Introducing Scalable Quantum Approaches in

Language Representation

Peter Wittek and Sándor Darányi

Swedish School of Library and Information Science
Göteborg University & University of Borås
Allégatan 1, 50190 Borås, Sweden
[email protected],[email protected]

Abstract. High-performance computational resources and distributed

systems are crucial for the success of real-world language technology ap-
plications. The novel paradigm of general-purpose computing on graph-
ics processors (GPGPU) offers a feasible and economical alternative: it
has already become a common phenomenon in scientific computation,
with many algorithms adapted to the new paradigm. However, appli-
cations in language technology do not readily adapt to this approach.
Recent advances show the applicability of quantum metaphors in lan-
guage representation, and many algorithms in quantum mechanics have
already been adapted to GPGPU computing. SQUALAR aims to match
quantum algorithms with heterogeneous computing to develop new for-
malisms of information representation for natural language processing in
quantum environments.

1 Introduction
Quantum mechanics is a very successful scientific theory for making pre-
dictions about systems with inherent ambiguity in them. That natural
language bears similarities with such a system is at least plausible. Re-
cent advances in theory and experimentation to apply quantum mechanics
to non-quantum domains include the use of quantum algorithms to ad-
dress, or to more efficiently solve, problems in such domains (including
contrasts between classical vs. quantum methods), such as applications
of artificial intelligence, information retrieval, and language modelling.
The quantum metaphor promises improved methodologies to capture
the subtleties and ambiguities of human language, resulting in optimised
algorithms for text processing. The purpose of SQUALAR is to investigate
methods borrowed from the field of quantum mechanics in a wide range of
large-scale language technology applications by seeking a match between
quantum algorithms and heterogeneous computing.
To this end, a scalable environment is a must. Latest trends indicate
the rise of a heterogeneous platform in which multi-core central processing
units (CPUs) and graphics processing units (GPUs) work together in a
distributed-memory parallelism. CPU-based parallelism has been utilized
for decades, and while not without its own problems, it is a mature field
and multicore CPUs enable developing faster algorithms with reasonable
effort. In this paradigm, there is a considerable overhead on dividing the
problem, distributing the bits along a small number of CPU cores, then
collecting and merging results. This type of parallelism is available in a
wide range of programming languages, although the source code needs to
be modified to some extent. GPU-based parallelism is a completely differ-
ent approach. The overhead of splitting the work is minimal, the number
of cores is massive, but the kind of computations that can be split is
limited to a simple, single-pass operation. This heterogeneous computing
environment has to be studied at different levels to find scalable imple-
mentations: low-level linear algebra, numerical methods, kernel methods
and manifold learning are candidates for testing, as well as higher level
load distribution such as MapReduce [1]. The constraints are as follows:

– Text processing is typically a data-intensive task, and several dis-

tributed algorithms have been proposed to deal with large-scale col-
lections on a grid or in a cloud computing environment. MapReduce1
was originally developed to this end, and mature libraries, such as
Cloud9, are readily available [2]. Other libraries, such as Mahout2 , fa-
cilitate the development of complex language technology applications.
– General-purpose computing on the GPU requires considerable effort
from developers. Initial results in text processing, however, indicate
that the improvement in execution time can be considerable [3–6].
– Quantum methods, on the other hand, rely on linear algebra and
other numerical libraries, many of which have already been optimized
to utilize the power of GPUs [7–10].

SQUALAR intends to bring the best of two worlds together. By bridg-

ing data-intensive text processing with sophisticated quantum modelling
of languages, we expect to see major advances in language technology.
The challenges, however, are far from trivial. The major frameworks
of GPGPU programming, CUDA and OpenCL, require wrapping in Java,
which is the environment of Hadoop, the most mature open source MapRe-
duce implementation. This paper offers an insight on the initial stage of
our ongoing investigation.
1
http://hadoop.apache.org/mapreduce/
2
http://mahout.apache.org
 This paper is organized as follows. Section 2 defines what we mean
by heterogeneous computing: a distributed system of nodes which are
equipped with multicore CPUs and GPUs. Section 3 gives a very short
overview of quantum approaches in language processing, with a focus
on methods that have the potential for acceleration. Section 4 discusses
how we intend to bridge heterogeneous computing and these quantum
approaches, and finally Section 5 concludes our paper.

