1Stylistic analysis is essentially a comparative process. An automatic method of comparing bodies of text in order to characterize their ‘differentness’ is provided by the Wmatrix software developed by Paul Rayson (for details, see Rayson 2008; also http://ucrel.lancs.ac.uk/​wmatrix/​). For my purposes, as I am interested here in the stylistic analysis of a single text, the comparison will be between that single text (the focal text) and a corpus (the reference corpus). The question is: How far can this automated procedure help to identify salient features of literary style? How far can phenomena which are statistically salient in the text be considered foregrounded from the point of view of literary theme and appreciation?

1. Virginia Woolf’s ‘The Mark on the Wall’: Our focal text

2‘The Mark on the Wall’, written in 1917, might be described as a story in which nothing happens – where nothing happens, that is, except in the mind of the narrator. (We use the term ‘narrator’ here, although it is the inner voice of the narrator that we experience throughout the story.) The narrator, sitting down after tea, notices a mark on the wall. Her mind explores in a myriad ways the significance of that mark – what it might be, and where it came from. This train of thought leads her by digressions of memory and imagination to such topics as the preceding occupants of the house – the nature of life – life after death – the oddities of experience – the mysteries of existence – always following the stream of the narrator’s consciousness. Every so often, however, the narrator’s attention comes back to the mark on the wall – and at last, she learns what it is. To give the flavour of the text, here are its opening paragraph and the final few lines:

Opening paragraph:

Perhaps it was the middle of January in the present year that I first looked up and saw the mark on the wall. In order to fix a date it is necessary to remember what one saw. So now I think of the fire; the steady film of yellow light upon the page of my book; the three chrysanthemums in the round glass bowl on the mantelpiece. Yes, it must have been the winter time, and we had just finished our tea, for I remember that I was smoking a cigarette when I looked up and saw the mark on the wall for the first time. I looked up through the smoke of my cigarette and my eye lodged for a moment upon the burning coals, and that old fancy of the crimson flag flapping from the castle tower came into my mind, and I thought of the cavalcade of red knights riding up the side of the black rock. Rather to my relief the sight of the mark interrupted the fancy, for it is an old fancy, an automatic fancy, made as a child perhaps. The mark was a small round mark, black upon the white wall, about six or seven inches above the mantelpiece.

Ending:

... – but something is getting in the way ... Where was I? What has it all been about? A tree? A river? The Downs? Whitaker's Almanack? The fields of asphodel? I can't remember a thing. Everything's moving, falling, slipping, vanishing ... There is a vast upheaval of matter. Someone is standing over me and saying:

‘I'm going out to buy a newspaper.’
‘Yes?’
‘Though it's no good buying newspapers. Nothing ever happens. Curse this war; God damn this war! ... All the same, I don’t see why we should have a snail on our wall.’ Ah, the mark on the wall! It was a snail.

2. Comparing the focal text and a reference corpus

3The focal text, ‘The Mark on the Wall’, will be compared quantitatively with a reference corpus which should be representative to some degree of the variety from which the text is taken. However, there are obviously different degrees of generality in defining the language variety meant to act as a reference standard. We have decided to use three different ‘reference varieties’ (the choice being determined, obviously, by the availability of suitable texts in electronic form):

5(B) A more general corpus of fiction, consisting of category K (General Fiction) in the Fiction subcorpus of the B-LOB corpus (a member of the Brown Family of corpora representing written (printed) British English over the period 1928-1934). This is more widely representative than (A), as it contains 29 text samples by different authors. However, it is less closely matched than (A) in time of publication, as the samples date from 1928-34.

101. Building the data: corpus design and compilation (in the case of our Wmatrix investigation, this has already been sufficiently described in terms of our focal text and the three reference corpora).

112. Annotating the data: analysing the corpus linguistically, using particular annotation tools: in the case of Wmatrix, the two annotation tools used are

(a) the CLAWS part-of-speech (POS) tagger, and
(b) the USAS semantic domain tagger.

