Will Allen w'h'a'allenncl'ac'uk Warren Maguire w'n'maguirencl'ac'uk Hermann Moisl hermann'moislncl'a

1

Phonetic variation in Tyneside exploratory
multivariate analysis of the Newcastle Electronic
Corpus of Tyneside English
Will Allen Warren Maguire Hermann Moisl
School of English Literature, Language, and
LinguisticsUniversity of Newcastle upon Tyne
2
Introduction

• The newly-created Newcastle Electronic Corpus of
  Tyneside English (NECTE) offers an opportunity to
  study an historically-recent sample of English
  spoken in the Tyneside region of North-East
  England.

3
Introduction

• This paper gives an overview of an exploratory
  analysis of phonetic data derived from NECTE.
• The analysis was undertaken with the aim of
  generating hypotheses about the main directions
  of phonetic variation among individual speakers
  and speaker groups in the corpus, and about how
  this variation correlates with associated social
  factors.
• The discussion is in four main parts.
• The first part outlines exploratory multivariate
  analysis in general, and in particular
  hierarchical cluster analysis, the method used in
  this study.
• The second describes the NECTE phonetic data used
  in the analysis.
• The third carries out a hierarchical cluster
  analysis of a sample of that data and states some
  hypotheses based on it.
• The fourth states raises some caveats and
  indicates directions for future work.

4
1. Multivariate analysis

• The proliferation of computational technology has
  generated an explosive production of
  electronically encoded information of all kinds.
• In the face of this, traditional paper-based
  methods for search and interpretation of data
  have been overwhelmed by sheer volume, and a wide
  variety of computational methods has been
  developed in an attempt to make the deluge at
  least tractable.
• As such methods have been refined and new ones
  introduced, something over and above tractability
  has emerged new and unexpected ways of
  understanding the data.
• The fact that a computer can deal with vastly
  larger data sets than a human is an obvious
  factor, but there are two others of at least
  equal importance
• one is the ease with which data can be
  manipulated and reanalyzed in interesting ways
  without the often prohibitive labour that this
  would involve using manual techniques
• the other is the extensive scope for
  visualization that computer graphics provide.

5
1. Multivariate analysis

• These developments have clear implications for
  the analysis of large bodies of text in corpus
  linguistics.
• Effective analysis of the large electronic
  corpora now being generated will increasingly be
  tractable only by adapting the interpretative
  methods developed by the statistical, data
  mining, and related communities.
• In the present paper we are interested in one
  particular type of tool multivariate analysis.
  What is multivariate analysis?

6
1. Multivariate analysis

• Observation of nature plays a fundamental role in
  science.
• In current scientific methodology, an hypothesis
  about some phenomenon is proposed and its
  adequacy assessed using data obtained from
  observation of the domain of inquiry.
• But nature is complex, and there is no hope of
  being able to observe it exhaustively. Instead,
  particular aspects of the domain are selected for
  observation.
• Each selected aspect is represented by a
  variable, and a series of observations is
  conducted in which, at each observation, the
  values for each variable are recorded. A body of
  data is thereby built up on the basis of which an
  hypothesis can be assessed.
• One might choose to observe only one aspect, in
  which case the data set consists of a set of
  values assigned to one variable such a data set
  is referred to as univariate.
• If two values are observed, then the data set is
  bivariate, if three trivariate, and so on up to
  some arbitrary number n the general term for
  n-variable data sets is multivariate.

7
1. Multivariate analysis

• As the number of variables grows, so does the
  difficulty of understanding the data, that is, of
  conceptualizing
• the interrelationships of variables within a
  single data item how, for example, are height,
  weight, and heart rate of any given person
  interrelated?
• the interrelationships of complete data items
  how do people measured on the above variables
  compare to one another?
• Multivariate analysis is the computational use
  of mathematical and statistical tools for
  understanding such interrelationships in data.

8
1. Multivariate analysis

• Numerous techniques for multivariate analysis
  exist. They can be divided into two main
  categories which are usually referred to as
  'exploratory' and 'confirmatory'.
• Exploratory analysis aims to discover
  regularities in data which can serve as the basis
  for formulation of hypotheses about the domain
  from which the data comes. Such techniques
  emphasize intuitively accessible, usually
  graphical representations of data structure.
• Confirmatory multivariate analysis attempts to
  determine whether or not there are significant
  relationships between some number of selected
  independent variables and one or more dependent
  ones.
• These two types are complementary in that the
  first generates hypotheses about data, and the
  second tries to determine whether or not such
  hypotheses are valid. Exploratory analysis is
  naturally prior to confirmatory this discussion
  is concerned with the former.

