed models.

1 Introduction

Since the early 1980s, a new idea of representing knowledge has emerged by the

name of distributed representation. It has been the answer to the problems of recog-

nition, reasoning, and language processing—people accomplish these everyday

tasks effortlessly, often with only noisy and partial information, while computational

resources required for these assignments are enormous. To this day many models

have been built, in which arbitrary variable bindings, short sequences of various

lengths, and predicates are all usually represented as ﬁxed-width high-dimensional

vectors that encode information throughout the elements. In 1990 Smolensky [14]

described how tensor product algebra provides a framework for the distributed rep-

resentation of recursive structures. Unfortunately, Smolensky’s tensor product does

A. Patyk (



)

Faculty of Applied Physics and Mathematics, Gda

nsk University of Technology, 80-233 Gda

nsk,

Poland

e-mail: patyk@mif.pg.gda.pl

A. Patyk

Centrum Leo Apostel (CLEA), Vrije Universiteit Brussel, 1050 Brussels, Belgium

E. Bayro-Corrochano, G. Scheuermann (eds.), Geometric Algebra Computing,

DOI 10.1007/978-1-84996-108-0_19, © Springer-Verlag London Limited 2010

401

402 A. Patyk

not meet all criteria of reduced representations as the size of the tensor increases

with the size of the structure. Nevertheless, Smolensky and Dolan [15]haveshown

that tensor product algebra can be used in some architectures as long as the size of

a tensor is restricted. In 1994 Plate [13] worked up his Holographic Reduced Rep-

resentation (HRR) that uses circular convolution and vector addition to combine

vectors representing elements of a domain in hierarchical structures. Elements are

represented by randomly chosen high-dimensional vectors. A vector representing

a structure is of the same size as the vectors representing the elements it contains.

In 1997 Pentti Kanerva [10, 11] introduced Binary Spatter Code (BSC) that is very

similar to HRR and is often referred to as a form of HRR. In BSC objects are rep-

resented by binary vectors, and the boolean exclusive OR is used instead of convo-

lution. The clean-up memory is an important part of any distributed representation

model as an auto-associative collection of all atomic objects and complex statements

produced by that system. Given a noisy extracted vector, such a structure must be

able to recall the most similar item stored or indicate that no matching object had

been found.

The geometric algebra (GA) model, which is the focus of this paper, is an alter-

native to models developed so far. It has been inspired by the well-known fact that

most people think in pictures, i.e., two- and three-dimensional shapes, not by using

sequences of ones and zeroes. As far as brain functions are concerned, geometric

computing has been applied thus far only in the context of primate visual system

[5, Chaps. 1 and 2].

In the GA model convolutions are replaced by geometric products, and su-

perposition is performed by ordinary addition. Sentences are represented by

multivectors—superpositions of blades. The concept of GA ﬁrst appeared in the

19th century works of Grassmann and Clifford but was abandoned for almost a

century until Hestenes brought up the subject in [8] and [9]. The Hestenes system

has recently found applications in quantum computation (Czachor et al. [1–4, 6]),

which appears to be a promising leap from cognitive systems based on traditional

computing.

Section 2 of this paper recalls basic operations that can be performed on blades

and multivectors, using the example Kanerva [11] gave to illustrate BSC. For fur-

ther details on multivectors and interesting exercises, the reader may refer to [7, 12].

Section 3 gives rise to discussion about various ways of asking questions and inves-

tigates the percentage of correctly recognized items under two possible construc-

tions. Section 4 introduces measures of similarity based not on only the inner prod-

uct of a multivector but also on its matrix representation. Finally, Sect. 5 studies

the inﬂuence of accidental blade equality on recognition, and Sect. 6 compares the

performance of the GA model with HRR and BSC.

2 Geometric Algebra Model

Distributed representation models developed so far were based on long binary or

real vectors. However, most people tend to think by pictures, not by sequences of

Geometric Algebra Model of Distributed Representations 403

numbers. Therefore geometric algebra with its ability to describe shapes is the most

natural language to mimic human thought process and to represent atomic objects

and complex sentences. Furthermore, geometric product of two multivectors is ge-

ometrically meaningful, unlike the convolution or a binary exclusive OR operation

performed on two vectors.

In this paper we consider the C

algebra generated by the orthonormal vectors

={0,...,0, 1, 0 ...,0} for i ∈{1,...,n}. The inner product used throughout the

paper is an extension of the inner product ·|· from the Euclidean space R

.For

blades X

k

= x

∧···∧x

and Y

l

= y

∧···∧y

, the inner product ·:C

C

→R is deﬁned as

X

k

l

=



x

x

l−1

 ··· x



x

x

l−1

 ··· x



x

x

l−1

 ··· x





for k =l, (1)

X

k

l

=0fork =l (2)

and is extended by linearity to the entire algebra.