133. Retrieving: extracting from the text data some analytic results, which may be displayed in a variety of formats for inspection or further processing. In the Wmatrix analysis, we are interested in three more or less standard listing formats:

(a) concordances, which list the occurrences of a particular word (or other feature) in their contexts of occurrence,
(b) frequency lists, which list words (or other features) in order of their frequency in a particular body of text data, and
(c) keyness lists, which list words (or other features) in order of their keyness in a given textual comparison.

144. Interpreting: This is the only stage of the process which is essentially non-automatic (‘manual’), although it can be aided by automatic procedures such as using the ‘Sort’ and ‘Collocation’ facilities of corpus software. Whereas stages 3(a) and 3(b) above are quantitative, stage 4 is qualitative: it makes use of the human ability to interpret texts and to explain the phenomena observed in them. In the case of the Wmatrix investigation, we may be interested here in examining the textual material more carefully, using especially the concordance displays, in order to explain the stylistic phenomena observed in the analysis.

15We now have to focus on the third, ‘Retrieving’ stage above, in order to explain in a little more detail what the software does. At the same time, we will avoid going into technical detail, which can be studied in Rayson (2008) and on the UCREL webpages already cited.

16To take the most basic case, the list of keywords is arrived at as follows:

i) Two word frequency lists are compiled: a list for the focal text (‘List X’), and a list for the reference corpus (‘List Y’).

• 9 The keyword list can include words which have 0 occurrences in List X or List Y. Negative keywords (...)

ii) List X and List Y are compared. This means that each word in List X is measured in terms of comparative frequency with the same word in List Y.9 ‘Comparative frequency’ means that the raw count of a word’s frequency is adjusted to a standard measure relative to corpus size, which in Wmatrix is the number of occurrences of the word as a percentage of all occurrences of words in the text/corpus.

• 10 The significance measure used in Wmatrix is log likelihood, which is considered preferable to the m (...)

iii) Each word’s keyness in the focal text is measured by a statistical formula, which calculates the degree to which the word is either ‘over-represented’ or ‘under-represented’ in this text, as measured against the reference corpus. The normal understanding of keyness is that the word is over-represented, that is, is relatively more frequent in the focal text than in the reference corpus, to a certain high degree of statistical significance.10

iv) The words in List X are re-ordered in order of keyness. This means that the words at the top of the list are most distinctive of that text.

17Concordance, frequency and key-feature lists of POS tags and semantic tags are extracted in the same way as the word lists described in 3(a)-(c) above. There are no particular difficulties in this, as the annotation (tagging) has meant that each word in each text is accompanied by label giving its grammatical and semantic classification.

3. Results: keywords, key POS tags, and key semantic domain tags

18To begin with, Table 3 shows the top 12 keywords, in order, when ‘The Mark on the Wall’ is compared with each of the reference corpora.

19Perhaps the most striking result is the amount of agreement that the three reference corpora show, in spite of their very different composition. Comparisons with A and B share all of their top 10 key words (out of 12); A and C share 9 of the 12; and B and C share 11. Perhaps this is a mild reflection of the degree of generality of the corpora. It seems that the keyword methodology is robust in showing up the ‘differentness’ of a text without respect to the exact make-up of the reference corpus.

20It is not surprising that mark is the ‘keyest’ of the keywords: it represents the theme of the story, as to a lesser extent does wall. These are words that, as we might imagine, occur relatively rarely in the reference corpora, and therefore their repeated use in ‘The Mark’ is salient, both statistically and thematically. Of the other words which occur in all three comparisons, one (typically used in the generic human sense) is perhaps a personal stylistic favourite of Virginia Woolf, representing as it does the objectification of the narrator’s personal experiences, as illustrated in the following passage:

because one will never see them again, never know what happened next ... as one is torn from the old lady about to pour out tea and the young man about to hit the tennis ball in the back garden of the suburban villa as one rushes past in the train.