9
1. Multivariate analysis hierarchical cluster
analysis

• Hierarchical cluster analysis is a variety of
  exploratory multivariate analysis.
• To understand how it works it is first necessary
  to understand the concept of distance between
  data points in vector space.
• Assume a domain of inquiry, say a linguistic
  corpus, which will be studied using six
  variables.
• If the six-dimensional data is to be analyzed
  using an exploratory method, it has to be
  represented mathematically.
• This is done in the form of vectors, where a
  vector is a sequence of values indexed by the
  positive integers 1, 2, 3.

10
1. Multivariate analysis hierarchical cluster
analysis

• Where the data consists of more than one case,
  which it usually does, then each case is
  represented by a vector, and the set of vectors
  is assembled into a matrix, which is a sequence
  of vectors arranged in rows and the rows are
  indexed by the positive integers 1, 2, 3 .
• In matrix M, case 2 is at row M2 and the value of
  the third variable for that case is at M2,3, that
  is, 0.1.

11
1. Multivariate analysis hierarchical cluster
analysis

• A vector space is a geometrical interpretation of
  a set of vectors
• The dimensionality n of the vectors, that is, the
  number of its elements, defines an n-dimensional
  space.
• The values in the vector define the coordinates
  of a point in that space

12
1. Multivariate analysis hierarchical cluster
analysis

• For example, a bivariate data set defines a
  2-dimensional space in which each vector
  specifies the coordinates of a point in that
  space.
• Take a data set consisting of vectors that
  specify the age and weight of some number of
  individuals. A single such vector might be v
  (36,160).
• In geometrical terms, the x or age axis is
  0..100, the y or weight axis is 0..200, and any
  vector in the data set can be plotted in the
  (x,y) space, as in

13
1. Multivariate analysis hierarchical cluster
analysis

• If more vectors are plotted in the space,
  nonrandom structure may or may not emerge,
  depending on the interrelationships of the
  real-world characteristics that the variables
  represent.
• Where there are no structured real-world
  interrelationships, the result will look
  something like the upper plot of random points.
• If there is structure, the plot might look
  something like the lower one, where two clusters
  have clearly emerged. These clusters say
  something substantive about the
  interrelationships of the represented entities.

14
1. Multivariate analysis hierarchical cluster
analysis

• Analogously, a trivariate (age, weight, height)
  vector  v (36, 160, 71) from a data set of
  length-3 vectors defines a point in 3-dimensional
  space, as in the upper plot.
• Multiple vectors representing a structured domain
  plotted in the space might look like the lower
  figure.

15
1. Multivariate analysis hierarchical cluster
analysis

• The structure of data with dimensionality higher
  than 3 cannot be directly visualized.
• How can it be represented in an intuitively
  accessible way?
• The various exploratory multivariate methods
  provide indirect visualizations.
• Hierarchical cluster analysis, in particular,
  constructs dendrograms or trees that show the
  constituency structure of clusters using relative
  distance between and among points in the
  high-dimensional data vector space.
• Distance can be understood quite literally
  distance between points A and B in the figure
  below can be measured, and it is less than the
  measured distance between A and C.

16
1. Multivariate analysis hierarchical cluster
analysis

• As an example of how dendrograms graphically
  represent the structure of higher-dimensional
  data, we use a benchmark data set that measures
  flowers on 4 variables.
• A hierarchical cluster analysis generates a tree
  in which the different lengths of the horizontal
  lines represent relativities of distance among
  data vectors in 4-dimensional space the longer
  the lines, the greater the distance.
• Knowing this, it can easily be seen that there
  are three main clusters, that is, three types of
  flower, and that each cluster has internal
  structure.

17
2. Data the corpus

• The NECTE corpus is based on two pre-existing
  corpora of audio-recorded speech, one of them
  gathered during Tyneside Linguistic Survey (TLS)
  undertaken in the late 1960s and the other
  between 1991 and 1994 for the Phonological
  Variation and Change in Contemporary Spoken
  English project, both at Newcastle University.
• NECTEs aim has been to enhance, improve access
  to, and promote the re-use of the TLS and PVC
  corpora by amalgamating them into a single,
  TEI-conformant electronic corpus.
• The result will shortly be made available to the
  research community in a variety of formats
  digitized sound, phonetic transcription, and
  standard orthographic transcription, all aligned
  and accessible on the Web.

