Roger Garside, Geoffrey Leech and Tamás Váradi

1. The Lancaster Parsed Corpus is a corpus of British English sentences excerpted from printed publications of the year 1961. The Parsed Corpus is a subset of the Lancaster-Oslo/Bergen Corpus (= LOB Corpus) (see S. Johansson, G. N. Leech and H. Goodluck, Manual of Information to accompany the Lancaster-Oslo/Bergen Corpus of British English, Department of English, University of Oslo, 1978). Each sentence in the Parsed Corpus has undergone a syntactic analysis in the form of a phrase marker, or labelled bracketing, following the scheme of analysis outlined in Part III below.¹

TABLE I: CONTENTS OF LANCASTER PARSED CORPUS

Text categories No. of sentences No. of words

A Press: reportage 728 8188

B Press: editorial 781 8832

C Press: reviews 594 7145

D Religion 721 8803

E Skills, trades and hobbies 563 8347

F Popular lore 584 7526

G Belles lettres, biography, essays 406 5987

H Miscellaneous (government documents, etc) 441 5798

J Learned & scientific writings 450 7537

K General fiction 1119 11279

L Mystery & detective fiction 1188 13010

M Science fiction 769 7525

N Adventure & western fiction 1344 12919

P Romance & love story 1369 13826

R Humour 720 8018

————— ——————

TOTAL: 11827 134740

The categories A-R above are those of the original LOB Corpus, which in turn were based on those of the Brown Corpus of American written English, for which the LOB Corpus was a British matching equivalent.²

3. The word-tagged LOB Corpus, which has provided the input to the Parsed Corpus, consists of 1,013,737 tagged running words, divided into 500 samples of approximately 2,000 words each.³ The Parsed Corpus is a sub-corpus of it, consisting of 13.29% of the tagged LOB Corpus.

4. The Parsed Corpus contains sentences from the first 10 text samples of each Text Category of the LOB Corpus, except for Categories M and R, which contained only 6 and 9 text samples (respectively) in the original corpus. Thus, roughly speaking, the Parsed Corpus represents, in smaller quantities, the range of text samples provided in the original LOB Corpus. However, there are qualifications to be made to this statement, as indicated in the paragraphs below.

5. Since the Parsed Corpus samples are restricted to the first ten (or less) samples in each Category, they are in fact more limited in genre/domain than the original corpus. These limitations are indicated in the descriptions in Table II below, which also lists the number of text samples per category both in the Parsed Corpus, and in the LOB Corpus:

TABLE II: TYPES AND NUMBERS OF TEXT SAMPLES IN THE PARSED CORPUS

Text Categories in Parsed Corpus Number of text samples in:

(More precisely defined than in Table I) Parsed Corpus (LOB)

A National daily papers (reportage) 10 (44)

B National daily papers (editorial) 10 (27)

C National daily papers (reviews) 10 (17)

D Books on religion 10 (17)

E Homecraft, handiman, and hobbies publications 10 (38)

F Popular politics, psychology, sociology 10 (44)

G Biography, memoirs 10 (77)

H Government documents 10 (30)

J Writing on natural sciences 10 (80)

K Novels (general fiction) 10 (29)

L Novels (mystery and detective) 10 (24)

M Novels and short stories (science fiction) 6 (6)

N Novels (adventure and western) 10 (29)

P Novels (romance and love story) 10 (29)

R Novels and articles from periodicals (humour) 9 (9)

——— ———

TOTAL number of text samples: 145 (500)

6. As will be evident from Tables I and II, the Parsed Corpus does not contain all the sentences of the c.2000-word text extracts in the LOB Corpus which were used as input. In fact, the sentences included in the Parsed Corpus are considerably shorter, on the average, than the average sentence length for the LOB Corpus as a whole (the average for Parsed Corpus is 11.39 words per sentence, as compared with an average of c.19 words per sentence for the entire LOB Corpus). On the whole, then, longer sentences have been excluded from the Parsed Corpus, with the result that the Parsed Corpus no longer contains LOB text extracts in their entirety.