Originally, the GA model was developed as a geometric analogue of BSC and

HRR and was described by Czachor, Aerts, and De Moor in [1] and [4]. Be-

fore switching from geometric product to BSC and HRR, one has to realize that

geometric product is a projective representation of boolean exclusive OR. Let

...x

and y

...y

be binary representations of two n-bit numbers x and y, and

let c

= c

...x

= b

...b

and c

= c

...y

= b

...b

be their corresponding

blades, b

being equal to 1. The following examples show that the geometric prod-

uct of two blades c

and c

equals, up to a sign, c

x⊕y

= c

10...0

=1 =c

0...0

(10...0)⊕(10...0)

, (3)

= c

10...0

110...0

010...0

(10...0)⊕(110...0)

, (4)

= c

110...0

10...0

=−b

=−c

010...0

=−c

(110...0)⊕(10...0)

=(−1)

(110...0)⊕(10...0)

, (5)

the number D being calculated as

D =y

+···+x

) +y

+···+x

) +···+y

n−1



k<l

. (6)

The original BSC is illustrated by an example taken from [4, 11]: atomic objects

are represented by randomly chosen strings of bits, “⊕” is a componentwise addi-

tion mod 2, and “” represents a thresholded sum producing a binary vector—the

threshold is set at one half of sentence chunks, and a random string of bits is added

in case of an even number of sentence chunks to break the tie. The encoded record

PSmith =(name ⊕Pat)  (sex ⊕male)  (age ⊕66), (7)

404 A. Patyk

and the decoding of name uses the involutive nature of XOR:

Pat



= name ⊕PSmith

= name ⊕



(name ⊕Pat)  (sex ⊕male)  (age ⊕66)



= Pat  (name ⊕sex ⊕male)  (name ⊕age ⊕66)

= Pat  noise →Pat. (8)

In order to switch from BSC to HRR, the x →c

map described in [4] is used.

In the GA model, roles and ﬁllers are represented by randomly chosen blades

PSmith =name ∗Pat +sex ∗male +age ∗66. (9)

The “+” is an ordinary addition, and “∗” written between clean-up memory items

denotes the geometric product—this notation will be traditionally omitted when

writing down operations performed directly on blades and multivectors. The “

”

written in the superscript denotes the reversion of a blade or a multivector. The

whole record now corresponds to the multivector

PSmith =c

...a

...x

...b

...y

...c

...z

, (10)

and the decoding operation “”ofname with respect to PSmith is deﬁned as follows:

PSmith  name = name

∗PSmith (11)

= c

...a

...x

...b

...y

...c

...z

]

= c

±c

a⊕b⊕y

±c

a⊕c⊕z

(12)

= Pat +noise. (13)

It remains to employ the cleanup memory to ﬁnd the element closest to Pat



—

similarity is computed by means of the inner (scalar) product. When using the de-

coding symbol “”, we assume that the reader knows which model is used at the

time. Therefore, there will be no variations of the “” symbol in BSC, HRR, or in

two possible GA models (depending on the way of asking a question).

For an actual example, let us choose the following representation for roles and

ﬁllers of PSmith:

Pat = c

00100

male =c

00111

66 =c

11000

⎫

⎬

⎭

ﬁllers (14)

name =c

00010

sex =c

11100

age = c

10001

⎫

⎬

⎭

roles (15)

Geometric Algebra Model of Distributed Representations 405

The whole record then reads

PSmith =name ∗Pat +sex ∗male +age ∗66

= c

00010

00100

11100

00111

10001

11000

=−c

00110

11011

01001

. (16)

The decoding of PSmith’s name will produce the following result:

name

∗PSmith =c

00100

11001

−c

01011

= Pat +noise =Pat



. (17)

At this point, inner products between Pat



and the elements of the clean-up mem-

ory need to be compared. Item in the clean-up memory yielding the highest inner

product will be the most likely candidate for Pat:



Pat|Pat





= c

00100

·(c

00100

11001

−c

01011

) =1 =0, (18)



male|Pat





= 0, (19)



66|Pat





= 0, (20)



name|Pat





= 0, (21)



sex|Pat





= 0, (22)



age|Pat





= 0, (23)



PSmith|Pat





= 0. (24)

A question arises as to how to extract information from a multivector, should a

question be asked on the left-hand-side of a multivector

name ∗ PSmith (25)

or on the right-hand-side

PSmith ∗name. (26)

Furthermore, should we use name or rather name

? Since we can ask about both

the role and the ﬁller, we should be able to ask both right-hand-side and left-hand-

side questions according to the principles of geometric algebra. Such an approach,

however, would make the rules of decomposition unclear, which is against the phi-

losophy of distributed representations. The problem of asking reversed questions on

the appropriate side of a sentence is that we should be able to distinguish roles from

