Retrieving Linguistic Information from a Corpus on the Example
OF NEGATION IN CHINESE
Luboš GAJDOŠ
Comenius University in Bratislava, Slovakia lubos.gajdos@uniba.sk
Abstract
The paper deals with corpus analysis of negation in Chinese, namely the negatives bù ^ and méi/méiyou S/SW. The adverbs BU and MEI are two of the most frequent negatives in Chinese. The aim of this study is to present statistical data together with linguistic analysis. The results provide empirical evidence of discrepancy between "authentic" language data versus linguistic prescription with practical implications for second-language acquisition. The findings inter alia suggest a new approach to verb categorisation.
Keywords: Chinese language; corpus linguistics; quantitative description; negation; potential complements
Povzetek
Članek obravnava korpusno analizo negacije v kitajščini, pri čemer se avtor osredotoča na prislova bu ^ in mei/meiyou	ki sta najpogostejši nikalnici v sodobnem kitajskem
jeziku. Namen prispevka je predstaviti statistične podatke v povezavi z jezikoslovno analizo. Rezultati študije prinašajo empirične dokaze o neskladju med jezikovno rabo in jezikovnimi normami, ta spoznanja pa je moč uporabiti tudi pri poučevanju kitajščine kot tujega jezika in za premislek o drugačnem pristopu na kategorizacijo glagolov.
Ključne besede: kitajščina; korpusno jezikoslovje; kvantitativni opis; negacija; zmožnostna dopolnila
Acta Linguistica Asiatica, 9(2), 2019. ISSN: 2232-3317, http://revije.ff.uni-lj.si/ala/ DOI: 10.4312/ala.9.2.103-115
@0g)
104 Lubos GAJDOS
1 Introduction
Generally speaking, there are a number of negatives in modern Chinese.1 In this article only two negative adverbs, namely bu ^ and mei/meiyou ^/^W2 (hereafter referred to as BU and MEI), are discussed. The Hanku corpus is used3 as the primary source of language material and statistical data. As the intention is to mainly use the corpus-driven4 approach to studying of negation, thus the previous linguistics research on this topic is left aside.
Let us start with some basic queries:5
[tag="AD" & word="^" ] [tag="W|AD" & word="i£|i£W"]6
The results are 7371142 (9897.85 per million),7 686352 (921.62 per million)8 respectively. These numbers only tell that the token BU is approximately 10 times more frequent than MEI. The difference is even more pronounced when searching in a certain variety of Chinese, e.g. in the corpus of legal Chinese, the occurrence of BU is 45254 (6 281.56 per million), the occurrence of MEI 720 (99.94 per million). Let us take a closer look at the tokens that collocate with these negatives. The following queries should return collocates at the position 1 on the right side:9
[tag="AD" & word="^"][] [tag="AD"& word="&|$;W"][]
1	For details see e.g. Liu (2004, pp. 253-258).
2	For the sake of simplicity, both negatives mei and meiyou are treated as two forms of one negative, namely MEI. On the other hand, their collocative partners may differ because of e.g. prosodic factors.
3	See more in Gajdos, Garabik and Benicka (2016, pp. 53-65).
4	See more in Baker, Hardie and McEnery (2006, p. 49).
5	In this article, the Corpus Query Language (hereafter CQL) is used to search for collocations. With CQL, complex criteria can be set to find one or many tokens. Criteria for each token must be between a pair of square brackets [ ], e.g. [attribute="value"]. See more at https://www.sketchengine.eu/documentation/cql-basics/
6	As there are more tags (e.g. VV = verbs, VE = YOU as the main verb, AD = adverbs) dedicated to tokens mei and meiyou it is rather difficult to accurately determine the value of the negative MEI. Thus, I only use tags VV, AD in this article. For more details on the tagset see Fei (2000, pp. 4-35).
