en:pojmy:din - Příručka ČNK

This page is read only. You can view the source, but not change it. Ask your administrator if you think this is wrong.
====== DIN ======

DIN (Difference index) is a so called effect-size metric, i.e. measure designed((see Fidler, M. - Cvrček, V.: {{:pojmy:josl-separat.pdf|A Data-Driven Analysis of Reader Viewpoints: Reconstructing the Historical Reader Using Keyword Analysis}})) for the purpose of quantifying the relevance of a difference between values. DIN is implemented for extracting prominent units from a text (keywords) in the [[en:manualy:kwords|KWords]] tool. 

===== Significance and relevance =====

When comparing values (e.g. frequencies of words) we should be interested not only in the statistical significance but also whether the difference under consideration is actualy relevant for the description. Statistical significance can be obtained by several tests (e.g. chi2 test, Fisher's test or log-likelihood test).((It is unimportant for the time being that these test can also be employed as association measures for the extraction of collocastions)) Significance is usualy expressed as a p-value, i.e. the probability that the difference is caused by chance or variation within the data.

Even if the difference is signifiacnt it does not necesarily entails that it is relevant for the description. Even a small difference can be significant when there is a lot of results available. That is why the statistical significance information is often combined with the effect-size.
===== How it works =====

In the model example of extracting prominent words (keywords) from a text we proceed in the following way. For units which display a statistically significant difference (according to the selected statistical test -- [[en:pojmy:chi2|chi2]] or [[en:pojmy:loglikelihood|log-likelihood]]), the **DIN** value is subsequently calculated (difference index), which is indicative of how relevant the difference is:

$$DIN = 100 \times \frac{RelFq(Ttxt) - RelFq(RefC)}{RelFq(Ttxt) + RelFq(RefC)}$$

where $RelFq(Ttxt)$ is the relative frequency of the phenomenon in the analyzed text (target text) and $RelFq(RefC)$ is the relative frequency of the same phenomenon in the reference corpus. 

The formula takes into account the difference between relative frequencies (numerator) in relation to the frequency level of the items under comparison. This reference frequency level can be represented by their average value, as can be seen in this formula which is equivalent with the above (the coefficient changed from 100 to 50):

{{:pojmy:vzorecdin2.png?nolink&350|}}

===== DIN Values =====

The DIN values are designed to reach values from -100 to 100, it being understood that:
  * a value of -100 means that the given phenomenon does not occur in the analyzed text and is only in the reference corpus (therefore the word is not prominent in the analyzed text)
  * a value of 0 means that the given phenomenon has approximately the same relative frequency in the analyzed text and in the reference corpus (therefore the word is not prominent in the analyzed text)
  * a value of 100 means that the word occurs only in the analyzed text (it can therefore be a very prominent word ((In such cases it is important to keep in mind that the complete absence of a word from a reference corpus is unusual and special attention should be paid to it; a word might not occur in a reference corpus e.g. because it is very rare, cited from another language etc.)))

In texts of up to 20 thousand words in length and for analysis of [[en:pojmy:word|word forms]], DIN values in the 75-100 range can be considered to be of interest and they indicate that the unit is probably prominent and can be used as a basis for the interpretation of the entire text.

Furthermore, the KWords application offers a whole range of additional information for work with keywords. Apart from the list of keywords and their values it is the data dispersion graph (showing the status of the individual keywords in the text), a graph of so-called keyword links, i.e. relations between keywords in the text and also a concordance of keywords for an analysis of their immediate context.

The KWords application was also designed for creating analyses of temporal (or other) data series. If the user inputs more texts into the application (the maximum amount is 20), the so-called **multi-analysis** regime is activated. This regime analyzes all the inserted texts and the results of the individual analyses are compared based on the DIN.
Trace: • tag • morfologicka_analyza • konkordance • strukturni_atributy_mluvene • mluveny • reseni_ukolu • genre • txtype_group • zaciname • afun