ﬁllers. This implies that atomic objects should be partly hand-generated, which is

not a desirable property of a distributed representation model. If we decide that a

question should always be asked on one ﬁxed side of a sentence, there is no point in

reversing the blade since there is no certainty that the ﬁxed side is the appropriate

one. Independently of the hand-sidedness of questions, in test results the moduli of

406 A. Patyk

inner products are compared instead of their actual (possibly negative) values. For

right-hand-side questions, we can reformulate (11)–(13) in the following way:

PSmithname =PSmith ∗name (27)

=[c

...a

...x

...b

...y

...c

...z

...a

(28)

=±c

±c

b⊕y⊕a

±c

c⊕z⊕a

(29)

=±Pat +noise =Pat



. (30)

The decoding of PSmith’s name will then take the form

PSmith ∗name =−c

00100

−c

11001

−c

01011

=Pat



, (31)

resulting in |PSmith|Pat



|= |−1|=1. We will study the effects of asking ques-

tions in various ways in the next section.

3 Recognition

Before we investigate the percentage of correctly recognized items, we need

to introduce the following deﬁnitions. Let S and Q denote the sentence and

the question, respectively. Let A be the set of all clean-up memory items A

for which SQ|A=0. We will call A a set of potential answers.Letm =

max{|SQ|A|:A ∈A} and T ={A ∈ A :|SQ|A|=m}.Apseudo-answer

is an answer belonging to the set T but different from the correct answer to SQ—

even if the difference is only in the meaning and not in the multivector. Of course,

the set T might also include the correct answer; therefore, it is called the set of

(pseudo-)correct answers and is actually the set of answers leading to the highest

modulus of the inner product. We assume that a noisy statement has been recognized

correctly if its counterpart in the clean-up memory is among the (pseudo-)correct

answers.

There are some doubts concerning how the sentences should be built—Plate [13]

adds an additional vector denoting action id (usually a verb) to a sentence, e.g.,

(eat

+eat

agt

 Mark +eat

obj

 theFish)/

√

3, (32)

where “” denotes circular convolution. We will distinguish between two types of

sentence constructions

• Plate construction, e.g., eat +eat

agt

∗Mark +eat

obj

∗theFish,

• agent-object construction, e.g., eat

agt

∗Mark +eat

obj

∗theFish.

The agent–object construction will often be denoted as “A–O” for short, especially

in table headings. Preliminary tests conducted on the GA model were designed to

investigate which type of construction suits GA better. The vocabulary set and the

sentence set for these tests are included in Table 1. The sentence set is especially

Geometric Algebra Model of Distributed Representations 407

Table 1 Contents of the clean-up memory used in tests described throughout this paper

Number Contents

of blades

1 A total of 42 atomic objects: 19 ﬁllers, 7 single-feature roles, and 8 double-feature roles

2(1a)bite

agt

∗Fido +bite

obj

∗Pat

(2a) ﬂee

agt

∗Pat +ﬂee

obj

∗Fido

3(3a)see

agt

∗John +see

obj

∗(1a)

(PSmith) name ∗Pat +sex ∗male +age ∗66

4 (1b) bite

agt

∗Fido +bite

obj

∗PSmith

(2c) ﬂee

agt

∗PSmith +ﬂee

obj

∗Fido

(4a) cause

agt

∗(1a) +cause

obj

∗(2a)

5 (3b) see

agt

∗John +see

obj

∗(1b)

(5a) see

agt

∗John +see

obj

∗(4a)

6(4c)cause

agt

∗(1b) +cause

obj

∗(2a)

7(DogFido) class ∗animal +type ∗dog +taste ∗chickenlike

+name ∗Fido +age ∗7 +sex ∗male +occupation ∗pet

8(1c)bite

agt

∗DogFido +bite

obj

∗Pat

(2b) ﬂee

agt

∗Pat +ﬂee

obj

∗DogFido

(4b) cause

agt

∗(1b) +cause

obj

∗(2c)

9(3c)see

agt

∗John +see

obj

∗(1c)

(5b) see

agt

∗John +see

obj

∗(4b)

10 (1d) bite

agt

∗DogFido +bite

obj

∗PSmith

(2d) ﬂee

agt

∗PSmith +ﬂee

obj

∗DogFido

11 (3d) see

agt

∗John +see

obj

∗(1d)

Each sentence carries a number (e.g., “(3a)”) to make further equations more readable

ﬁlled with similar sentences to test sensitivity of the GA model to confusing data.

Each sentence carries a number (e.g., “(3a)”) to make further equations more com-

pact and readable.

3.1 Right-Hand-Side Questions

In the previous section we commented on the use of reversions and the choice of

the side of a statement that a question should be asked on. The argument for right-

hand-side (generally: ﬁxed-hand-side) questions without a reversion was that rules