7. The reason for the omission of longer sentences is as follows. Our original hope was that we would be able to parse the whole of the LOB Corpus automatically. The prototype probabilistic parser developed for this purpose is described in R. Garside, G. Leech and G. Sampson (eds), The Computational Analysis of English: a Corpus-based Approach, London: Longman, 1987, Chapter 6. In practice, this goal proved too ambitious, and we ran the parser over only the 145 text extracts as listed in Table II. Even then, the parser was unable to achieve a parse of most sentences over 20-25 words in length. These unparsed sentences were therefore omitted from the Parsed Corpus. For the remaining sentences, which had been parsed, we accepted the "winning parse" (that to which the parser had assigned the highest probability) as the input to a manual post-editing stage. During manual post-editing, many sentences which the parser had incorrectly parsed (according to the procedure mentioned above) were corrected by hand, and a number of final checks were made to ensure that the Parsed Corpus is reasonably error-free. Therefore the Lancaster Parsed Corpus can be regarded as a treebank broadly representative of the syntax of written (printed) English across a great variety of styles and text types. It may hopefully be used, for example, as a testbed for wide-coverage general-purpose grammars and parsers of English, as well as for quantitative linguistic studies of English syntax.

8. The fact that long sentences are under-represented in the corpus does, of course, limit the corpus's value as a testbed. In coverage of grammatical rules and structures, however, it will probably be more adequate than any other treebank publicly available to date.

9. In another respect, too, the corpus is unrepresentative: it cannot be regarded as a balanced sub-sample of the LOB Corpus. Hence any quantitative comparison between features observed in the parsed corpus and other features independently observed in the LOB Corpus are likely to be unreliable. The reason for this is (as can be gleaned from Tables I and II) that the Text Categories are not represented in the Parsed Corpus according to their proportion of the full LOB Corpus. For example (to take the most extreme cases), in Category G, the Parsed Corpus contains only 3.84% of the original LOB Corpus material, whereas in Category M, it contains 62.13%. Disparities such as this are due to (a) the greater average length of sentences in some categories than in others, and (b) variation across Text Categories in the number of text extracts in the LOB Corpus and in the number of text samples in the Parsed Corpus.

10. In practice, the consequences of the factors just mentioned are that the Parsed Corpus contains a higher proportion of material from imaginative (largely fictional) texts (Categories K-R), whereas the LOB Corpus as a whole has a higher proportion of expository text material (Categories A-J). The figures are given in Table III:

TABLE III

Expository Categs. (A-J) Imaginative Categs. (K-R)

LOB Words: 756,293 Words: 257,444

Corpus %age of Corpus: 74.60 %age of Corpus: 25.39

Parsed Words: 68,163 Words: 66,577

Corpus %age of Corpus: 50.59 %age of Corpus: 49.41

Acknowledgements

We are grateful for help received from the following sources:

Notes to Part I

The text is contained in plain ASCII text files. Each sentence is in a separate paragraph and is followed by an empty line. At the head of each paragraph is a line that serves to identify the sentence that follows. It contains an alphanumeric code standing for the genre and text sample in the LOB Corpus that the sentence belongs to. Next to this, is the ordinal number of the sentence in the present corpus. Numbering is restarted at the beginning of each genre. To ensure readability of text on 80 column wide computer screens, text lines have been formatted so that they do not reach over 72 characters. Below is a sample sentence:

B06 666

[S[N it_PP3 N][V would_MD become_VB V][J easy_JJ J][Ti&[Vi

to_TO be_BE Vi][J cynical_JJ J][Ti+ and_CC [Vi to_TO

despair_VB Vi]T+]Ti&] ._. S]

Each word in the text is followed by an underscore character and a sequence of symbols (normally capital letters) which represents a wordtag, i.e. a label giving the grammatical class of a word. (See Part II for a full list.) For example, in the sample text above PP3 means `third person personal pronoun' and MD means `modal auxiliary'. For the present purposes punctuation marks are treated as words and have their own tags.

The syntactic structure of the sentence is laid out in the form of labelled bracketing. The type of grammar used in the parsing was a Phrase Structure Grammar, in which the structure of a sentence can be shown as a tree, with `S' (= `Sentence') as the root of the tree and the wordtags `PP3', `MD', etc. as the leaves of the tree. Here, for instance, is the sentence above represented as a tree:

S

PP3 MD VB JJ TO BE JJ CC TO VB .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

It would become easy to be cynical and to despair .

