Heriot-Watt University
ROS: The Research Output Service. Heriot-Watt University EdinburghNot a member yet
4689 research outputs found
Sort by
Semantic consistency in text generation
Get PDFAutomatic input-grounded text generation tasks process input texts and generate human-understandable natural language text for the processed information. The development
of neural sequence-to-sequence (seq2seq) models, which are usually trained in an end-to-end fashion, pushed the frontier of the performance on text generation tasks expeditiously. However, they are claimed to be defective in semantic consistency w.r.t. their
corresponding input texts. Also, not only the models are to blame. The corpora themselves always include examples whose output is semantically inconsistent to its input.
Any model that is agnostic to such data divergence issues will be prone to semantic inconsistency. Meanwhile, the most widely-used overlap-based evaluation metrics
comparing the generated texts to their corresponding references do not evaluate the
input-output semantic consistency explicitly, which makes this problem hard to detect.
In this thesis, we focus on studying semantic consistency in three automatic text
generation scenarios: Data-to-text Generation, Single Document Abstractive Summarization, and Chit-chat Dialogue Generation, by seeking for the answers to the following research questions: (1) how to define input-output semantic consistency in different
text generation tasks? (2) how to quantitatively evaluate the input-output semantic
consistency? (3) how to achieve better semantic consistency in individual tasks?
We systematically define the semantic inconsistency phenomena in these three
tasks as omission, intrinsic hallucination, and extrinsic hallucination. For Data-to-text Generation, we jointly learn a sentence planner that tightly controls which part
of input source gets generated in what sequence, with a neural seq2seq text generator,
to decrease all three types of semantic inconsistency in model-generated texts. The
evaluation results confirm that the texts generated by our model contain much less
omissions while maintaining low level of extrinsic hallucinations without sacrificing
fluency compared to seq2seq models. For Single Document Abstractive Summarization, we reduce the level of extrinsic hallucinations in training data by automatically
introducing assisting articles to each document-summary instance to provide the supplemental world-knowledge that is present in the summary but missing from the doc ument. With the help of a novel metric, we show that seq2seq models trained with as sisting articles demonstrate less extrinsic hallucinations than the ones trained without
them. For Chit-chat Dialogue Generation, by filtering out the omitted and hallucinated
examples from training set using a newly introduced evaluation metric, and encoding
it into the neural seq2seq response generation models as a control factor, we diminish
the level of omissions and extrinsic hallucinations in the generated dialogue responses
Multidisciplinary use of LCA and empirical data to improve the environmental sustainability and industrial viability of Kraft pulp mill
Get PDFAbout 4% of the world energy is used to produce paper and it takes 300 -2600 m3
of
water to produce 1 tonne of A4 sheet (Van Oel and Hoekstra 2012). Hence, the key aim
of this research is to use a multi-disciplinary approach involving life cycle assessment
(LCA) and empirical methodology to improve the environmental sustainability and
industrial viability of the conventional Kraft pulp mill. The life cycle assessment
methodology (LCA) used to study the Kraft pulp mill identified energy as the elementary
flow with most impact closely followed by water. The chemical recovery unit was
identified as the process unit of highest concern followed by the washing and screening
process unit. The LCA study suggested that using smaller wood particle size during
cooking would reduce energy consumption and emission to the air. Furthermore, the use
of a re-usable liquid as part of the washing process could reduce water consumption and
enable the recovery of extractable chemicals in the form of a green crude. The empirical
data showed that reducing the particle size from 8mm currently used in the conventional
Kraft process to < 2mm could save up to 20% energy during pulp production. Replacing
part of the water during washing with recoverable hexane could reduce water
consumption during the pulp washing close to 37%. In addition, using hexane as a
replacement for water enabled approximately 19% green crude recovery from particle
size <2mm cooked in 30mins based on initial dry wood weight. The recovery of green
crude could transform the Kraft pulp mill into a potential biorefinery. The LCA outcome
indicated that the impact of energy generation on climatic change is equivalent to about
713 kgCO2eq for a conventional Kraft mill and this was reduced by 20% to about 570
kgCO2eq for the proposed Kraft biorefinery. In conclusion, the combination of LCA and
empirical methodology showed how a Kraft pulp mill could be transferred into a more
sustainable and industrially viable biorefinery resulting in lower environmental impact
and producing green chemicals from wood
The traditional marketplace : creating memorable, engaging and authentic cultural consumption experience
Get PDFCreating memorable experiences and offering unique services have become pivotal in the
tourism industry in order to enhance competitiveness and sustainable success. The extant
literature has recognised the fundamental change in contemporary consumers’ behaviour.