2 Heterogeneous computing

Heterogeneous computing aims to combine the parallelism of traditional

multicore CPUs and GPU accelerator cores to deliver unprecedented lev-
els of performance [11]. While the phrase typically refers to single node,
a distributed environment may be constructed from such heterogeneous
nodes.
CPUs excel in running single-threaded processes, or in multithreaded
applications in which a thread often consists of fairly complicated se-
quential code. Graphics processors are ideally suited for computations
that can be run on numerous data elements simultaneously in parallel.
This typically involves arithmetic on large data sets (such as matrices)
where the same operation can be performed across thousands of elements
at the same time. This is actually a requirement for good performance:
the software must use a large number of threads. The overhead of creating
new threads is minimal compared to CPUs that typically take thousands
of clock cycles to generate and schedule, and a low number of threads will
not perform well on GPU [12]. The decomposition and scheduling of com-
putation among CPU cores and GPUs are not trivial even on a single node
[13–15], and the task is even more complicated for clusters [16]. In order to
issue work to several GPUs concurrently, a program needs the same num-
ber of CPU threads, each with its own context. All inter-GPU communi-
cation takes place via host nodes. Threads can be lightweight (pthreads,
OpenMP, etc. [17]) or heavyweight (MPI [18]). Any CPU multi-threading
or message-passing API or library can be used, as CPU thread manage-
ment is completely orthogonal to GPGPU programming. For example,
one can add GPU processing to an existing MPI application by porting
the compute-intensive portions of the code without changing the commu-
nication structure [19]. However, the efficient utilisation of all CPU and
GPU cores remains an open question.
While research is being carried out to develop the formal founda-
tions of efficient scheduling and decomposition in multiple heterogeneous
nodes, GPU-based clouds are becoming available 34 , and initial investi-
gations have been carried out to develop an efficient MapReduce frame-
work [20]. Like OpenMP and MPI, MapReduce provides an abstraction,
a means to distribute computation without burdening the programmer
with the details of distributed computing; however, the level of granu-
larity is different [2]. These frameworks are mostly designed to deal with
processor-intensive problems and have only rudimentary support for deal-
ing with very large amounts of input data. The strength of MapReduce is
data-intensive distributed parallel processing on a massive scale [1]. The
potential of combining a data-intensive cloud-based approach with the
compute-intensive GPGPU paradigm for sophisticated, large-scale natu-
ral language processing is enormous.