Alternatively, the same phrase structure (omitting the wordtags at the bottom) can be represented as a labelled bracketing as follows:

[S [N] [V] [J] [Ti& [Vi] [J] [Ti+ [Vi] Ti+] Ti&] S]

The interpretation of the labelled bracketing on the output is as follows:

`[' represents the opening of a constituent, and `]' represents the closing or completion of a constituent. The symbols alongside these brackets are labels for constituents. For example, `[V' means `a verb phrase opens here' and `V]' means `a verb phrase closes here'. The labels are called constituent tags, as distinct from the wordtags, which label words and are attached to them directly. Constituent tags label higher constituents such as sentences, clauses and phrases, and are attached to the corresponding brackets.

For example, the following are the meanings of the constituent tags in our sample sentence:

You will notice that these constituent tags sometimes consist of just one capital letter, and at other times there are additional symbols, such as a lower-case letter, an `&' or a `+'.

The capital letters indicate the major class of the constituent that the tag labels. Where they occur, additional lower-case letters indicate a subclassification. A `&' represents a compound constituent (e.g. `N&' indicates a compound noun phrase, in which two or more noun-phrase-like structures are joined by a coordinating conjunction). A `+' represents a coordinated constituent, e.g. `N+' indicates a coordinated noun phrase.

The grammar on which this parsing scheme is based is somewhat simplified. It is simplified (i) because very few subclassifications are included (`Vi' being an exception), and (ii) because the constituent tags indicate formal categories only. They do not represent functional concepts such as `subject', `object', `complement' and `adverbial'.

It is important to note that this parsing scheme treats coordination in a rather unorthodox way (which is, nevertheless, convenient for automatic parsing). Suppose we take a typical compound noun phrase such as:

`my brother and his wife'

Most phrase structure grammars would represent this construction roughly in the following way:

[N& [N+ my brother N+] and [N+ his wife N+] N&]

The construction is symmetrical and may be shown in the form of a tree diagram as follows:

N&

my brother and his wife

However, the present parsing scheme treats the second coordinated noun phrase as subordinate in relation to the first, as follows:

[N& my brother [N+ and his wife N+] N&]

Notice that the `and' is treated as part of the coordinated noun phrase that it introduces,

As a tree-diagram, this looks like this:

N&

my brother and his wife

If the compound noun phrase contains three (or more) noun-phrase-like structures, the second and subsequent ones are treated as subordinated, as in:

`my brother and my sister-in-law and their children'

This is analyzed:

It is common, however, for the first `and' to be omitted in such structures, and in that case, the constituent tag `N-' is used for a coordinated noun phrase not introduced by `and':

The same convention is used in constructions with parataxis, i.e. constructions where the `and' is omitted altogether:

One more oddity about coordination is that if a sentence begins with a coordinating conjunction, the whole sentence is tagged `S' not `S+'. E.g.:

[S but in his letter 12 days later he retracted. S]

Sq and Si

Pattern A:

"Nothing will change my mind", said Pat.

Pattern B:

Pat said, "Nothing will change my mind".

In these cases, the direct speech is analyzed as [Sq]:

Pattern A: [S "[Sq]" , [V] [N] S]

Pattern B: [S [N] [V] , "[Sq]" S]

Pattern C:

"Nothing," said Pat, "will change my mind".

In this case, Sq isn't used. Instead, the reporting clause is treated as an Si:

Pattern C: [S "..." , [Si] , "..." S]

Here is a further example of the use of Si:

S&, S+ and S-

`S&' represents a compound sentence, `S+' represents the second or subsequent conjoin of a compound sentence, if it begins with a coordinating conjunction, and `S-' represents such a conjoin when it does not begin with a coordinating conjunction. (See the discussion on coordination in 3. above).

`Fa' is a finite adverbial clause (e.g. a finite subordinate clause of time, of condition, of reason etc.)

`Fc' is a comparative clause, normally beginning with `than' or `as'.

E.g.: `I know [Fn that you saw them Fn].'

E.g.: `the house [Fr in which I was born Fr]'

N.B. a `fused' or `nominal relative clause' as in `I will do [what you want]' is treated as `Fn'.

F&, F+, F-, etc.

These tags, which will also occur in combination with the letters `a', `n', `r' etc., are used for coordinated finite subordinate clauses in accordance with the rules detailed in 3.

Nonfinite clauses are indicated by `T'. However, `T' does not normally occur alone. It is combined with the subscripts below.

