Metaphor, Popular Science and Semantic Tagging: Distant Reading with the _Historical Thesaurus of English_

1. Introduction
This paper describes and implements a computational procedure for semantically analysing analogy in large bodies of text using a semantic annotation system based on the database of the Historical Thesaurus of English.1 In so doing, it demonstrates the value of a comprehensive and fine-grained semantic annotation system for English within corpus linguistics. Using log-likelihood measures on its semantically-annotated corpus of abstract popular science, the paper therefore demonstrates the existence, the extent, and the location of significant metaphorical content in this corpus. In so doing, it applies a version of Franco Moretti’s ‘distant reading’ programme in the analysis of literary history to non-narrative texts, as well as continuing work on integrating meaning into the methodologies of corpus linguistics.2

1.1. Analogy and Popular Science
Following the 1980 publication of George Lakoff and Mark Johnson’s Metaphors We Live By,3 it has been frequently stated that human beings, as embodied minds perceiving the mental, social and physical worlds around them, understand abstractions in terms of concrete entities. While this is a well-explicated concept in cognitive linguistics and psychology, few studies have yet aimed to establish both the extent and operation of this in a large corpus of discourse. The standard methodology in cognitive linguistics tends to rely on introspection and the intuitions of native speakers, at the expense of empirical data.4 This lack of rigour has resulted in results which, though "intuitively appealing", are criticized "for lacking a clear set of methodological decision principles".5 Following earlier work we have undertaken on the investigation of analogy and metaphor in English from empirical groundings,67 in this paper we discuss a methodology for identifying these textual phenomena automatically, and in so doing aim to open up cognitive linguistics to more digital humanities techniques, in addition to demonstrating the use of automated semantic annotation and disambiguation techniques at an unprecedented level of granularity.

1.2. The Corpus
We take as our initial data two book-length popular science texts which focus on explaining abstract concepts to a non-specialist audience, and therefore provide the greatest potential for the analysis of non-literary analogy - metaphor theory tells us that these should therefore be rich in non-abstract analogies. The corpus is therefore made up of Brian Greene's 2004 The Fabric of the Cosmos and Marcus du Sautoy's 2003 The Music of the Primes, although we have subsequently tested the methodology on other popular science texts.

Through the procedure we describe in 3.1 below to analyse metaphor and analogy in these texts, we identify a range of domains which are unusually frequent in these texts and which are not pertinent to their subject matter (that is, not in the areas of physics, mathematics or general science). We then demonstrate in the remainder of section 3 that these domains are those analogies used systematically and consistently across the texts to elucidate and explicate the abstract concepts the books are focused on discussing. In order to do this, we identify all the semantic domains mentioned in these texts at very high levels of precision, using an annotation system built around the unprecedented detail found in the database of the Historical Thesaurus.

2. Semantic Annotation
Semantic tagging and annotation is, we argue, the best solution we have to address the problem of searching and aggregating large collections of textual data: at present, historians, literary scholars and other researchers must search texts and summarize their contents based on word forms. These forms are highly problematic, given that most of them in English refer to multiple senses – for example, the word form "strike" has 181 Historical Thesaurus meaning entries in English, effectively inhibiting any large-scale automated research into the language of industrial action; "show" has 99 meanings, prohibiting effective searches on, say, theatrical metaphors or those of emotional displays. In such cases, much time and effort is expended in manually disambiguating and filtering search results and word statistics.

To resolve this problem, we use in this paper an early version of the Glasgow-Lancaster Semantic Annotation System, which we are currently developing at both of those universities. GL-SAS is a tool for annotating large corpora with meaning codes from the Historical Thesaurus, enabling us to search and aggregate data using the 236,000 precise meaning codes in that dataset, rather than imprecise word forms. These Thesaurus category codes are over one thousand times more precise than USAS, the current leader in semantic annotation in English corpus linguistics.8 The system automatically disambiguates these word meanings using existing computational disambiguation techniques alongside new context-dependent methods enabled by the Historical Thesaurus' dating codes and its fine-grained hierarchical structure. With our data showing that 60% of word forms in English refer to more than one meaning, and with some word forms referring to close to two hundred meanings, effective disambiguation is essential to GL-SAS.

3. Results
3.1. Methodology
The 600,000 word corpus we outline above were lemmatised and then processed through our annotation system, resulting in texts with each word being annotated with a Historical Thesaurus meaning code. We then aggregated those codes into a dataset which summarised the frequency of each meaning code in the text, and took that frequency list and compared it to a reference corpus made up of a 14m word corpus of random selections from Wikipedia, to provide a comparison against standard expository text. Our comparison was based on a log-likelihood significance measure,9 which identifies, to an acceptable degree, those semantic domains which are mentioned unusually frequently in our popular science texts by comparison to the reference corpus, and therefore indicates a text's "key" domains (where the log-likelihood values are greater than around 20)10 - those domains which reflect what a text is "about".11

3.2. The Fabric of the Cosmos
Brian Greene’s 2004 The Fabric of the Cosmos discusses theoretical physics and its relation to the concepts of space and time. Its key semantic domains are given in Table 1:

HT Category Category Name Log-Likelihood Value
01.05.07 Space 13655.8
01.05.07.01 Distance 6344.8
01.04.07.05.04.08 Photon 4912.5
01.05.06.07 Computation of time 3603.5
01.02.09.15 Spinning textiles 3193.5
03.11.03.01.08.02 Stringed instruments 2277.7
03.11.03.02.09.14 Pattern/design 1949.8
01.02.09.14.01.03 Woven fabric 1922.2
While the first four domains are within the Thesaurus categories which refer to the text's topic, and therefore expected, the next four (in bold) are not immediately relevant to the book's topic. Looking for these domains in the text itself, chunked into 591 smaller files of 320 words each, we get a distribution like this:

Fig. 1: Analogical textual clusters in The Fabric of the Cosmos, shown by frequency of key semantic domains

(Here, the Thesaurus codes have been replaced by words representing those categories, for ease of reading.)

The peak three-quarters of the way through the text indicates an area rich in mentions of textiles, and looking at this point in the text we find passages such as:

Since we speak of the ‘fabric’ of spacetime, the suggestion goes, maybe spacetime is stitched out of strings much as a shirt is stitched out of thread. That is, much as joining numerous threads together in an appropriate pattern produces a shirt’s fabric, maybe joining numerous strings together in an appropriate pattern produces what we commonly call spacetime’s fabric. Matter, like you and me, would then amount to additional agglomerations of vibrating strings.12

The areas we have identified through the log-likelihood analysis are therefore those areas rich in metaphors of fabric and strings (as other examples show) which are used by the author to discuss physics. We can therefore use this technique to pinpoint areas of significant use of metaphor or analogy in a text.

3.3. The Music of the Primes
As a check of the methodology, the same technique shows that in this particular book, which discusses prime number theory, there are highly key domains of travel and landscape in use alongside mathematical terms. Going to sections particularly rich in these domains gives analogical content over a long stretch, introduced by the following extract:

Gauss’s two-dimensional map of imaginary numbers charts the numbers that we shall feed into the zeta function. The north-south axis keeps track of how many steps we take in the imaginary direction, whilst the east west axis charts the real numbers. We can lay this map out flat on a table. What we want to do is to create a physical landscape situated in the space above this map. The shadow of the zeta function will then turn into a physical object whose peaks and valleys we can explore.13

4. Conclusion
We therefore demonstrate in this paper the use of a very fine-grained semantic annotation system, and establish the utility of such detailed annotations by describing a digital technique for discovering not only the existence of systematic metaphorical content but also its location and where it clusters. We believe that this result is significant in its own right, particularly for scholars of metaphor or cognitive linguistics, but we will also show that this represents only one of the uses to which highly-granular semantically annotated data can be put.

References
1. Kay, C., J. Roberts, M. Samuels, and I. Wotherspoon (eds). (2009). Historical Thesaurus of the Oxford English Dictionary. Oxford: Oxford University Press. See also historicalthesaurus.arts.gla.ac.uk.

2. Rayson, Paul. (2008). From Key Words to Key Semantic Domains. International Journal of Corpus Linguistics 13.4. 519-549.

3. Lakoff, George & Mark Johnson. (1980). Metaphors We Live By. Chicago: University of Chicago Press.

4. Gibbs, Raymond W. (2006a). Introspection and Cognitive Linguistics: Should We Trust Our Own Intuitions? Annual Review of Cognitive Linguistics 4(1). 135-151.

5. Evans, Vyvyan & Melanie Green. (2006). Cognitive Linguistics: An Introduction. Edinburgh: Edinburgh University Press. Page 780.

6. Alexander, Marc & Christian Kay. (2011) [2010]. Mapping Metaphors Across Time with the Historical Thesaurus. Conference paper at Helsinki Corpus Festival: The Past, Present, and Future of English Historical Corpora, University of Helsinki, Finland. Based on an earlier paper at The 3rd UK Cognitive Linguistics Conference, University of Hertfordshire.

7. Alexander, Marc. (2011). Meaning Construction in Popular Science An Investigation into Cognitive, Digital, and Empirical Approaches to Discourse Reification. University of Glasgow: Ph.D. thesis.

8. ucrel.lancs.ac.uk/usas

9. Dunning, Ted. (1993). Accurate methods for the statistics of surprise and coincidence. Computational Linguistics 19(1). 61–74.

10. Rayson, Paul, Damon Berridge, & Brian Francis. (2004). Extending the Cochran Rule for the Comparison of Word Frequencies between Corpora. 7th International Conference on Statistical Analysis of Textual Data.

11. McIntyre, Dan & Brian Walker. (2010). How can Corpora be Used to Explore the Language of Poetry and Drama? In Anne O’Keeffe & Michael McCarthy (eds.), The Routledge Handbook of Corpus Linguistics. London: Routledge. 516-530.

12. Greene, Brian R. (2004). The Fabric of the Cosmos: Space, Time and the Texture of Reality. Alfred A Knopf: New York. Page 486-7.

13. du Sautoy, Marcus. (2003). The Music of the Primes: Why an Unsolved Problem in Mathematics Matters. London: Harper Perennial. Page 85.