3 Quantum approaches in language processing

Metaphors of quantum theory in linguistic applications arose over the last

decade [21–26]. The vector space model of information retrieval was first
adopted largely because it allowed for a naturally continuous ‘relevance
score’ by using the cosine dissimilarity, as opposed a mere binary decision
between relevant and irrelevant documents. In a similar fashion quantum
mechanics yields a continuous probability that a particular event will be
observed, a feature making it useful to reflect on possible similarities with
natural language [22]. Moreover, it appears likely that quantum interac-
tion would be of a type where the context of the interaction itself must be
incorporated into the model. For example, a measurement in a quantum-
scale system will have an impact on the result. If the system is displaying
contextual behaviour such as natural languages, then a quantum approach
often incorporates this behaviour very naturally [27].
Quantum phenomena in languages may be present at different levels.
At sub-word level, terms and documents can be regarded as linear com-
binations of their semantic features [28], which can account for semantic
priming [29].
At word level, a word in semantic space may be likened to a quantum
particle. In the absence of context it is in a superposed state, it is a
collection of all the possible meanings of the word: ρ = p1 ρ1 + . . . + pm ρm ,
where ρ is the word in the semantic space as a density matrix, and each
i is a basis state representing one of the m senses of the word and the
3
http://www.hoopoe-cloud.com/
4
http://aws.typepad.com/aws/2010/11/new-ec2-instance-type-the-cluster-gpu-
instance.html
probabilities pi sum to unity. Encountering the word in context, however,
gives rise to a ‘collapse’ of potential meanings onto an actual one. The
context is modelled a projection operator which is applied to a given
density matrix corresponding to the state of a word meaning resulting in
its ‘collapse’ [30].
Turning to combinations of words, at least two approaches offer so-
lutions. One uses the operator algebra of quantum theory to construct a
‘semantic calculus’ [24, 31]. The other approach encodes word order re-
lying on random indexing [32, 33], using either permutation [34, 35] or
circular convolution [29, 35]. The order can also be encoded by tensor
product [23, 36].
Using different units of analysis, quantum approaches find their way
to applications, most notably:
– Information retrieval: Vector space logic and quantum logic (Neumann
algebra) are very similar [24]. In particular, negation has been inves-
tigated in depth in [22]. These models may allow new types of queries
and also inference [37].
– Memory models: Two schools of memory models are relevant to quan-
tum theory: matrix memory [38, 39, 30], and convolution-correlation
memory (holographic-like) [40, 41, 29]. Matrix models are not directly
related to QT, but there can be a connection through Heisenberg’s
matrix mechanics, which was the first complete and correct definition
of quantum mechanics. It is equivalent to the Schrödinger wave for-
mulation of quantum mechanics, and is the basis of Dirac’s bra-ket
notation for the wave function. Matrix models can incorporate hierar-
chical sentence and paragraph representation [42], bridging distribu-
tional and symbolic approaches [43], encode location [44], or include
sense in a term-space approach [45, 46]. The other approach, convo-
lution memory is particularly useful to encode syntactic information
[47, 35].
– Semantic regions: Regions meant to solve the problem to be able to
say that apple is a kind of fruit (apple is part of the fruit region), as
opposed to modelling that apples and fruit have something to do with
one another. Separating hyperspaces may define a semantic region
[31]. As an alternative, [48] measures the distance between subspaces
spanned by documents by projecting them into one another.
– Spectral theory in mathematics is key to the success of as diverse
application domains as quantum mechanics and methods of latent se-
mantic analysis (LSA, [49]) for language processing. In other words,
both quantum mechanics and LSA rely on eigenvalue decomposition
 for the localization of their respective entities in observation space.
This neglected fact, supported by a high number of papers in dif-
ferent disciplines describing the dynamic behaviour of documents and
their index terms over time, points at some implicit “energy” inherent
in them and in need of quantification. Prominently, theories of word
meaning (contextual [50, 51] and referential [52, 53]), currently used
in applications trying to capture and exploit semantic content, fall
back on quantities of qualities, but quite possibly miss the underlying
framework. LSA is just one spectral approach in language represen-
tation: [30] demonstrate the quantum collapse of meaning using the
hyperspace analogue to language (HAL, [54]).

4 Methods and planned outcomes

With the above plethora of approaches available for testing, the funda-
mental task of SQUALAR is bridging scalable linear algebra and nu-
merical methods that are widely used in scientific computing with the
emerging theories in quantum interaction to enable practical, real-world
language technology applications.
The hardware and basic software infrastructure is what we described
in section 2: a distributed system consisting of heterogeneous nodes which
combine multicore CPUs and GPUs (top part of Figure 1). Since hardware
virtualization is already at consumer level, the distributed system can be
either a privately owned cluster or grid, or a high-performance computing
cloud provided by a third-party.
Without going into details, algorithms in linear algebra are the most
obvious candidates for acceleration on graphics hardware (middle part of
Figure 1, left). Vector space models of semantics can be implemented by
accelerated BLAS libraries [7, 9], including operator algebra for semantic
inference [22, 24]. Matrix decompositions and dimension reduction that
also play an important role in understanding semantics are currently lim-
ited to matrices of limited sizes [10]. Convolution, which plays an im-
portant part in encoding term positions [29, 35], can be mapped to the
frequency domain by Fourier transformation, where the operation simpli-
fies to a simple multiplication. Fast Fourier transformation on GPUs is
a classical area for acceleration [55]. More complex examples in acceler-
ated quantum methods [56, 57] and related visualization [58] are awaiting
appropriate metaphors in language processing.
Approaching from existing language processing algorithms, if a suf-
ficient metaphor cannot be found or if the method does not lend itself
 General-purpose
Parallel Distributed
computing on
programming on systems: grids
graphics
multicore CPUs and clouds
processing units

Computing environment

Scalable and Low-level parallel

Existing language
parallel environments
processing
algorithms in (OpenMP, CUDA,
systems
quantum theory etc.)