`Ti' stands for a to-infinitive clause (e.g. an infinitive construction in which to+infinitive may or may not be followed by an object, a complement and/or adverbials).

`Tg' stands for an -ing clause (i.e. a participial or gerundival construction in which the -ing form of the verb may or may not be followed by an object, a complement, and/or adverbials).

`Tn' stands for a past participle clause (i.e. a construction in which the past participle form of the verb may or may not be followed by an object, a complement and/or abverbials).

`Tb' stands for a `bare infinitive clause' (i.e. a construction in which the `bare infinitive' — infinitive without to — may or may not be followed by an object, a complement and/or abverbials).

`Tf' is used as a variant of the infinitive clause, where the subject of the infinitive is introduced by `for'.

N.B. Nonfinite clauses generally have no subject: but it is also possible for a subject to occur;

`W' stands for a nonfinite or verbless clause introduced by with.

`L' stands for a verbless clause not introduced by with or by a subordinating conjunction.

Note: If an adverbial verbless clause or nonfinite clause is introduced by a subordinating conjunction (e.g. if, when), it is treated as a `Fa':

If an adverbial verbless clause or nonfinite clause is introduced by a wh- word why, what, how, it is treated as a `Fn':

`V' means `finite Verb Phrase', in the narrow sense, in which `verb phrase' excludes objects, complements, etc. Thus `V' may include simple verb phrases such as is, have, did and also more complicated ones with modals, progressive aspect, perfect aspect or passive.

Vo and Vr

In general, no subscript is used with `V'. However, `Vo' and `Vr' are exceptions. They are used when a verb phrase is split into two parts by subject-auxiliary inversion. The first part is labelled `Vo' (o = `operator') and the second part is labelled `Vr' (r = `remainder'). E.g. in `Have you seen Mary?' `have' is `Vo' and `seen' is `Vr'.

Note that `V' includes the negative word `not' as well as adverbs. E.g. the whole of `have not seen' (or `haven't seen' or `have recently seen') is a `V'. But if the subject noun phrase occurs between the auxiliary and the main verb, this is treated as a separate noun phrase. Accordingly, `have you seen' consists of `Vo' followed by `N' followed by `Vr'.

Vi, Vg, Vn

These are labels for nonfinite verb phrases, i.e. verb phrases which are the verb phrases of nonfinite clauses `Ti', `Tg' or `Tn'.

means `to-infinitive verb phrase', e.g. `to eat' or `to have eaten'.

means `-ing participle verb phrase', e.g. `eating' or `having eaten'.

means `past participle verb phrase', e.g. `eaten'.

N is the label for a noun phrase, whether it is a single word (such as the pronoun it) or a sequence of words.

In general, `N' has no subscripts. One major exception is `Na', which stands for a noun phrase marked as subject of the verb. In practice, `Na' almost always indicated one of the pronouns `I', `she', `he', `we', `they'. (N.B. `you' and `it' as subject are not marked `Na' because their status of subject is not unambiguously shown by their form.)

Another exceptional use of `N' + subscript, meaning a wh- noun phrase, such as `who', `which', `which car', `what time' etc.

`J' means an adjective phrase such as `happy', `very tall', `too happy for words', etc. If an adjective occurs as the head of a noun phrase, e.g. `the wealthy', `the unemployed', the phrase is marked `N' not `J'.

Here, as with `Nq', the `q' means `a phrase beginning with a wh-word', e.g. `Jq', an adjective phrase beginning with a wh-word, is in practice a phrase such as `How old'.

`P' stands for `prepositional phrase', e.g. `in London' or `on arriving at the station', `with it', `for what we are about to receive', i.e. a preposition followed by its complement or completive element. Prepositional phrases also sometimes contain adverbs like `just' in `just inside the door'.

stands for `prepositional phrase with a wh-word, e.g. `on whose behalf', `in which case', `for whom'.

stands for a `prepositional phrase beginning with the preposition of'.

`R' is the symbol for an adverb phrase, which may be a single word such as `there' or `quickly' or may be a sequence such as `quite often', `too fast', `further than I expected', etc.

stands for an adverb phrase beginning with a wh-word. This would include such phrases as `how' in `How do you feel?', or `how long' in `How long have you been waiting?'

