The web-site is now in readonly mode. Login and registration are disabled. (28 June 2019)

Computation / rhetoric / text analysis / argumentation Workshop

Cloud created by:

11 October 2010

Workshop at The Open University in October 2010 to discuss computation, rhetoric, text analysis and argumentation.

Extra content

Approximate agenda!...

9.30 - Simon Buckingham Shum (KMI) - Welcome

9.35 - Simon Buckingham Shum & Anna De Liddo (KMI)

10.05 - Rebecca Ferguson (IET/KMI)

10.35 - Ágnes Sándor (Xerox, & visiting OLnet Project Fellow, IET/KMI)

11.05 - Go for a coffee

11.35 - Denise Whitelock (IET)

12.05 - Paul Piwek & Svetlana Stoyanchev (Computing)

12.35 - Sandra Williams & Richard Power (Computing)

1.05 - Lunch and decide what to do next...

Simon Buckingham Shum
05:38 on 13 October 2010 (Edited 05:44 on 13 October 2010)

Liveblog of talk by Simon Buckingham Shum

Very interested in conversation. Lots of learning happens through discourse.

Lots of different types of discourse.

Fundamentally, people don’t really disagree. Most platforms have no idea about contested knowledge, and about levels of agreement and disagreement

Lots of tools to detect and render potentially significant patterns in information.

Starting to see tools that can lift out patterns from the text and from the people behind the text.

We make meaningful connections between information elements. Need to be able to represent these.

Meaning is dependent on individuals and context. There are different interpretations of the same event.

Want to be able to track ideas.

Ideas are first-class objects in the system. Need to be able to reference, link, embed and share. They may not all be contained within one document.

Work with hypertext because this allows representation of an idea as a node in a network.


Start to see connections between ideas.

Some of these are agreements, and some are challenges.

Start to see a discourse emerging.

People looking at documents, or in groups, tutorials, seminars, do this already, but there is no means of rendering it at present.

Can currently add notes, but there are no semantic links.

Want to introduce a language for argumentation – eg questions, answers, arguments, challenges…

Cohere  introduces a set of icons to represent ideas, and provides a set of relationships to connect them.

Allows people to connect anything to anything in a Web 2.0 context.

Users can add any connection type they want. Helpful if they can add a polarity (eg pro or con) as this allows the program to make sense of new connections.

Icons can also be customised.

Question: How can this semantic network be linearised as a textual narrative? Can we parse these connections to produce near-natural0language text?

Is there a theoretical basis that can be used to derive these relationships.

Coherence relations theory.

Is there any grounds for constraining relationships?

How is the ontology of these relationships ordered?

Are there basic types of relationship eg causal , temporal and additive. These have polarity.

See work by Clara Mancini and Neil Benn.

Now have a dataset of the relationships that people make – crowdsourced many types of semantic relationships Can they now be parsed and tied in with cognitive coherence relations.

Detailed tree diagram on cognitive coherence relations.

Diagram applies to contested domains.

This is an examination of semantic relations, rather than of dialogue.

Rebecca Ferguson
09:22 on 13 October 2010

Anna De Liddo: Talk on Cohere

Tools with which users can create different types of annotation about things they are exploring on the web.

What if we think about this idea as something that could be placed in an asynchronous discussion.

For example, I make a comment on a document. What if I express the relationship this comment could have into an online dialogue with other users? Can this be represented to show the rhetorical moves?

Anna shows demonstration video of Cohere.

Can associate an icon with the comment and can identify its rhetorical role eg argument in favour, question, opinion. Can also add contextual information.

Cohere supports webpage annotation. Users can highlight text and then make a comment in a sidebar – again, identifying the rhetorical role of their comment. Other users who visit the page can see these comments and their rhetorical roles.

Can connect ideas using a connection builder. These don’t have to be just your own ideas – you can connect them with other people’s ideas.

Can create a discussion group and share annotations and ideas within that group.

Cohere then allows to visualise the network of comments and connections as a social semantic network. This represents the ideas and their rhetorical role.

Long discussions can become difficult to make sense of.

Can use semantic filtering of the network. For example, can focus on one question and the ideas that are consistent with it.

Question: Have you used it to produce a text? Interested in doing this in the future. Are these networks quite hard for humans to make sense of?

There is a literature about how to make concept maps and argument maps more comprehensible for users.

Understanding the visual language gives an entry into hitherto hidden meanings within the dialogue.

This can be a tool for knowledge discovery.

Can this be used to create a multi-authored text, or a representation of a debate. It is not easy to make sense of a dialogue that is going on online. There are many ways of creating a linear text from varied sources of knowledge.

Want to provide easier to read summaries of online discourse.