Supporting Technologies

Novel language Scalable Further high-level

processing based language applications:
on quantum engineering in digital libraries,
theory clouds text mining, etc.

Applications

Fig. 1. An overview of the SQUALAR framework

easily to any of the methods described above, lower level libraries can be
used for developing multithreaded, GPU-based implementations (middle
part of Figure 1, right and middle).
If we focus on a single computer, we will be able to perform oper-
ations several folds faster, gaining new insights on language technology
(bottom part of Figure 1, left). By providing a high-level load balancing
mechanism, the potential of compute and data-intensive processing can
be released in a distributed environment for web-scale applications (bot-
tom part of Figure 1, middle). Some machine learning algorithms, such
as support vector machines, have already been adopted to graphics hard-
ware [59]. Combining these with the above, we gain powerful text mining
applications (bottom part of Figure 1, right). Since Information Retrieval
has already began experimenting with a wide range of quantum theory
based metaphors, this field has the most to benefit.

5 Conclusion

Whether language to some extent shares a conceptual framework with

quantum mechanics, and if thereby some linguistic phenomena could be
eventually modelled on physical ones, is a research question of interest to
SQUALAR. We trust that by better mastering the match between quan-
tum algorithms and GPU computing, web-scale applications will become
feasible.
The fundamental tasks and challenges of the project are the following:
– Rephrasing natural language processing and text mining algorithms
in quantum domain to use compute-intensive heterogeneous program-
ming model;
– Data and compute-intensive distributed and cloud computing appli-
cations with heterogeneous hardware;
– Performance evaluation of heterogeneous hardware for natural lan-
guage processing tasks;
– Trade-offs of using scalable quantum models in language engineering;
– Exploiting heterogeneous architectures to accelerate sophisticated lan-
guage processing.

6 Acknowledgement

We would like to thank Lance de Vine (Queensland University of Tech-

nology) for discussions related to ideas presented in this paper. This work
was also supported by Amazon Web Services.

