This past summer, I’ve had the pleasure of working with Professor Chen Ding, Lucinda Liu, and Professor Daniel Gildea on an all-timescale co-occurrence analysis algorithm for optimizing microprocessor storage usage. This analysis can efficiently provide co-occurrence ratios for multiple timescales, which are time allowed for or taken by a process or sequence of events. The timescale is used to determine the likelihood of two events occurring together within that time window, giving a co-occurrence ratio. The developed algorithm is unique in its ability to calculate co-occurrence ratios for multiple timescales more efficiently than a brute-force method.
An initial pass of a trace of objects stores the distance between different elements of the trace. These distances can be used to see where gaps between trace objects are to wide to fall in the same time window, giving us a count of absence windows. The beauty of this algorithm is that the expensive operation, the initial pass of the trace, only needs to be done once to calculate timescale analysis of all pairs in a trace. By counting absence windows, joint and single frequencies can be determined.
We define the co-occurrence ratio to be the joint frequency of the pair divided by the single frequency of a given element. The given element is known to be present in a time window, meaning that the co-occurrence ratio represents the probability of finding the other element in the same time window. This makes the co-occurrence ratio a conditional probability.
This analysis was initially intended for use in optimizing processor cache utilization, by placing frequently co-occurring memory objects together in memory. The precise explanation can be found this paper. However, this concept can be extrapolated to any trace, whether it be a trace of words, or music notes, or pixels in a photo.
We applied our algorithm to Shakespeare’s Sonnets and the King James Bible, treating each character as a memory access taking one unit of time, skipping symbols and spaces. The letters, ignoring capitalization, are treated as memory accesses. We used timescales of 2, 5, and 10 to look at adjacent pairs, what one hand can type, and what two hands can type. Below are links to the data files and analysis of Shakespeare’s Sonnets and the King James Bible for multiple timescales using our algorithm:
|Shakespeare’s Sonnets||Timescale 2||Timescale 5||Timescale 10|
|King James Bible||Timescale 2||Timescale 5||Timescale 10|
The data can be interpreted as the conditional probability of seeing the second character given the first character is present in a given time window. Each individual character appears the same number of windows as the timescale size. So, if a second character is always adjacent to the first character on the same side the maximum co-occurrence ratio should be (timescale – 1)⁄timescale. The reason for this is that there will be one window that includes the first character but not the second.
For example, the character ‘q’ should always precede the character ‘u’. So, given a time window with ‘q’ the probability seeing ‘u’ in the same window should be at least (timescale – 1)⁄timescale. This means that all pairs with a co-occurrence ratio above that value have the second character on both sides of the first character like in the case of all word containing the substring “eve”. This is why for a timescale of 2 in Shakespeare’s Sonnets the pair (q, u) can be beaten by pairs like (x, e), but with larger timescales u’s from other words boost the (q, u) co-occurrence to put it in the lead.
An example from Shakespeare’s Fourth Sonnet. Note how with the pair (q, u) the first window contains ‘q’ but not ‘u’ while the other nine windows have both characters. Also, the ‘q’ present in this trace appears in the same number of windows as the timescale size.
It is important to note that this method of co-occurrence analysis is not tallying frequencies. So when the pair (q, u) has the highest co-occurrence ratio, it does not mean that is the most common pair. Since the character ‘q’ is rarely used in the English language, its single occurrences are low, while its joint occurrences with the character ‘u’ are high. This is why for all the tests above many of the pairs with the highest affinities are uncommon consonants paired with a much more common vowel, like ‘e’. In both texts where the timescale 10 analyses were used, the top 10 pairs, excluding (q, u), have the character ‘e’ as a second character.
The data collected through this analysis has numerous uses outside of literature. The co-occurrence information can be used to evaluate the querty keyboard layout, or be used in speech therapy. Both the texts used here are from the same era, the turn of the sixteenth century into the seventeenth. It is possible that the character co-occurrence ratios could change with modern text or transcribed speech. Other than analyzing text, this analysis can be applied to images or videos for facial or object recognition.
I hope that this interesting re-application of memory layout analysis was entertaining to read about and I’d like to thank Professor Chen Ding, Lucinda Liu, and Professor Daniel Gildea for working with me on this project.
Image Sources: https://upload.wikimedia.org/wikipedia/commons/f/f6/Sonnets1609titlepage.jpg