Hence, experiences have been increasingly evolving in the tourism industry. In exploring
the current developments and future directions in the tourism literature, this thesis offers
a theoretically rich and well-validated conceptual model, particularly in a culturally
specific field. In doing so, this thesis sheds light on three stages of the cultural
consumption experience by integrating the consumer-based model of authenticity with the
value creation theory.
More broadly, this thesis integrated cultural motivation, sociability, host sincerity, object-based and existential authenticity, tourist engagement, perceived value, and memorable
tourism experience into the consumer-based model of authenticity and embedded them
into the larger perspective of service logic in a Turkish heritage context. In particular, the
conceptual model proposed that sincere host-guest interactions, perceived authenticity and
tourists’ engagement are influenced by cultural motivation and sociability, impacting
upon perceived value and memorable tourism experience. To provide a holistic
understanding of interrelationships between three stages of the consumption experience,
data was collected in two sequential phases. The first phase consisted of a qualitative
research approach that involves semi-structured interviews, personal observations, and
field notes. In Phase II, data was collected through a questionnaire to provide a wider
perspective and analysed using Partial Least Squares Structural Equation Modeling (PLS-SEM). Qualitative findings contributed to the factors shaping overall heritage
experiences. The findings of the data supported the conceptual model in determining the
pre/on-site/post phases of the heritage consumption experience. Following this,
quantitative results show that the hitherto separate concepts make a substantial
contribution to the consumer-based model of authenticity. In particular, the quantitative
data demonstrate the interrelationship between these factors, offering complementary
ways of understanding the phenomenon within the non-Western service industry. More
broadly, this thesis identifies components and issues that are significant for tourists
visiting heritage destinations and attractions. The findings of this thesis could have
practical implications for planners, destination managers and tourism policy-makers to
develop competitive advantage and sustainable success
Behind the chain of obscurity : methodologies for cryptocurrency forensic analysis
Get PDFBitcoin and alternative cryptocurrencies are decentralised digital currencies that
allow users to anonymously exchange money without requiring the presence of a
trusted third party. The privacy components of cryptocurrency can facilitate illegal
activities and present new challenges for cybercrime forensic analysis. Tackling such
challenges motivates new research interest in cryptocurrency tracking. This thesis
explores and proposes novel methodologies and improvements to existing cryptocurrency tracking and analysis methodologies.
Our first contribution explores the most commonly used cryptocurrency tracking methodology named Taint Analysis and investigates a potential improvement to
the methodology’s tracking precision with the implementation of address profiling.
We also introduce two context-based taint analysis strategies and hypothesise behaviours related to the tracked Bitcoins context to create a set of evaluation metrics.
We conducted an experiment using sample data from known illegal Bitcoin cases to
illustrate and evaluate the methodology, and the results reveal distinct transaction
behaviours in tracking between the results with and without address profiling for all
of the metrics. Our second contribution proposes a cryptocurrency tracking methodology named Address Taint Analysis that is capable of tracking zero-taint coins created by Privacy-Enhancing Technologies (PETs) called centralised mixer services,
which are untrackable with taint analysis tracking. Our results indicate that our proposed address taint analysis can trace the zero-taint Bitcoins from nine well-known
mixer services back to the original Bitcoins. Our third contribution investigates and
proposes a detection method for Wasabi Wallet’s CoinJoin transactions, which is one
of the most recent well-known PET services. Our fourth contribution introduces an
open-source library for cryptocurrency tracking and analysis named, TaintedTX ,
that we utilised to perform our research experiments. The library supports a variety of taint analysis strategies that users can select to track targeted transactions
or addresses. The library also includes a compilation of utility functions for address
clustering, website scraping, transaction and address classifications
The computational modelling of the structure and reactivity of Rhodium σ-Alkane complexes in the solid-state
Get PDFA series of solid-state [Rh(Cy2PCH2CH2PCy2)(alkane)][BArF
4] (alkane = C3-C6 linear,
isobutane, 2-Methylbutane, 3-Methylpentane, Cyclohexane and NBA) σ-alkane
complexes were optimised with periodic DFT and investigated with a range of electronic
structure analysis techniques including QTAIM, NBO and NCI plots. Intramolecular
interactions of the alkane with the Rh centre show a range of hapticities from η1
through
to η2 and some η2
interactions supported by varying degrees of interactions with geminal
C–H bonds. Intermolecular interactions of the Rh cation with its environment were also
explored where the importance of the solid-state environment in alkane binding is shown.