21We will not dwell on the items in this list, some of them uncommon words, like Precedency, which gain idiosyncratic prominence in Woolf’s narrative – see Leech (2008: 168-71) for further discussion. But there are some interesting points to observe about the similarities and differences between the lists. For example, is is very much overrepresented when compared with the fictional reference corpora (but not with the more general reference corpus C), and this is probably because Woolf, in capturing the immediacy of the interior monologue, tells much of her story in the historic present, instead of using the past tense narrative convention of the majority of fictional writers. This choice of the present tense is understandably not so salient when compared with the full range of written texts (scientific, journalistic, etc.) in the 1901±3 corpus. On the other hand, the pronoun I, frequent in Woolf’s first-person narrative, stands out as over-represented when compared with the cross-section of written texts in 1901±3, but is less salient in the two fiction reference corpora, where first person reference occurs frequently, for example in dialogue.

22We move on now to the lists of key part-of-speech tags, reflecting the different grammatical choices made by Virginia Woolf as compared with the writers in the other reference corpora.

Key: AT – article neutral for number; chiefly the definite article the. AT1 – singular article; chiefly the indefinite article a/an
DDQ – wh-determiner or wh-pronoun (e.g. what, which)
IO – the preposition of
NN2 – plural common noun (e.g. tables, women, thoughts) NPD1 – singular weekday noun (e.g. Sunday, Monday)
PN1 – singular indefinite pronouns (e.g. one, anything, nobody) PNX1 – indefinite reflexive pronoun (i.e. oneself)
PPH1 – third person personal pronoun it
PPIS1 – the first person subject pronoun I
RGQ – wh-adverb of degree (how when modifying another word) RPK – about used in the expression be about to.
VBZ – present tense –s form of the verb to be (i.e. is)
VVG – ing-form of lexical verb (e.g. saying, wishing)
VVZ – present tense lexical verb ending in –s (e.g. says, wishes) VV0 – present tense lexical verb not ending in –s (e.g. say, find)

23The amount of shared ‘key tags’ between the comparisons here is the same: nine tags are shared by the top twelve in A, B and C. What brings A and B closer together, however, is the fact that the top four tags are the same and in the same order. As mentioned above, the present tense (represented in the keyness of the s-form of lexical verbs VVZ as well as of VBZ and VV0), is a distinctive feature of ‘The Mark’, as opposed to fiction written in the more conventional past-tense narrative. More difficult to explain is the second-keyest tag, the plural noun tag NN2; however, the following passage illustrates how W oolf’ s style may favour plural nouns in describing the multitudinous particularity of her experiential world:

let me just count over a few of the things lost in one lifetime, beginning, for that seems always the most mysterious of losses – what cat would gnaw, what rat would nibble – three pale blue canisters of book-binding tools? Then there were the bird cages, the iron hoops, the steel skates, the Queen Anne coal-scuttle, the bagatelle board, the hand organ – all gone, and jewels, too. Opals and emeralds, they lie about the roots of turnips.

24It is striking, also, that this passage contains four examples of another key tag, IO (representing the preposition of in the tagging system). The word, of course, has many functions – but its main function, in the most general terms, is to signal the interconnectedness of things. It is noticeable in this list that IO stands out as a key tag in relation to the fictional reference corpora A and B, but not in relation to the most general reference corpus C, which is predominantly non-fictional. Elaboration of noun phrases by means of of is likely to be a characteristic of informational texts, which oddly here seem to be more akin to Woolf’s own elaborative style. Of the other key tags, we will comment only on PN1, PNX1 and RGQ. PN1 chiefly represents the pronoun one already noted as favoured in ‘The Mark’; and PNX1, normally a very rare tag (representing the word oneself) stands out in this text even though there are only two occurrences of it. RGQ represents the adverb How as a modifier, in this text especially associated with exclamations:

How readily our thoughts swarm...
How shocking, and yet how wonderful it was to discover...
How peaceful it is down here.