Analogy – report-backs from break-out groups – one person explaining what has happened with reference to a flipchart.


Rebecca Ferguson
10:13 on 13 October 2010

Q&A on Rebecca's talk:

Is counting the words in the whole contribution misleading, since it might be just one sentence that triggers? (but analysis suggests that bigger contributions do have more exploratory dialogue)

Segment the dialogue into sections further than whole sessions?

You could predict the likelihood of exploratory dialogue statistically

Variance may be critical

Bales' group talk categories may be of use, perhaps more than Mercer. Bales argued that there was typically an intellectual leader and socio-emotive leader in a group

Social network analysis of this data might illuminate further

*Questions* are central to sensemaking. (but complicated by online dialogue punctuation)

Look at content words in questions, and compare to content of the talk?

Simon Buckingham Shum
10:17 on 13 October 2010 (Edited 10:18 on 13 October 2010)

Longer version of Simon & Anna's presentation and demo: Visual Analytics summer school lecture and demos.

Simon Buckingham Shum
10:21 on 13 October 2010

Notes from Agnes Sandor's Talk

Wants to identify the key ideas in scientific texts.

Parses texts into syntactic relationships. Analyse about twenty languages.- eg English, French, Portuguese, Italian.

Mainly do dependency parsing. Have a number of applications eg opinion mining and knowledge representation

Cover a number of domains eg biology, litigation and health.

Analysis level is the sentence.

Gives example of a sentence that can be split down into various propositions. Also identifies argumentation and rhetorical function.

Information extraction is used to analyse the propositions.

Rhetorical analysis identifies coherence.

Rhetorical function (eg ‘In contrast with previous hypotheses’) is handled by concept matching.

Everything that has a predicate is identified as a proposition – this is a linguistic analysis. Predicates are extracted using syntactic rules and patterns.

Interested in rhetorical functions that are related to research problems. If we know what problems there are, we know what the article is about.

Background knowledge, generalising, contrast (contrasting ideas, novelty, significance, surprise), open question, summarising.


  • Background knowledge indicators eg ‘recent studies indicated’, ‘is universally accepted’
  • Generalising: ‘emerging as a promising approach’
  • Contrasting: ‘unorthodox view’
  • Novelty ‘we suggest a new approach’
  • Surprise: ‘We have identified unusual’
  • Summarising: ‘the goal of this study’


No common semantic patterns, not collocations, not paraphrases or synonyms.

These are units in conceptual terms rather than in linguistic terms.


For example, for contrasting ideas, you have ideas of contrast, past and ideas. This is not jut down to identifying words, but also to identifying syntactic relationships (using the syntactic parser).

Coherence is identified by syntax in language.

Rhetorical functions have constituent concepts that are used within scientific argument (eg scope, temporality, contrast). Ultimately, they all come down to words. (eg indicators of temporality are early, traditional, remain. Indicators of generality are many, frequently, range)

Table shows how rhetorical functions are matched by constituent concepts and are matched by words.

Rebecca Ferguson
10:45 on 13 October 2010

Notes from Agnes talk:

Rhetorical moves in scientific texts are not linguistic units which can be picked up by current computational linguistic approaches, but conceptual units. We need a new way to detect them.

EERQI - European Educational Research Quality Indicators project:
Educational science research publications analysed by the XIP Xerox parser, looking for key segments in the text, and feeding into a scholarly search engine

IET/KMI OLnet Project: >100 OER project reports being analysed by humans and XIP, to compare and contrast

Simon Buckingham Shum
11:09 on 13 October 2010

Notes from Denise's talk:


  • Found a correlation between types of tutor feedback (using Bales framework) and pass level
  • Bales categories generalise across disciplines, which confirms that the tool is picking up ways of teaching
  • Bar chart giving the tutor feedback on the appropriate quantitiy of their different comments on TMAs


  • Focus on causal reasoning in History
  • Moodle module giving feedback on free text student responses


Simon Buckingham Shum
11:27 on 13 October 2010

Paul & Svetlana's talk

  • research suggests that for learning and persuasion, presenting info as a dialogue is more effective than a monologue
  • dialogue for exploring complex ideas has a long history of course, from Plato onwards


  • Denise's work might contribute new kinds of dialogue acts, with pedagogical intent (Bales)
  • Argumentation theory classifies different kinds of dialogue with different kinds of argumentation schemes and Critical Questions (Doug Walton)
  • KMi has annotated multiparty dialogues and identified different kinds of Qs

Simon Buckingham Shum
11:58 on 13 October 2010

Embedded Content


Contribute to the discussion

Please log in to post a comment. Register here if you haven't signed up yet.