However, there is a lack of correlation with the experimental lifetimes of the complexes.
Both dehydrogenation and H/D exchange of [Rh(Cy2PCH2CH2PCy2)(Isobutane)][BArF
4]
was investigated using an isolated cation model and a solid-state model using periodic
DFT. For dehydrogenation two pathways were identified differing by initial 1º
or 3º
C–
H activation. The isolated cation model showed both pathways to have comparable
barriers however, in the solid-state, the two pathways become energetically distinct.
Hirshfeld surfaces showed destabilising contacts between isobutene and an aryl ring of a
BArF
4
-
anion at the rate determining H2 reductive coupling transition state in the
disfavoured 3º
C–H activation pathway. Another case where the solid-state environment
controlled reactivity was the H/D exchange of both the 1º
and 3º
C–H bond participating
in a σ-interaction. Here different pathways were favoured in the two models.
Finally, the electronic structure of anagostic interactions in Rh((DPEphos-o-R)(NBD)+
(R = H, Me, OMe, and iPr) were studied showing the effect of changing the ortho R group.
The largest substituent, iPr, gives the shortest and strongest anagostic interactions by
QTAIM and NBO analysis and in general a correlation between H...Rh distance and
interaction strength was seen. The anagostic interactions were also found to be
predominantly covalent in nature. Computed 1H NMR chemical shifts show a downfield
shift experienced by the anagostic proton, as seen experimentally, however the extent of
which does not correlate with the strength of the anagostic interaction. Plots of the
Laplacian visually rationalised the NMR data, and the occurrence of the downfield shift,
with H...Rh bond paths lying over an area of charge accumulation in all cases
Machine learning and anomaly detection for insider threat detection
Get PDFWith the ever-growing online presence of businesses technology is ubiquitous in our
daily life. Whether through social media, online shops, working remotely, or talking to
a smart speaker to order dinner, cyber security will be only more important. While
external intrusions to information systems still constitute the vast majority of cyber
hazards, there is one type of threat that has been growing over the recent years and
drawing more attention, both from the industrial and academic worlds: a threat from
within.
Insider threat is a case of an employee or anyone with security access to an organisation
negatively affecting the company: through fraud, spreading misinformation and chaos
among other employees, or theft. Insiders are already past the security perimeter of an
organisation, giving them advantage and often easy access to sensitive files. That makes
insider threat a very dangerous and costly type of cyber crime.
Insider threat detection is a growing niche of cyber security. It faces many interesting
challenges, including problems with finding real data for experiments, the high point difficulty of anomaly detection, and difficulties encountered when trying to model attacks.
These problems have played a big role in shaping the current landscape of insider threat
detection.
In this thesis we describe our work with various machine learning and anomaly detection
models for insider threat detection. We describe the current state of the art and the
background of insider threat detection. We describe the datasets we use, both real
and synthetic, and the attacks we generate and insert into benign data to create data
for additional verification of our models. We outline three guidelines of good conduct
regarding experiment protocol, motivated by literature and experience gathered during
this research: use varied and multiple quality metrics to describe results; design and
create models without prior knowledge of data; report on all known limitations and
assumptions.
Out of the models we developed, the ensembles and local outlier factor (LOF) had consistently the best results out of the entire suite, and were comparable to state of the art
research. Our focus on multiple metrics helped us pinpoint a situation where models
were performing worse than the most popular metric, area under receiver operating characteristic curve (AUC), would suggest. Our method for generating and inserting attacks
was a valuable tool of verification of model results. Finally, for future research in the
field we suggest improving heterogenous ensembles, using deep learning and especially
autoencoders for anomaly detection, and further work on generating data
Impact investments in microfinace in India (2010-2016)
Get PDFA key tenet in the definition of impact investing is that the investment must be made with the
intention of making a social impact. Yet, there is little literature that addresses the constitution
of such intentionality. Without clarity, the prospect of credibly implementing intention is bleak,
which renders the whole definition of impact investing weak and practically useless. By
specifically studying the intention of making investments in microfinance institutions (MFIs)
in India, this research addresses the question of whether such investments qualify as ‘impact
investments’ or not.