18
2. Data the corpus

• This discussion is concerned with the TLS
  component of NECTE
• It originally consisted of 150 loosely-structured
  30-minute interviews with Tyneside informants
  that were recorded onto analog reel-to-reel
  tapes.
• As part of its research activity based on these
  recordings, the TLS produced highly detailed
  phonetic transcriptions of about 10 minutes of
  each of 64 recordings, of which 61 survive.
• These 61 transcriptions are the basis for the
  data used in this presentation.

19
2. Data the corpus

• One of the main aims of the TLS project was to
  see whether systematic phonetic variation among
  Tyneside speakers of the period could be
  significantly correlated with variation in their
  social characteristics.
• To this end they developed a methodology which
  was radical at the time and remains so today
• in contrast to the then-universal and
  still-dominant theory driven approach, where
  social and linguistic factors are pre-selected by
  the analyst,
• the TLS proposed a fundamentally empirical
  approach in which salient factors are extracted
  from the data itself and then serve as the basis
  for model construction.

20
2. Data the corpus

• To realize its research aim using its empirical
  methodology, the TLS had to compare the audio
  interviews it had collected at the phonetic level
  of representation.
• To be able to do this, the TLS phonetically
  transcribed a substantial sample of its audio
  corpus, as noted.
• The TLS invented its own transcription scheme. It
  is too complex for presentation here, but its
  main features are
• It captures variation in the distribution of
  phonetic segments across lexical environments.
• There are two levels of transcription relatively
  broad and very narrow.

21
2. Data abstraction

• The analysis reported in this talk is at the
  broad phonetic transcription level.
• It is based on comparison of phonetic profiles
  associated with each of the TLS speakers.
• A profile for any speaker S is the number of
  times S uses each of the phonetic segments
  defined by the TLS transcription scheme in his or
  her interview.

22
2. Data abstraction

• More specifically, the profile P associated with
  S is a vector having as many elements as there
  are codes such that
• Each vector element Pj represents the jth
  segment, where j is in the range 1..number of
  phonetic segments in the TLS scheme
• The value stored at Pj is an integer representing
  the number of times S uses the jth code code.
• There are 156 codes, and so a profile is a
  length-156 vector. For example

23
2. Data abstraction

• There are 61 TLS speakers, and their profiles are
  represented in matrices having 61 rows, one for
  each profile

24
2. Data preprocessing

• Prior to analysis, this matrix modified in two
  ways.
• Normalization for variation in text length
• Dimensionality reduction elimination of
  superfluous variables
• The result is a 61 x 80 length-normalized matrix,
  which is the data for the analysis that follows.

25
3. Analysis

• This section analyzes the data matrix developed
  in the preceding section.
• One particular variety of hierarchical cluster
  analysis is used squared Euclidean distance
  measure and the increase in sum of squares
  clustering algorithm, or, more simply, Wards
  method.
• The implications of this choice are discussed in
  due course.
• The aim is to get an initial indication of the
  structure of the TLS phonetic data using the
  matrix of speaker profiles normalized for length
  and dimensionality reduced as described. The
  cluster tree looks like this

26
3. Analysis

• Four main clusters emerge, labelled A-D.
• D clusters markedly against the rest, and
  comprises the Newcastle group of speakers.
• On the basis of the phonetic segment frequency
  distribution evidence, therefore, Newcastle
  speakers are strongly distinguished from
  Gateshead ones.
• Gateshead ones can be further analyzed, but for
  present purposes we stay with the Newcastle /
  Gateshead distinction

27
3. Analysis

• Knowing that there are well defined clusters is
  one thing.
• Knowing why is another what are the main
  phonetic segmental determinants of the clusters?
• Several ways of answering this question exist we
  take a graphical approach.
• The phonetic profiles for all 61 speakers were
  simultaneously plotted with the aim of visually
  identifying systematic differences between the
  Newcastle and Gateshead clusters.

