Brainstorming the Future of Corpus Tools

Since arriving at the Centre for Corpus Approaches to Social Science (CASS), I’ve been thinking a lot about corpus tools. As I wrote in my blog entry of June 3, I have been working on various software programs to help corpus linguists process and analyse texts, including VariAnt, SarAnt, TagAnt. Since then, I’ve also updated my mono-corpus analysis toolkit, AntConc, as well as updated my desktop and web-based parallel corpus tools, including AntPConc and the interfaces to the ENEJE and EXEMPRAES corpora. I’ve even started working with Paul Baker of Lancaster University on a completely new tool that provides detailed analyses of keywords.

In preparation for my plenary talk on corpus tools, given at the Teaching and Language Corpora (TaLC 11) conference held at Lancaster University, I interviewed many corpus linguists about their uses of corpus tools and their views on the future of corpus tools. I also interviewed people from other fields about their views on tools, including Jim Wild, the Vice President of the Royal Astronomical Society.

From my investigations, it was clear that corpus linguists rely on and very much appreciate the importance of tools in their work. But, it also became clear that corpus linguists can sometimes find it difficult to see beyond the features of their preferred concordancer or word frequency generator and attempt to look at language data in completely new and interesting ways. An analogy I often use (and one I detailed in my plenary talk at TaLC 11) is that of an astronomer. Corpus linguists can sometimes find that their telescopes are not powerful enough or sophisticated enough to delve into the depths of their research space. But, rather than attempting to build new telescopes that would reveal what they hope to see (an analogy to programming) or working with others to build such a telescope (an analogy to working with a software developer), corpus linguists simply turn their telescopes to other areas of the sky where their existing telescopes will continue to suffice.

To raise the awareness of corpus tools in the field and also generate new ideas for corpus tools that might be developed by individual programmers or within team projects, I proposed the first corpus tools brainstorming session at the 2014 American Association of Corpus Linguistics (AACL 2014) conference. Randi Reppen and the other organizers of the conference strongly supported the idea, and it finally became a reality on September 25, 2014, the first day of the conference.

At the session, over 30 people participated, filling the room. After I gave a brief overview of the history of corpus tools development, the participants thought about the ways in which they currently use corpora and the tools needed to do their work. The usual suspects—frequency lists (and frequency list comparisons), keyword-in-context concordances and plots, clusters and n-grams, collocates, and keywords—were all mentioned. In addition, the participants talked about how they are increasingly using statistics tools and also starting programming to find dispersion measures. A summary of the ways people use corpora is given below:

  • find word/phrase patterns (KWIC)
  • find word/phrase positions (plot)
  • find collocates
  • find n-grams/lexical bundles
  • find clusters
  • generate word lists
  • generate keyword lists
  • match patterns in text (via scripting)
  • generate statistics (e.g. using R)
  • measure dispersion of word/phrase patterns
  • compare words/synonyms
  • identify characteristics of texts

Next, the participants formed groups, and began brainstorming ideas for new tools that they would like to see developed. Each group came up with many ideas, and explained these to the session as a whole. The ideas are summarised below:

  • compute distances between subsequent occurrences of search patterns (e.g. words, lemmas, POS)
  • quantify the degree of variability around search patterns
  • generate counts per text (in addition to corpus)
  • extract definitions
  • find patterns of range and frequency
  • work with private data but allow  for powerful handling of annotation (e.g. comparing frequencies of sub-corpora)
  • carry out extensive move analysis over large texts
  • search corpora by semantic class
  • process audio data
  • carry out phonological analysis (e.g. neighbor density)
  • use tools to build a corpus (e.g. finding texts, annotating texts, converting non-ASCII characters to ASCII)
  • create new visualizations of data (e.g. a roman candle of words that ‘explode’ out of a text)
  • identify the encoding of corpus texts
  • compare two corpora along many dimensions
  • identify changes in language over time
  • disambiguate word senses

From the list, it is clear that the field is moving towards more sophisticated analyses of data. People are also thinking of new and interesting ways to analyse corpora. But, perhaps the list also reveals a tendency for corpus linguists to think more in terms of what they can do rather than what they should do, an observation made by Douglas Biber, who also attended the session. As Jim Wild said when I interviewed him in July, “Research should be led by the science not the tool.” In corpus linguistics, clearly we should not be trapped into a particular research topic because of the limitations of the tools available to us. We should always strive to answer the questions that need to be answered. If the current tools cannot help us answer those questions, we may need to work with a software developer or perhaps even start learning to program ourselves so that new tools will emerge to help us tackle these difficult questions.

I am very happy that I was able to organize the corpus tools brainstorming session at AACL 2014, and I would like to thank all the participants for coming and sharing their ideas. I will continue thinking about corpus tools and working to make some of the ideas suggested at the session become a reality.

The complete slides for the AACL 2014 corpus tools brainstorming session can be found here. My personal website is here.