Corpus Linguistics & Statistics @ UoBham, 11/02/16 by John Williams https://padlet.com/johnxwilliams1/y7v657kn6p3o Scroll down and along to see all notes en-us 2016-02-11 13:58:30 UTC 2025-11-30 21:24:49 UTC hello@padlet.com Michaela Mahlberg: Intro johnxwilliams1 https://padlet.com/johnxwilliams1/y7v657kn6p3o/wish/94590102
  • Language as a social phenomenon
  • Meaning and form are associated (lexico-grammar)
  • CL prioritizes lexis
    • - in texts and between texts
  • Meaning is based on evidence of interaction: selection of linguistically relevant patterns, depends on RQ

    • "Scholars don't pay enough attention to what non-scholars think about the world." (Proctor 2012)


    ]]>     2016-02-11 14:07:27 UTC https://padlet.com/johnxwilliams1/y7v657kn6p3o/wish/94590102     Simon Preston: Corpus Analysis from Math perspective johnxwilliams1 https://padlet.com/johnxwilliams1/y7v657kn6p3o/wish/94595397
    • Use old existing (easier?) solutions
    • Corpus analysis in terms of input (math representation of corpus X) & output - f(X)
    • Deciding on X (eg. 'bag of words' representation --> matrix) forces us to decide what we retain and discard
    • Try to represent it as a relationship between simpler matrices ('matrix factorization')
    • 'bag of words' representation discards information about the order of words
    • get round that with co-occurrence matrix --> network visualization
    • Challenges:
      • How to analyse time-structured corpora & co-occurrence networks (or combinations thereof - 'time-dependent networks')

    ]]>     2016-02-11 14:22:34 UTC https://padlet.com/johnxwilliams1/y7v657kn6p3o/wish/94595397     Thompson, Murakami, Hunston: Topic Modelling johnxwilliams1 https://padlet.com/johnxwilliams1/y7v657kn6p3o/wish/94603737
    • 'Bottom-up' approach, no prior model of corpus
    • 'Topic' defined in terms of probability distribution of fixed vocabulary
    • We can model a document in terms of rolling a 'die' for choosing a topic, then rolling second die for the choice of words within that. 'Topic modelling' sort of reverses that process and tries to recreate the (irregular) shape of these dies.
    • Tested on research papers in the environmental domain, with a roster of 60 topics ('documents' may be component parts of papers, 'text chunks' --> topics may be recurring at characteristic points in research papers)
    • Topics may be prominent at particular time periods
    • Topics can be envisaged in terms of individual words,co-occurrence of invididual words, or n-grams
    ]]>     2016-02-11 14:43:26 UTC https://padlet.com/johnxwilliams1/y7v657kn6p3o/wish/94603737     Viola Wiegand: Identifying surveillance discourses johnxwilliams1 https://padlet.com/johnxwilliams1/y7v657kn6p3o/wish/94613171
  • The journal 'Surveillance & Society' mined for key words in the domain
  • Looked for common features for common patterns across the 13 volumes of the journal
  • Used 'key keywords', 'lockwords' & co-occurrence
  • 'KKs' are keywords that occur across a large number of texts in the corpus
  • 'Lockwords' - words that are stable in frequency across texts
  • Found 28 items that were both KK and lockwords --> looked at co-occurrent pairs, these can be mapped ('co-occurrence networks')
  • Linguistic interpretation still necessary - "No purely statistical analysis of language can reveal meaning".
    • ]]>     2016-02-11 15:06:00 UTC https://padlet.com/johnxwilliams1/y7v657kn6p3o/wish/94613171     Yves van Gennip: Graphical representations of a corpus, clustering johnxwilliams1 https://padlet.com/johnxwilliams1/y7v657kn6p3o/wish/94629564
  • Co-occurrence: need to make decisions about window-size, directionality, weighting - before making a graph
  • Graph = network of nodes/vertices and edges/links
  • Thickness of edge can indicate weightedness
  • Using a block matrix to indicate strength of co-occurrence
    • ]]>     2016-02-11 15:48:06 UTC https://padlet.com/johnxwilliams1/y7v657kn6p3o/wish/94629564     Hennessy: Time-dependency in corpora johnxwilliams1 https://padlet.com/johnxwilliams1/y7v657kn6p3o/wish/94638677
  • 'binning' = grouping documents in an axis of your matrix (eg. corresponding to time periods, then making simpler matrices based on each bin
  • width of bins depends on RQ & data
  • bins can overlap
  • 'kernel' - a kind of scaled bin to give a more 'realistic' view of effect size
  • this is a form of statistical smoothing
  • possible kernels come in a set of classic mathematical shapes

    • ]]>     2016-02-11 16:11:33 UTC https://padlet.com/johnxwilliams1/y7v657kn6p3o/wish/94638677     Smyth, Bull: The right to read = The right to mine johnxwilliams1 https://padlet.com/johnxwilliams1/y7v657kn6p3o/wish/94644894 (Librarians)

    • Data-Asset-Method: Harnessing the Infinite Archive --> a set of protocols for storing corpora uniformly, available for networks of researchers
    • UK copyright legislation 2014: computational analysis and transformation of data does not infringe copyright ('the right to mine' - for non-commercial purposes; may not apply to some international collaboration)
    • lot of publishers making their content available digitally
    • CORE portal for accessing open-access articles
    • CROSSREF is making these articles available for data mining
    ]]>     2016-02-11 16:28:20 UTC https://padlet.com/johnxwilliams1/y7v657kn6p3o/wish/94644894     Laurence Anthony: DIY corpus tools creations: pros and cons johnxwilliams1 https://padlet.com/johnxwilliams1/y7v657kn6p3o/wish/94652968 Design of new corpus tools tends to be dominated by computer scientists
    Corpus users tend to do research 'inside the box' --> AntConc downloads still going up --> Why aren't the new generation developing their own tools? Why are we not all programming?

    LA presents both sides of the debate:
    Pros:
    - (Biber) if you learn programming, you can do what you want, you are in the driver's seat, it can be cheaper
    - (Gries) "inflexible software creates inflexible research"
    - (Davies) distinguishes 'corpus users' from 'corpus creators'
    Advice:
    - Pick a popular language, eg. Python, Scratch (on Raspberry Pi), Java, R (very ugly)
    - Read a programming book
    - Join Stack Overflow

    Against:
    - Most corpus users can 'get by' with current tools
    - Researchers in many fields do not develop their own tools: it tends to be the fruit of collaboration between researchers & engineers
    - Programmers are a different world. DIY tools risk being less accurate, slower
    Advice:
    - Decide your research question before selecting your tool/method
    - Learn to use a *good* text editor, eg, Notepad++, TextWrangler
    - Read the user guide
    - Be proactive in contacting specialists
    - Provide motivation for getting specialists involved, treat them as part of the team
    - Understand the limitations of your tools and potential alternatives
    "Life is short"

    LA comes down on pro-programming side, but with teams integrated from the start. Even if users use only standard tools, they would benefit from a bit of programming nous.



    Check out: wordwanderer.org]]>     2016-02-11 16:47:37 UTC https://padlet.com/johnxwilliams1/y7v657kn6p3o/wish/94652968     Programme johnxwilliams1 https://padlet.com/johnxwilliams1/y7v657kn6p3o/wish/94722307 Accuracy of my notes not guaranteed, esp where statistics is concerned !  Apologies for any misrepresentation]]> 2016-02-11 20:00:16 UTC https://padlet.com/johnxwilliams1/y7v657kn6p3o/wish/94722307