7	Unless stated otherwise, frequencies are given in absolute occurrence in the Hanku corpus.
8	The occurrence of mei ® is 342190 (459.49 per million) and 344162 (462.14 per million) for meiyou
m
9	The regular expressions may match the following patterns:
adverb bu + any token or adverb mei/meyou + any token. As it is rather difficult to identify the collocates at the position further to the right using only POS tags, this topic will be left for future research. See e.g. Gajdos (2018).
Retrieving Linguistic Information from a Corpus on the Example
105
The results are summarised in the tables below.10
Table 1: The most frequent POS at the position 1 (Corpus: web-zh)
#	Query: word, [tag="AD" & word="W"][]
#	Query: word, [tag="AD|VV" & word="S | SW"][]
No. of results: 7371142	No. of results: 686352
tag		frequency	tag	frequency
AD	VV	5092363	VV VV	326984
AD	VA	827140	AD VV	193781
AD	VC	695511	VV AD	35762
AD	AD	444055	VV P	30138
AD	P	155025	AD AD	16246
AD	PU	25581	VV AS	15136
AD	AS	19730	VV PU	9902
AD	JJ	19525	VV CD	5771
AD	PN	13121	VV PN	5350
AD	NN	10044	VV VA	4975
AD	BA	9276	VV BA	4870
AD	VE	8798	VV NN	4449
AD	LB	8553	VV DT	4223
AD	CD	7535	AD VA	3511
AD	SB	6815	AD P	3488
AD	NR	5930	VV SB	2795
AD	DT	4226	VV DEC	2660
AD	DEC	3170	VV LB	1914
AD	DEV	2380	AD SB	1541
AD	SP	1987	VV DER	1460
AD	LC	1744	AD BA	1334
AD	M	1709	VV JJ	1209
AD	MSP	1545	VV NR	1106
AD	NT	1097	VV NT	959
AD	OD	1032	AD CD	861
AD	CC	1023	AD PN	691
AD	CS	1021	AD VC	663
AD	DEG	941	VV VC	526
AD	DER	118	VV OD	499
AD	ETC	88	VV LC	380
AD	FW	45	AD VE	354
AD	IJ	14	AD LB	338
10 The results are calculated using the NoSketch Engine UI - Node tags.
106 Lubos GAJDOS
The table indicates different collocability for the negative BU and MEI, e.g. the negative BU exhibits a strong preference for copulas (here VC).11 For practical reasons, only the POS tags, which are more frequent than 1% of each group (here in bold), are included in the analysis. The PU tag is also to be excluded from further analysis as it stands for punctuation.
Table 2 shows 10 of the most frequent collocates for each negative. The results are calculated using the NoSketch Engine UI - Node forms.12
Table 2: The most frequent tokens at the position 1 (# Corpus: web-zh)
#	Query: word, (meet [tag="VV|VA| VC|AD|P"]2:[tag="AD" & word="*"]-1 -1)
#	Query: word, (meet [tag="VV| AD|P|AS"]2:[tag="AD | VV" & word="S | >St"]-1 -1)
	*			mm	
	No. of results:	7214094		No. of results: 621555	
word		frequency	word		frequency
£		693223	M		56260
m		377529	*		14539
		311133	m		13999
M		288756	7		12223
		227692	M		9843
		125328	£		9249
		125212	U		9215
		124252	rn		9141
rn		91523	*3\		7189
m		85127			6744
At first sight, it is surprising that the collocation MEI+ neng m is the third most frequent, despite the fact that most grammars and textbooks deny this possibility.13 Similar findings may provide the impetus for further research which would take greater
11	The co-occurrence of MEI+VC is caused by the misspelling of the character shi in most cases, e.g. mei shi instead of mei shi
12	To find collocative partners of both negatives, the operator meet is used. That means that the corpus is search for the following patterns:
adverb BU + verbs (VV) or adverb BU + adjectives (VA) or adverb BU + copulas (VC) or adverb BU + adverbs (AD) or adverb BU + prepositions (P).