Towards the objective, the Reasoned Action Approach (RAA) is applied to elicit a causal
analysis to explain the behaviour of MFIs. The investors’ intention and influence on the
behaviour of MFIs is germane. 67 semi-structured interviews and questionnaires were
administered to five strata of strategic decision-makers of influencers viz., asset owners,
investment managers, banks and service providers. Social impact is characterised by corporate
governance, business behaviour, human resource and community involvement. Financial
returns are analysed using the Du Pont formula - Net Profit Margin, Asset Turnover Ratio and
Financial Leverage. Taken together, these seven characteristic behaviours unveil the
behavioural predilection of each stratum of respondents. The study factors that the behaviour
of MFIs is determined not only by the willingness but also by the ability of investees to manage
social impact alongside financial returns. The Analytical Hierarchy Process (AHP) is deployed
to disentangle complex decisions that are faced by MFIs. The AHP decrypts respondents’
choices, the ease with which such choices are made and unravels that there is a disposition to
select financial return vis-à-vis social impact. MFI promoters and donors are brought in for
validation.
The key findings are: (i) the intention of social impact was limited, and the focus was primarily
on financial returns; (ii) the relationship between social impact and financial return is complex
and inextricably inter-dependent; therefore, ‘standards’ of evaluation will have to combine
social impact and financial return (iii) both impact investors and MFIs, show a high degree of
isomorphism, on both intention and behaviour of social impact. These findings have direct
applications to (a) public policy on impact investments in India and elsewhere, (b) social
enterprises’ strategy, (c) corporate policies of impact investors and (d) bridge the gap between
theoretical definitions and practical applications. The literature section can also be applied for
studies on behavioural finance, institutional structures and regulation of impact investments
Geometry and BPS spectra of 4d N = 2 theories
Get PDFIn this thesis we present work done on N = 2 theories of class S with a focus on
two projects. N = 2 theories have a large enough amount of supersymmetry to
allow for exact results, while also being less restricted in terms of field contact as
theories with more supercharges. The theories of class S can be obtained from string
theory constructions, which assign a Riemann surface C to each theory. Physical
phenomena and computations are then related to the geometry of C, in particular
to the space of flat connections on C.
One of the projects is about computing the effective twisted superpotential Weff,
using a geometric method based on a proposal by Nekrasov, Rosly, and Shatashvili.
We give a detailed explanation of the geometric recipe to compute Weff by first constructing Darboux coordinates for the moduli space of flat connections of the corresponding UV-surface C, and then evaluating these coordinates on oper connections
by solving the differential equations on C, which can be found from quantization of
the Seiberg-Witten curve Σ → C. We present results for the pure SU(2) theory in
its weak coupling limit, the AD2 theory, and the Minahan-Nemeschansky E6 theory.
The other project is about the BPS spectrum of the SU(2) N = 2∗
theory.
We start with the theory on specific walls Ei
in the Coulomb branch, on which the
spectrum is known. We reproduce the results on these with quiver methods and
then examine the perturbation from the wall, first by again using quiver methods
and then by a python algorithm which implements the Kontsevich-Soibelmann wall
crossing formula. We find very intricate behavior including highly unexpected wall
crossings with negative pairing. We show a way to interpret these crossings as the
reverse of the usual pairing 2 wall crossing formula, involving a collection of a priori
unrelated states
Synthesis and catalytic applications of titanium amine bis(phenolate) complexes
Get PDFMany catalytic processes rely on the use of transition metal complexes, frequently utilising
precious metals, such as ruthenium, iridium, palladium or platinum. While these metals exhibit
excellent performances in well-established processes, they also bear the inherent disadvantages
of high cost and low abundance. It is therefore desirable to move towards more sustainable
alternatives. In this work, molecular titanium complexes were investigated as potential
replacements for noble metals in photocatalysis and as polymerisation catalysts.
