Inconsistency

The NewsTools Project




Analysing and acting on inconsistency in news


If news reports are in the form of structured text, knowledge-based systems technology can be used for automated reasoning with the information in each news report. This may include tools for looking for inconsistencies in news reports and for merging potentially conflicting news reports from multiple sources.

It is common to spot inconsistencies between a news report and domain knowledge. Some of these inconsistencies may be more significant than others, and some news reports may contain more inconsistencies than others.

Having some understanding of the "degree of inconsistency" of a news report can help in deciding how to act on it. Moreover, inconsistencies between information in a news report and domain knowledge can tell us important things about the news report. An inconsistency may indicate an error in the news report, or an anomaly with respect to some expectation, or that the news report is part of some unexpected trend. To address this, at UCL we are developing Inconsistent News Measurement Tools to help users of news to spot inconsistencies in news, to localize inconsistencies, to determine the degree of inconsistency, and to act on them.

Since the nature of inconsistency is with respect to the user of the news report, consider a user who is an analyst of listed companies. Some inconsistencies are likely to be significant for an analyst. For example, suppose in a news report on the latest quarterly results by an FT100 company it states that the current pre-tax profit is GBP 45 million and that this is 90% fall from the previous quarter, but the domain knowledge has the pre-tax profit for the previous quarter as GBP 45 million. In this case, it seems that in the news report either the value GBP 45 million or the 90% fall is erroneous. This error could be important for a trader considering transactions in this share.

In constrast some inconsistencies are likely to be insignificant for an analyst. For example in a news report on the takeover of Orange by France Telecom, the information that Orange is based in London is inconsistent with the domain knowledge that Orange is actually based in Bristol. We may say this is insignificant because it is unlikely a trader is that bothered about whether it is London or Bristol.

Some inconsistencies arise from violation of hard constraints. For example, a news report may have BP Ltd as being listed on the London Stock Exchange, but this violates the constraint that all companies listed here are Plcs.

More interestingly, some news reports violate expectations. For example, Vodafone was trying to buy Mannesmann for some time, when there was a news report on FT.com that said that Mannesmann had just agreed to buy Vodafone. In this case, the violation of expectation was a typographical error.

But there are plenty of examples where unexpected news is announced that is true, and increasingly often is announced via the web where typographical errors are prevalent. For example, consider a dotcom company, which has been growing at an enourmous rate and this growth is used to justify its enourmous capitalization. Expectations are it is a well-run company with much promise, and so this growth will continue strongly. This can be represented in a knowledge-base. Now suppose there is a news report that reveals rumours that its sales have decreased sharply since last quarter. This information would cause an inconsistency with the default expectation about sales growth, and as such may be useful if flagged for the analyst.

Whilst some of this is the subject of ongoing research, it is clear that there is a trend now for tools and services with increased functionality for managing news. Technologies including improved web indexing, information extraction, XML, and knowledge-based reasoning systems, are creating opportunities for adding value and helping users.

For more information about the NewsTools Project, contact Tony Hunter (a.hunter@cs.ucl.ac.uk or +44 20 7679 7295), or look at the NewsTools Project home page.