26The third level of analysis, that of semantic tagging, produces lists of key semantic domains as follows:

27Key semantic domains tell us something about the ‘aboutness’ of texts, rather than about their stylistic characteristics in the strict sense. They are therefore less relevant to style, and there is less agreement between the different reference corpus comparisons: only half of the key semantic domains listed are shared by all three lists. On the other hand, there are some features which are salient not so much in style as in the authorial world view. The domain of colour is high on the list of key domains in all three comparisons, as are the domains relating to the natural world: ‘Plants’ and ‘Living creatures’. Readers of Virginia Woolf will probably agree that these traits have a ‘key’ role in her writing. Other, more abstract domains are more difficult to interpret, but arguably reflect her exploration of the nature of reality and the ontological concerns of her writing. At the other extreme, the domain of ‘Smoking’ must be regarded as incidental to the text, in that it results from the semantic tagging of four words only: one of the drawbacks of choosing such a short focal text for analysis is that such haphazard results can occur. Here is another excerpt, which contains a reference to smoking, but is also relevant to some other key features:

Even so, life isn’t done with: there are a million patient watchful lives for a tree, all over the world, in bedrooms, in ships, on the pavement, lining rooms, where men and women sit after tea, smoking cigarettes. It is full of peaceful thoughts, happy thoughts, this tree.

28This passage illustrates representation of some of the key features high on the list above: Plants (tree), Life and living things (life, lives), Mental object; conceptual (thoughts), Parts of buildings (bedrooms, rooms). Obviously there is much more to be said about this story, and the extent to which the ‘key’ analysis succeeds in highlighting stylistically important features. But the main point of this section of my paper has been to illustrate the potential of such analyses, using a chosen text and three alternative reference corpora of different generality.

4. Conclusion

29In this article I have briefly explored a method of computer-aided stylistic analysis, involving the comparison of a focus text and one or more reference corpora. The technique is to employ the WMatrix software to identify and display items in order of keyness, or distinctiveness in the focal text, as contrasted with the reference corpus, measured in terms of the significance ratio of Log Likelihood. The main difficulty with this was the relative shortness of the ‘The Mark’, which gave undue prominence to some features occurring only a few times.

30It is worthwhile, finally, noting some of the limitations as well as the future possibilities of this stylistic method. It is only too obvious, to begin with, that this type of analysis when applied to very large quantities of electronic text would be virtually impossible without the power of the modern computer. The great advantage of the techniques illustrated here is that they can be carried out automatically and at great speed. Wmatrix also shows great adaptability to the use of a wide range of corpora. The variety of corpora capable of being used is limited only by the user’s ability to assemble the corpora and load them as ‘personal folders’ onto the Wmatrix website.

31The corresponding disadvantage is that any activity involving human scrutiny of the data is immensely slow by comparison. Although POS tagging and semantic tagging are relatively accurate, there are still plenty of ‘mistakes made by the computer’ that ideally need to be manually checked. Further, although at present Wmatrix can operate with grammatical tags and semantic tags, there are many other levels of analysis that at present it cannot undertake – most importantly, parsing: the systematic syntactic analysis of a text in terms of phrases, clauses and so forth. There are also some more meaning-oriented stylistic analytic tasks (e.g. identifying metaphor or irony) that cannot (yet) be achieved by a computer.

32The present situation, then, is that certain tasks can be undertaken fast but fallibly by computer, while other tasks can be undertaken more reliably but more slowly by human beings. Wmatrix already has the advantage that it can undertake a multi-level linguistic analysis of English corpora. Some of the items highlighted by the statistical analysis can clearly be seen to have thematic and literary significance, although without the help of WMatrix, they probably would not have been noticed.

33One of the things suggested by this analysis is that there is no need to worry unduly about choosing an exactly appropriate reference corpus. None of the three reference corpora used in this experiment were ideal for the purpose, and yet the differences between the results of using the different reference corpora were rather minor.

34Obviously this small experiment is far from exhaustive. I believe that present results, although lacking in detail, are promising, and that we can look forward to a future in which more revealing analyses of style can be achieved by computer at a more abstract level.