References
1. Dean, J., Ghemawat, S.: MapReduce: Simplified data processing on large clusters.
In: Proceedings of OSDI-04, 6th International Symposium on Operating Systems
Design & Implementation, San Francisco, CA, USA, ACM Press, New York, NY,
USA (2004)
2. Lin, J., Dyer, C.: Data-Intensive Text Processing with MapReduce. Morgan &
Claypool (2010)
 3. Cavanagh, J., Potok, T., Cui, X.: Parallel latent semantic analysis using a graphics
processing unit. In: Proceedings of GECCO-09, 11th Annual Conference Compan-
ion on Genetic and Evolutionary Computation Conference: Late Breaking Papers,
Montreal, QC, Canada, ACM Press, New York, NY, USA (2009) 2505–2510
4. Ding, S., He, J., Yan, H., Suel, T.: Using graphics processors for high performance
IR query processing. In: Proceedings of WWW-09, 18th International Conference
on World Wide Web, Spain, Madrid, ACM Press, New York, NY, USA (2009)
421–430
5. Zhang, Y., Mueller, F., Cui, X., Potok, T.: Large-scale multi-dimensional docu-
ment clustering on GPU clusters. In: Proceedings of IDPDS-10, 24th International
Parallel and Distributed Computing Symposium, Atlanta, GA, USA, IEEE Com-
puter Society (2010)
6. Byna, S., Meng, J., Raghunathan, A., Chakradhar, S., Cadambi, S.: Best-effort
semantic document search on GPUs. In: Proceedings of GPGPU-10, 3rd Workshop
on General-Purpose Computation on Graphics Processing Units, New York, NY,
USA, ACM (2010) 86–93
7. Krüger, J., Westermann, R.: Linear algebra operators for GPU implementation
of numerical algorithms. In: Proceedings of SIGGRAPH-05, 32nd International
Conference on Computer Graphics and Interactive Techniques, Los Angeles, CA,
USA, ACM Press, New York, NY, USA (2005) 234–242
8. Galoppo, N., Govindaraju, N., Henson, M., Bondhugula, V., Larsen, S., Manocha,
D.: Efficient numerical algorithms on graphics hardware. In: Proceedings of EDGE-
06, Workshop on Edge Computing Using New Commodity Architectures, Chapel
Hill, NC, USA (2006)
9. Barrachina, S., Castillo, M., Igual, F., Mayo, R., Quintana-Orti, E.: Evaluation
and tuning of the level 3 CUBLAS for graphics processors. In: Proceedings of
IPDPS-08, 22nd International Symposium on Parallel and Distributed Processing,
Miami, FL, USA, IEEE (2008) 1–8
10. Lahabar, S., Narayanan, P.: Singular value decomposition on GPU using CUDA.
In: Proceedings of IPDPS-09, 23rd International Symposium on Parallel and Dis-
tributed Processing, Rome, Italy, IEEE (2009)
11. Brodtkorb, A., Dyken, C., Hagen, T., Hjelmervik, J., Storaasli, O.: State-of-the-art
in heterogeneous computing. Scientific Programming 18(1) (2010) 1–33
12. Kirk, D., Hwu, W.: Programming massively parallel processors: A hands-on ap-
proach. (2009)
13. Jiménez, V., Vilanova, L., Gelado, I., Gil, M., Fursin, G., Navarro, N.: Predic-
tive runtime code scheduling for heterogeneous architectures. High Performance
Embedded Architectures and Compilers (2009) 19–33
14. Lee, S., Min, S., Eigenmann, R.: OpenMP to GPGPU: a compiler framework
for automatic translation and optimization. In: Proceedings of PPOPP-09, 14th
Symposium on Principles and Practice of Parallel Programming, ACM Press, New
York, NY, USA (2009) 101–110
15. Luk, C., Hong, S., Kim, H.: Qilin: Exploiting parallelism on heterogeneous multi-
processors with adaptive mapping. In: MICRO-42, 42nd Annual IEEE/ACM In-
ternational Symposium on Microarchitecture, New York, NY, USA, IEEE (2009)
45–55
16. Phillips, J., Stone, J., Schulten, K.: Adapting a message-driven parallel applica-
tion to GPU-accelerated clusters. In: Proceedings of SC-08, 21st Conference on
Supercomputing, Austin, TX, USA, IEEE Press (2008) 1–9
17. Kuhn, B., Petersen, P., O’Toole, E.: OpenMP versus threading in C/C++. Con-
currency: Practice and Experience 12(12) (2000) 1165–1176
18. Koop, M., Sur, S., Gao, Q., Panda, D.: High performance MPI design using unre-
liable datagram for ultra-scale InfiniBand clusters. In: Proceedings of ISC-06, 21st
