PhD Notes

Contents:

\[\gdef{\S}{\mathcal{S}} % Note: overrides existing command \gdef{\O}{\mathcal{O}} % Note: overrides existing command \gdef{\F}{\mathcal{F}} \gdef{\N}{\mathcal{N}} \gdef{\R}{\mathbb{R}} \gdef{\Nat}{\mathbb{N}} \gdef{\sign}{sign} \gdef{\L}{\mathcal{L}} \gdef{\C}{\mathcal{C}} \gdef{\V}{\mathcal{V}} \gdef{\W}{\mathcal{W}} \gdef{\G}{\mathcal{G}} \gdef{\vle}{\preceq} % Source and fact orderings \gdef{\sle}{\sqsubseteq} \gdef{\slt}{\sqsubset} \gdef{\sge}{\ge} \gdef{\seq}{\simeq} \gdef{\fle}{\preceq} \gdef{\flt}{\prec} \gdef{\fgt}{\succ} \gdef{\fge}{\succeq} \gdef{\feq}{\approx} \gdef{\src}{\mathsf{src}} \gdef{\facts}{\mathsf{facts}} \gdef{\obj}{\mathsf{obj}} \gdef{\mut}{\mathsf{mut}} \gdef{\orderings}{\mathcal{L}} \gdef{\num}{\mathcal{T}_{Num}} \gdef{\rec}{\mathsf{rec}} \gdef{\norm}{\mathsf{norm}} \gdef{\ord}#1{\langle{#1}\rangle} % Shortcuts \gdef{\limn}{\lim_{n \to \infty}} \gdef{\voting}{\emph{Voting}} \gdef{\sums}{\emph{Sums}} \gdef{\usums}{\emph{UnboundedSums}} \gdef{\avlog}{\emph{Average$\cdot$Log}} \gdef{\scvoting}{\emph{SC-Voting}} \gdef{\scoh}{\mathrel{\lhd}} \gdef{\fcoh}{\mathrel{\blacktriangleleft}} \gdef{\tuple}#1{{\langle{#1}\rangle}}\]

Future areas

Ideas which require a lot of time, background reading and catching up with the literature, but which are still relevant to truth discovery and may be tackled eventually.

Philosophical issues and problems with the model

So far little attention has been given to the philosophical interpretation of ‘trust’, ‘belief’ and ‘truth’ for truth discovery. Trust seems to be applied loosely, with different interpretations in different algorithms. In fact, for most truth discovery algorithms we could replace ‘trust’ with ‘reliability’ and nothing would change. This indicates there is no specific notion of trust in play. The same applies to our paper, where the particular notions of truth and trust are left undefined.

Related to this, Martin offered the following feedback on the paper

Something isn’t true just because many people believe in it.

The notion of truth needs to be grounded in something.

I suppose the first point is related to our Unanimity axiom, where we appear to be saying that facts claimed by all sources are ‘true’. In fact, we mean that the fact is the most plausible in some sense. I think we need to work out in which sense this is, and make it clear in any future work. The problem could also be related to subjective vs objective truth.

So, perhaps I should read up on trust and truth on a more philosophical level to understand what it means to say that some source is more trustworthy than another, and to say that a fact is true. Even if there is no single answer, we could at least attempt to define what we mean by these terms. See Marsh’s thesis (1994) for this and probably some other references.

When it comes to our model for truth discovery there are more philosophical issues. Firstly, we try to give a global ranking of trust and belief, whereas surely trust should depend on one’s point of view. Maybe the stuff on group recommendations, where an agent’s recommendations are based on their trusted neighbours, could apply to truth discovery, and we have personalised trust and belief rankings.

Related to the above is that the trust ranking covers all domains. Surely in reality a source can be reliable and trustworthy in one domain (say, sports) but not reliable in another (say, medicine). It seems weird to give a single ranking across all domains. I think Pasternack’s thesis discusses this, and also Richard’s paper on trust in belief revision uses this idea.

Who knows more?

See Martin’s email and suggested reading list. The basic idea is to work out which agent is more ‘knowledgable’ given their beliefs in a nonmonotonic setting.

This could be related to truth discovery. Maybe a source is trustworthy if he knows lots about his domain and doesn’t ‘bullshit’ (could incorporate ideas from Martin’s work on lying vs bullshit etc).

Logic based work

Maybe we could give truth discovery a logic-based formalism instead of (or extending the current one). Current vague ideas of things to think about:

  • Multi-agent epistemic logic? Or maybe doxastic logic (referring to beliefs rather than knowledge) is a better term. E.g. write \(B_i\phi\) to mean that source \(i\) believes \(\phi\) is true. As well as expressing the source claims as is currently done, the sources could make claims about other source’s claims; e.g. \(B_iB_j\phi\) means that source \(i\) claims that source \(j\) claims \(\phi\). This would be some kind of modal logic.

Combined NLP and argumentation?

There seems to be some work on ‘fake news detection’ which uses NLP to extract claims (e.g. news articles). On the other hand, argumentation mining extracts structured arguments from textual documents. Combining fake news detection with argumentation could be interesting for truth discovery, if it has not already been done.

This 2016 paper might be a good place to look: Argumentation Mining: State of the Art and Emerging Trends.