28
3. Analysis

• The problem is immediately apparent too much
  detail. The level of detail was reduced as
  follows
• A significantly smaller number of the
  highest-variance variables was selected, thus
  reducing the density of information on the
  x-axis.
• The frequency vectors for all the Gateshead
  speakers 1-57 were averaged, yielding a mean
  frequency vector for these speakers.
• The same was done for the Newcastle speakers
• The mean frequency vectors were plotted against
  each other on the same graph

29
3. Analysis

• For the 30 highest-variance variables shown here,
  those for which the mean vectors differ most are
  the most important in differentiating Newcastle
  and Gateshead speakers.

30
3. Analysis

• The table on the right gives the interpretation
  of this plot
• Rank indexes the 6 largest variable-differences
  between the two clusters 1 is the largest
  difference, 2 the second largest, etc.
• Variable nr shows, for each Rank, the
  corresponding variable number in the range 1..40
  on the x-axis of the plot.
• Variable symbol identifies the phonetic symbol
  corresponding to the Variable nr.

31
3. Analysis

• The variables that are most important for
  distinguishing Gateshead from Newcastle speakers
  can be read off from the table. They are
• ? in words like standard and interview
  (localised Tyneside English often has ? here)
• closely related to this, ? in words like baker
  and china (localised Tyneside English often has
  ? here)
• ? in the KIT lexical set
• ? in words like houses and places
• ?? in the GOAT lexical set (RP type English has
  ?? here)
• e? in the PRICE lexical set (RP type English
  has a? here).

32
3. Analysis

• We have conducted similar analyses on the
  Gateshead subcluster.
• There isnt time to go into the details, but the
  essence of the results is that, on the basis of
  the phonetic frequency profiles
• The main phonetic segmental determinants for the
  clustering can be extracted
• The main subclusters can be correlated with
  social factors, as follows.

33
3. Analysis

• 1. There is a correlation between gender and
  cluster
• Cluster A is almost exclusively female
• cluster BC is almost exclusively male
• cluster DE is mostly female.
• 2. There is a correlation between socio-economic
  status and cluster
• speakers in cluster A are, by and large, from the
  lowest of the three socio-economic groups
• speakers in cluster BC are from socio-economic
  groups 1 and 2
• speakers from cluster DE are from the two higher
  socio-economic groups, 2 and 3.

34
3. Analysis

• When the interaction between these two social
  variables in considered, the picture becomes even
  clearer. The following table summarises the
  typical social characteristics of the clusters

35
4. Discussion

• The foregoing analysis has used only one of a
  wide range of possible distance measure /
  clustering algorithm combinations available under
  the rubric hierarchical cluster analysis.
• In general, with respect to any given data set,
  different combinations of distance measure and
  clustering algorithm typically differ from one
  another to greater or lesser degrees.
• Given that there is no obvious way of selecting
  the best analysis that captures the true
  structure of the data, the question is how
  reliable a tool is hierarchical cluster analysis
  for linguistic research?
• Our own cluster analyses have shown such
  variation with respect to the TLS data.
• How can we claim any validity for our results?

36
4. Discussion

• In principle, the response is that the purpose of
  exploratory multivariate analysis is to generate
  hypotheses rather than to provide definitive
  answers, and different analyses of the same data
  simply generate different hypotheses to be
  tested.
• In practice, if different selections of distance
  measure / clustering algorithm produce wildly
  different analyses of one and the same data set,
  its difficult to see what advantage they offer
  over unguided hypothesizing.
• Therefore, the next step is to analyze the data
  using
• various other distance measure / clustering
  algorithm combinations in hierarchical cluster
  analysis
• Different clustering methods such as
  multidimensional scaling and self organizing maps
• to see if a stable analysis emerges.

37
4. Conclusion

• The hierarchical analyses we have undertaken so
  far look promising. The next steps are
• Detailed phonetic and sociolinguistic analysis of
  the cluster tree generated by the distance
  measure / clustering algorithm used in this talk.
• Comparison of this tree with structural analyses
  generated by other varieties of hierarchical
  cluster analysis and by alternative clustering
  methods.
• Relation of our results to existing work on
  Tyneside English, and in particular to that of
  Val Jones-Sargent, one of the original TLS
  researchers, on whose work our own is based.
• V. Jones-Sargent, Tyne Bytes. A computerised
  sociolinguistic study of Tyneside, Peter Lang,
  1983