Trending words, neologisms

Neologisms and diachronic analysis of word usage

Trends is a feature for detecting words which undergo changes in the frequency of use in time (diachronic analysis). Trends identify words whose use increases or decreases in time.

Lexicologists can use trends to identify new words (neologisms) and  historians can use trends to identify the point time when a word started to be used, stopped being used or when it saw an unusually increased or decreased use.

For a detailed description of the algorithms and comparison of their performance, see Ondřej Herman’s thesis.

Which corpora can be used?

Trends will only work with a corpus with time stams, i.e. a corpus whose documents are annotated with a (publication) date. The largest corpus in Sketch Engine compatible with trends is the Timestamped corpus in many languages.

Using Trends

A timestamped corpus has to be selected for Trends to work. Use the BASIC or ADVANCED tab to set the criteria if necessary.

Use the tooltips attached to the elements of the input forms to learn about the individual settings and controls.

The default settings do not have to be changed to produce a result.

Trends feature - result page

By default, the results are sorted by the absolute value of the change. The words with the biggest change will be at the top irrespective of whether the change was positive (=growing usage) or negative (=decreasing usage).

Size and balance

There is no minimal requirement for corpus size but to receive usable result, the corpus should contain a decent amount of data for each period. At present, our smallest corpus with trends is the Anthology Reference Corpus with 38 million words. It is unlikely to receive usable results from a corpus of 1 million words.

The balance of periods is also important. If most texts belong to only a few periods from a wide range of periods, the results will be biased towards those periods.

Configuring a corpus for use with Trends

Data have to be annotated with time stamps. The timestamp attribute is user-defined (e.g. pub_date). Most corpora only have one timestamp attribute but the same corpus can contain several . The user can select which timestap should be used for the computation.  It is not necessary to have the timestamp in each document, but trends will be computed only from documents with the time stamps.

Annotations timestamps

The beginning of the configuration file must contain the definition of the timestamp attribute, for example:

DIACHRONIC "doc.pub"

The number of attribute values should not exceed 500. The values must be composed of the same number of characters, the longest time period (e.g. year) must come first, the shortest last (e.g. day). Non-numerical characters will not produce an error but will be ignored.

Examples of valid values

All values within the same corpus must have the same format.

1958
1964-05
1994-03-27
year2004
Y2011M04D12

Invalid values

2004Mar14 – the month will be lost

If you have problems with setting things up, please ask for assistance.

(Words in bold refer to the screenshot above.)


word
clicking the word will bring up a frequency list for the word like this broken up by the time intervals (see details on the right)

Trends - frequency list

Frequency list broken down by the year, the time interval available depends on how the corpus is annotated, corpora can be annotated for smaller or bigger units such as months or decades.

P – positive filter – will show a concordance from the given period
N – negative filter – will show a concordance from all periods but not the one that was clicked
Frequency – absolute frequency
Rel [%] – relative text type frequency


trend
the value indicating the degree of change

p-value
expresses the degree of change, see p-value

freq
the frequency in the corpus (number of hits), clicking the number will bring up a concordance

p-value (statistical significance) – a function of the observed sample results that is used for testing a statistical hypothesis

slope (trend direction) – a number that describes the direction and the steepness of a line

Theil-Sen – a non-parametric method that finds the median slope among all lines designed by pairs of sample points

Linear regression – a method used for modeling the linear relationship between a dependent variable and one or more explanatory variables

Adam Kilgarriff, Ondřej Herman, Jan Bušta, Pavel Rychlý and Miloš Jakubíček. DIACRAN: a framework for diachronic analysis (presentation). In Corpus Linguistics (CL2015), United Kingdom, July 2015.

Ondřej Herman and Vojtěch Kovář. Methods for Detection of Word Usage over Time. In Seventh Workshop on Recent Advances in Slavonic Natural Language Processing, RASLAN 2013. Brno: Tribun EU, 2013, pp. 79–85. ISBN 978-80-263-0520-0.

Ondřej Herman (2013). Automatic methods for detection of word usage in time. Bachelor thesis. Masaryk University, Faculty of Informatics.

A script “mktrends” is part of Sketch Engine. If you have a local installation, you can call the script directly in your command-line. Without parameters, it gives this information.

usage: /usr/bin/mktrends CORPUSNAME STRUCTATTR ATTR STATISTIC LIMIT EPOCH_LIMIT [SUBCORPFILE]
	EPOCH_LIMIT: attribute values with smaller norm are discarded
	LIMIT: words with fewer occurrences are skipped
	METHODS: comma-separated list of methods to compute
		possible values are: linreg_nonzero, mkts_all, mkts_nonzero, linreg_all
example: /usr/bin/mktrends bnc bncdoc.year word mkts_all 100 100000