`M' stands for a `numeric phrase' when such an expression is part of a noun phrase. Examples are `five thousand' in `five thousand young people'; `another hundred' in `another hundred calories'. Numeric phrases have a numerical word at their head (e.g. `hundred'), and consist of at least two words. (N.B. if numerical expressions such as `five thousand' occur on their own as noun phrases, they are labelled `N'.)

`D' stands for a `determiner phrase', i.e. a phrase consisting of at least two words, in which the determiner is a head, and which is part of a noun phrase. E.g. `too many' in `too many people'; `a good few' in `a good few people'. (N.B. `too many' or `a good few' on their own, acting as a noun phrase, are labelled `N'.)

stands for a determiner phrase (as defined above) beginning with a wh-word. E.g. `how many' and `how much', when they are part of a noun phrase, as in `How many apples (did you buy)?'

`G' stands for `genitive phrase' i.e. a phrase which consists of two or more words acting as the genitive in a noun phrase. E.g.`the earth's' in `the earth's rotation around the sun'; `my mother's' in `my mother's greatest wish'; `last Friday's' in `last Friday's Evening Standard'; `someone else's' in `someone else's bedroom'; `the Vicar of Bray's' in `the Vicar of Bray's famous dictum'.

`X' is the negative word `not' when acting as an independent element of clause structure; e.g. in `He told us [Fn what not to do Fn]', not follows the object of the subordinate clause what and precedes the verb phrase to do. Thus, the clause what not to do has the three constituents `Nq', `X', and `Vi'. Generally, `not' is part of the verb phrase (see under `V' above) and therefore does not require an `X'.

`E' is the label used for existential `there', i.e. the unstressed `there' in the `there is/are` construction. E.g. `[E There E] is nothing wrong'.

`U' is the tag used for an exclamatory word such as `Oh' or a grammatical isolate such as `yes' or `no'.

As already mentioned in 3. above, `&', `+', and `-' are suffixes used to mark constituents in coordinated constructions. They may be added to any of the constituent tags. Because of their uniform function, the above list does not contain these composite tags. Merely as an illustrative example, this is how `Tg&' and `Tg+' can be used as follows:

A problem of labelling arises when the two or more elements coordinated do not belong to the same class. In such cases, we resort to the device of a `slashtag', i.e. a composite tag in which the classes of the items concerned are separated by a slash (/). E.g.:

No slashtag is used when the first constituent tag and the second or subsequent tags all share the same capital letter. For example, there is no need to use `N/Na&' in the parsing of `my wife and I'

Slashtags can get very complicated when there is a sequence of three or more coordinated elements. Here is one complex, but very rare, example:

If coordination takes place on the word level only, (e.g. where two or more words inside the same constituent are linked by `and'), the coordination is indicated in the same way as described so far except for the fact that in such cases the relevant wordtags are attached to the brackets. E.g.

The reasoning behind this arrangement is that the set of coordinated words behave, with regard to the rest of the sentence, like a single word.

Word level coordinations of words of different classes are treated by slashtags in the same way as described in the previous section. E.g.

Note, however, that if a word is linked to a phrase of two or more words, then the word counts as a phrase for the purposes of coordination. In other words, if any of the coordinated elements is a phrase, the whole string can no longer be a word level coordination. E.g.

The sequence `[very young and friendly] dog', on the other hand, would be ambiguous between coordination of words and coordination of phrases; word coordination (where `very' is understood to apply to both adjectives) is more likely.

The above problem about coordination brings up the more general question of when to use constituent tags.

In general, constituent tags are used whenever, to generalize about grammatical structure, we need to recognize constituents intermediate between the `root node' `S' and the `leaf nodes' represented by grammatical wordtags such as `NN', `VB' and `IN'. This general rule needs to be supplemented by the following rules which apply to the parsing of the Lancaster Parsed Corpus.

a) `the deep blue sea' and b) `her deep blue eyes'

A) means `the sea which is both blue and deep'. B) means `her eyes which are deep blue'. We could show the difference by treating `deep blue' as a constituent, with its own tag, in the second case:

b) `[N her [J deep blue J] eyes N]'

Of course, if `deep-blue' occurred hyphenated, which is quite likely, it would then be treated as a single word adjective (wordtagged `JJ'), rather than a phrase.

But the difference between a) and b) is difference only detectable by the criterion of meaning; it is not signalled by any special structural characteristics. Therefore, our parsing scheme does not differentiate between a) and b), and the constituent `[J deep blue J]' is not recognized.