Annual International Conference on Supercomputing, Dresden, Germany, ACM
(2006) 180–189
19. : NVida Compute Unified Device Architecture Best Practices Guide 3.2 (2010)
20. Shirahata, K., Sato, H., Matsuoka, S.: Hybrid map task scheduling on GPU-based
heterogeneous clusters. In: Proceedings of CloudCom-10, The 2nd International
Conference on Cloud Computing, Indianapolis, IN, USA (2010)
21. Aerts, D., Aerts, S., Broekaert, J., Gabora, L.: The violation of bell inequalities
in the macroworld. Foundations of Physics 30(9) (2000) 1387–1414
22. Widdows, D., Peters, S.: Word vectors and quantum logic: Experiments with
negation and disjunction. In: Proceedings of MoL-03, 8th Mathematics of Language
Conference. Volume 8., Bloomington, IN, USA (2003) 141–54
23. Aerts, D., Czachor, M.: Quantum aspects of semantic analysis and symbolic ar-
tificial intelligence. Journal of Physics A: Mathematical and General 37 (2004)
L123–L132
24. van Rijsbergen, C.J.: The Geometry of Information Retrieval. Cambridge Univer-
sity Press, New York, NY, USA (2004)
25. Widdows, D.: Geometry and meaning. (2004)
26. Bruza, P., Widdows, D., Woods, J.: A quantum logic of down below. In Engesser,
K., Gabbay, D., Lehmann, D., eds.: Handbook of Quantum Logic and Quantum
Structures. Volume 2. Elsevier (2009)
27. Kitto, K.: Why quantum theory? In: Proceedings of QI-08, 2nd International
Symposium on Quantum Interaction, Oxford, UK, Springer-Verlag (2008) 11–18
28. Lyons, J.: Semantics. Cambridge University Press, New York, NY, USA (1977)
29. Jones, M., Mewhort, D.: Representing word meaning and order information in a
composite holographic lexicon. Psychological review 114(1) (2007) 1–37
30. Bruza, P., Woods, J.: Quantum collapse in semantic space: interpreting natural
language argumentation. In: Proceedings of QI-08, 2nd International Symposium
on Quantum Interaction, Oxford, UK, College Publications (2008)
31. Widdows, D.: Semantic vector products: Some initial investigations. In: Proceed-
ings of QI-08, 2nd International Symposium on Quantum Interaction, Oxford, UK,
College Publications (2008)
32. Kanerva, P., Kristofersson, J., Holst, A.: Random indexing of text samples for
latent semantic analysis. In: Proceedings of CogSci-00, 22nd Annual Conference
of the Cognitive Science Society. Volume 1036., Philadelphia, PA, USA (2000)
33. Sahlgren, M.: An introduction to random indexing. In: Proceedings of TKE-05,
Methods and Applications of Semantic Indexing Workshop at the 7th International
Conference on Terminology and Knowledge Engineering, Copenhagen, Denmark,
Citeseer (2005)
34. Sahlgren, M., Holst, A., Kanerva, P.: Permutations as a means to encode order in
word space. In: Proceedings of CogSci-08, 30th Annual Meeting of the Cognitive
Science Society, Washington, DC, USA (2008)
35. De Vine, L., Bruza, P.: Semantic oscillations: Encoding context and structure in
complex valued holographic vectors. In: Proceedings of QI-10, 4th Symposium on
Quantum Informatics for Cognitive, Social, and Semantic Processes, Arlington,
VA, USA (2010) 11–13
36. Mitchell, J., Lapata, M.: Vector-based models of semantic composition. In: Pro-
ceedings of ACL-08, 46th Annual Meeting of the Association for Computational
Linguistics, Columbus, Ohio, ACL, Morristown, NJ, USA (2008) 236–244
37. Song, D., Lalmas, M., van Rijsbergen, C., Frommholz, I., Piwowarski, B., Wang,
J., Zhang, P., Zuccon, G., Bruza, P., Arafat, S., et al.: How quantum theory is
developing the field of Information Retrieval. In: Proceedings of QI-10, 4th Sym-
posium on Quantum Informatics for Cognitive, Social, and Semantic Processes,
Arlington, VA, USA (2010) 105–108
38. Humphreys, M., Bain, J., Pike, R.: Different ways to cue a coherent memory sys-
tem: A theory for episodic, semantic, and procedural tasks. Psychological Review
96(2) (1989) 208–233
39. Wiles, J., Halford, G., Stewart, J., Humphreys, M., Bain, J., Wilson, W.: Tensor
models: A creative basis for memory retrieval and analogical mapping. In Dartnall,