Titanium is an abundant, inexpensive metal with low toxicity. Consequently, TiO2 is used widely
in ‘low cost’ applications, for example as white pigments and as a photocatalyst, although it only
absorbs in the ultraviolet region of the electromagnetic spectrum. Ideally, a photocatalyst
should be able to utilise visible light (i.e. sunlight) to drive chemical reactions, therefore the
application of molecular, visible-light-absorbing titanium complexes as photosensitisers was
explored here. A series of mono- and bimetallic complexes with amine bis(phenol) ligands was
synthesised, characterised and subsequently trialled in the generation of singlet oxygen under
visible light irradiation (420–450 nm). The active, singlet oxygen species was trapped using -
terpinene, cyclopentadiene, 1,3-cyclohexadiene and anthracene. All complexes were found to
be active photosensitisers, however, the presence of a ‘TiO2-like’ titanium-oxo moiety was found
to be crucial for the activity of the complexes. These experiments were conducted both under
continuous flow conditions and as batch reactions. Structure-activity relationships were
investigated, giving rise to a better understanding of the key features of these complexes and
targets for further development of the photosensitisers.
Moreover, the synthesised complexes were utilised as polymerisation catalysts in the ring opening polymerisation of the cyclic ester ɛ-caprolactone. The polymeric form,
poly(ɛ-caprolactone), is biodegradable and of interest in food packaging and medical
applications. Using benzyl alcohol initiator, the complexes afforded polymers, with molecular
weights of up to 21,000 g mol-1 and reasonably narrow dispersities (1.09–1.73), suggesting
moderate to good control over the polymerisations. Two vanadium complexes and two
manganese complexes were synthesised for comparison, although it was found that the
titanium complexes typically had a more favourable balance between catalytic activity and
properties of the resulting polymer. Mechanistic aspects of these polymerisation reactions were
explored and a first indication of potential reaction mechanisms was uncovered.EPSRC fundin
The integration of heat pipe technology into photovoltaic panels to increase the operational efficiency
Get PDFThe Sultanate of Oman growth in population and infrastructure expansion in recent years
resulted in increased energy consumption. As a solution to meet the increasing energy
demand, energy management strategies and renewable energy-driven technologies are the
most viable alternatives. Among these technologies solar photovoltaics (PV) are the most
promising technology due to the strong support from the Omani government. This study
introduces a Heat Pipe Heat Exchanger (HPHE) technology as a passive cooling mechanism
to be integrated within PV terminals. The aim of this research is to increase the energy
capacity of rooftop PV modules in hot and arid climates like the Sultanate of Oman. The
performance of the existing grid photovoltaic system was benchmarked using a thermal
collector and data loggers which monitored the PV modules temperature. The experimental
investigation resulted in the establishment of the site solar irradiation of 911.11 ± 143.43
W/m2
and the Nominal Operating Cell Temperature (NOCT) of 61.4 ℃ which produced
the Peak PV Power efficiency of 54.8 %. The recorded findings of 63.8 ℃ in the NOCT had
reduced the PV Power efficiency by 2.19%. The Computational Fluid Dynamics (CFD)
modelling of the HPHE using a single independent PV panel and its analysis was made using
different methodology of investigation to specify the optimum configuration. The CFD
modelling results were used to identify that the efficient physical set up is made of PV-HPHE-DSCD (Double Sided Condenser) orientation with screen mesh wick. The optimum
configuration was made of 20 units HPHE arranged on 50 mm on centers at an angle of
inclination of 3 degree in the middle installation of the PV back surface. Water was used as
a refrigerant with a fill ratio of 65% which equates to 59 ml loaded into the evaporator section.
The results of the final stage of the experimental set up had an average PV-HPHE power
performance of 29.03 ± 0.047 % and an average power generation of 71.94 ± 2.41 W that
comprised of 23.98 % of the rated PV power capacity. The validation of the CFD model
using experimental testing was carried out by determining the error which was found to be
within the accepted range with a mass flow rate of 2.07e-05 kg/s equivalent to an average
flow rate of 8.13 e-05 m/s (Al-Mabsali et al., 2021) in the evaporator to condenser flow
direction. The significance of the research data indicates that if the heat pipe technology is
incorporated in typical outdoor conditions and the power efficiency of the device can be
improved to a maximum of 7.94% from an average power efficiency of 5.09%
Access Repository Dashboard
Do you manage ROS: The Research Output Service. Heriot-Watt University Edinburgh ? Access insider analytics, issue reports and manage access to outputs from your repository in the CORE Repository Dashboard!