On the other hand, `her very blue eyes' would be analyzed as having an adjective phrase (`J') in it, because in this case the adjective phrase can be distinguished structurally by the fact that there is a qualifying adverb (`QL') before the adjective `blue':

For example, in the sentence, d) `I saw her twice', each of the words is a separate element of sentence structure: `I' is the subject, `saw' is the verb element, `her' is the direct object, and `twice' is an adverbial. So each of these words is assigned a phrase tag:

If we draw a tree diagram of this (inserting the wordtags as well), the result is that the tree contains `unary branching' beneath each of the phrase tags:

S

Na V N R

PPIA VBD PP30 RB

I saw her twice

This seems wasteful and goes against the general requirement that constituent tags are inserted only when necessary. If we got rid of the phrase tags, the resulting structure would be much simpler.

S

PPIA VBD PP30 RB

I saw her twice

But this structure would fail to recognize the potential expansion of each of these words into phrases; e.g.:

In fact, the rule that elements of clause/sentence structure are given phrase tags also extends to E (the tag for existential `there') and to `U' (the tag for exclamations), even though these cannot be (easily) expanded into multi-word phrases.

Exception: However, coordinating and subordinating conjunctions do not count as clause/sentence elements for these purposes. When they introduce a clause or sentence, these words (wordtagged `CC' and `CS') are direct constituents of the clause or sentence without receiving tags of their own.

In other words, except for the `unary branching' allowed in accordance with Rule 2 above, `unary branching' is not tolerated. For example, in a noun phrase (`N'), there are various constituents of two or more words which are given tags:

`the city [P of New York P]'

`the [J most difficult J] job'

`[G last night's G] election'

i) Determiner phrases (`D')

`[D so much D] trouble'

j) Numeric phrases (`M')

`[M five hundred M] pounds'

k) Noun phrases (`N')

`a [N two seater N] sports car'

But if these two-words were replaced by single words, then the tags would no longer be used. Examples parallel to some of the examples above have no constituent tags:

g') `the hardest job'

h') `yesterday's election result'

i') `five pounds'

k') `a two-seater sports car'

Note that if the two words are joined by hyphenation, as in k'), this no longer counts as a phrase, but as a single word.

Note also that in k'), if there were no hyphenation, there would a structural rather than semantic reason for recognizing `two seater' as an `N': the phrase is marked structurally by the fact that the indefinite article (which is singular) is followed by the number `two' (which is plural).

Exception: There is one important exception to Rule 3: viz. that a preposition in a prepositional phrase is followed by a constituent (normally a noun phrase), which is labelled by a constituent tag, even when the constituent consists of one word. In l) and m)

l) `This present is [P for [N your sister N] P]'

m) `This present is [P for [N you N] P]'

the prepositional complements `your sister' and `you' are both treated as noun phrases, even though one of them consists of just a single word. The reason for this exception is that the nominal expression which follows a preposition has all the potentiality of a noun phrase.

Punctuation points and brackets (`.' `?' `!' `:' `;' `,' `-' `(' `)' ) are treated as words for the purposes of wordtagging. Contrary to orthodox practice, punctuation points are treated as parts of a sentence in their own right. The full stop which ends most sentences is treated as the last constituent of the highest `S'. Other punctuation points and brackets are placed as high in the constituent structure tree as they can be, subject to other parsing conventions and rules.

Quotation marks, however, are ignored in parsing except when they mark an `Sq' (direct quotation) constituting an independent sentence.

The above paragraphs may have reinforced the idea that every wordtag is treated as significant in parsing: it is indeed true, in the main, that parsing operates on a word-by-word basis, treating the word-separating spaces of the written text as sacrosanct. However, apart from quotation marks, there is an important exception to this principle.

A sequence of words which behaves syntactically in an idiomatic way (e.g. `in order that', `as well as') is wordtagged by using the wordtag appropriate for the sequence as a whole for each of its components, together with two digits, the first of which indicates the total number of elements in the idiom, the second standing for the number of the position of that particular element within the sequence. E.g.:

`in_CS31 order_CS32 that_CS33'

`as_CC31 well_CC32 as_CC33'

`up_IN21 to _IN22'

The implication of this arrangement for parsing, of course, is that such sequence is treated as a unit for the purposes of parsing.