T., ed.: Artificial intelligence and creativity. Kluwer Academic (1994) 145–159
40. Plate, T.: Holographic reduced representations: Convolution algebra for composi-
tional distributed representations. In: Proceedings of IJCAI-91, 12th International
Joint Conference on Artificial Intelligence, Syndey, Australia, Citeseer (1991) 30–
35
41. Plate, T.: Holographic reduced representations. IEEE Transactions on Neural
Networks 6(3) (1995) 623–641
42. Antonellis, I., Gallopoulos, E.: Exploring term-document matrices from matrix
models in text mining. In: Proceedings of SDM-06, Text Mining Workshop in
conjuction with the 6th SIAM International Conference on Data Mining, Bethesda,
MD, USA (2006)
43. Rudolph, S., Giesbrecht, E.: Compositional matrix-space models of language. In:
Proceedings of ACL-10, 48th Annual Meeting of the Association for Computational
Linguistics, Uppsala, Sweden, Association for Computational Linguistics (2010)
907–916
44. Rölleke, T., Tsikrika, T., Kazai, G.: A general matrix framework for modelling
information retrieval. Information Processing & Management 42(1) (2006) 4–30
45. Swen, B.: A sense matrix model for information retrieval. Technical report, Tech-
nical report TR-2004-2 of ICL-PK (2004)
46. Novakovitch, D., Bruza, P., Sitbon, L.: Inducing shades of meaning by matrix
methods: a first step towards thematic analysis of opinion. In: Proceedings of
SEMAPRO-09, 3rd International Conference on Advances in Semantic Processing,
Sliema, Malta, IEEE (2009) 86–91
47. Jones, M., Kintsch, W., Mewhort, D.: High-dimensional semantic space accounts
of priming. Journal of memory and language 55(4) (2006) 534–552
48. Zuccon, G., Azzopardi, L., Rijsbergen, C.: Semantic spaces: Measuring the distance
between different subspaces. In: Proceedings of QI-09, 3rd International Sympo-
sium on Quantum Interaction, Saarbruecken, Germany, Springer-Verlag (2009)
225–236
49. Deerwester, S., Dumais, S., Furnas, G., Landauer, T., Harshman, R.: Indexing by
latent semantic analysis. Journal of the American Society for Information Science
41(6) (1990) 391–407
50. Wittgenstein, L.: Philosophical Investigations. Blackwell Publishing, Oxford, UK
(1967)
51. Harris, Z.: Distributional structure. In Harris, Z., ed.: Papers in structural and
transformational Linguistics. Formal Linguistics. Humanities Press, New York,
NY, USA (1970) 775–794
52. Peirce, C.: Logic as semiotic: The theory of signs. In Peirce, C., Buchler, J., eds.:
Philosophical Writings of Peirce. Dover Publications (1955) 98–119
53. Frege, G.: Sense and reference. The philosophical review 57(3) (1948) 209–230
54. Lund, K., Burgess, C.: Producing high-dimensional semantic spaces from lexical
co-occurrence. Behavior Research Methods Instruments and Computers 28 (1996)
203–208
55. Govindaraju, N., Lloyd, B., Dotsenko, Y., Smith, B., Manferdelli, J.: High per-
formance discrete Fourier transforms on graphics processors. In: Proceedings of
SC-08, 21st Conference on Supercomputing, Austin, TX, USA, IEEE (2008)
56. Ufimtsev, I., Martı́nez, T.: Graphical processing units for quantum chemistry.
Computing in Science & Engineering 10(6) (2008) 26–34
57. Watson, M., Olivares-Amaya, R., Edgar, R., Aspuru-Guzik, A.: Accelerating corre-
lated quantum chemistry calculations using graphical processing units. Computing
in Science & Engineering 12(4) (2010) 40–51
58. Stone, J., Saam, J., Hardy, D., Vandivort, K., Hwu, W., Schulten, K.: High per-
formance computation and interactive display of molecular orbitals on GPUs and
multi-core CPUs. In: Proceedings of GPGPU-09, 2nd Workshop on General Pur-
pose Processing on Graphics Processing Units, Washington, DC, USA, ACM (2009)
9–18
59. Catanzaro, B., Sundaram, N., Keutzer, K.: Fast support vector machine training
and classification on graphics processors. In McCallum, A., Roweis, S., eds.: Pro-
ceedings of ICML-08, 25th Annual International Conference on Machine Learning,
Helsinki, Finland, Omnipress (2008) 104–111