See more at https://www.sketchengine.eu/documentation/cql-meet-union/
13	There are some exceptions, e.g. Svarny and Uher (2014, p. 48) describe this phenomenon and Liu (2004, p. 257) also suggests this possibility, however, they do not further elaborate this point.
Retrieving Linguistic Information from a Corpus on the Example
107
account of actual language use. The "new" grammars or textbooks should be then based on such research.
After searching the first hundred examples manually, it turns out that the cooccurrence of some tokens with BU is higher than one would expect based on the frequency of affirmation, e.g. liâo 7 (70691), zhù ft (65400), qï M (40086) etc., furthermore, these verbs typically serve as so-called complements.14 That means that only tentative conclusion may be drawn from this evidence, nevertheless, it should play a role when comparing the overall frequency of both negatives. I discuss this topic further in the chapter Potential complements.
2 Potential complements
Let us return to the examples that have been mentioned in Chapter 1 and analyse them.
(1) m
Tä He
'He cannot do this.'
& ^/AD
bàn bù/AD manage BU-neg,
7/vv W
liäo/VV cï shi compl. this matter
(2)	U ^/AD	ft/VV	UT
jieke zhongyu	ren bu/AD	zhu/VV	shuo le
Jack finally	endure BU-neg.	compl.	speak LE
'Jack finally couldn't help saying it.'
(3)	m	^/AD	Ê/VV	7
zhè bèizi fangzi	mai	bù/AD	qï/VV	le
this life house	buy	BU-neg.	compl.	LE
'(One) cannot afford to buy a house for the entire life.'
(4)	^/AD U/VV
duiyu shuangfang bu neng bu/AD shuo/ VV for to both sides not able to BU-neg. speak 'regarding (things that) both sides cannot but speak'
Randomly selected samples suggest that many examples may be considered as so-called potential complements with the "morphological" structure VV + BU + VV while
14 See e.g. Yip (2009, pp. 234-241).
108 Lubos GAJDOS
the first morpheme (verb) is not equal to the third. The following query meets this condition:15
(meet (meet 1:[tag="W"][tag="AD" & word="*"]-1 -1) 2:[tag="W|VA"]-2 -2) & 1.word!=2.word
The examples below show that the regular expression does not always match the desired pattern and therefore must be modified.
(5)	I^AW^-, & * M/VV? Nage ren ne, jiao Li YT, zh! bu zhTdao That person is Li Yi know BU-neg. know 'Do you know that that person is called Li Yi?'
(6)	m m * mww
Nin neng bu	nenggou zai juti de gen women jiang yTxia?
You able to BU-neg. able to tell us more specifically 'Can you tell us this again more specifically?'
(7)	m * n/vv
YTding chengdu shang, bu neng bu shuo/VV to a certain extent, not able BU-neg. speak 'To a certain extent, one cannot but speak.'
(8)	n * ^/VV
YTge shuo bu yao one say BU-neg. want 'One says no.'
(9)	n * Mrw xfc Zhexie ren zhao bu dao/VV gongzuo these people find BU-neg. compl. work 'These people cannot find work.'
There is a point worth noting here as well - auxiliary verbs (e.g. modal verbs) must be removed from the search pattern. As there is no dedicated tag for modal or auxiliary
15 This regular expression matches the following pattern:
verb2/adjective2 + BU + verbl and verb1 * verb2, the verbl is KWIC (Key Word in Context).
Retrieving Linguistic Information from a Corpus on the Example
109
verbs (except VE, VC), each of the verbs must be enumerated in the query with the attribute "word".16 A double negative must be excluded too. The refined query is:17
(meet (meet (meet 1:[tag="VV"& word!= "g | f£" "& word="(?i).{1,2}"] [tag="AD" & word="^"]-1 -1) 2:[tag="VV|VA"& word!= "g| M"]-2 -2)[word!="^"]-3 -3) & 1.word!=2.word
The following table shows the result. The overall frequency is 828224 (1112.13 per million).
Table 3: The most frequent potential complements - the negative form (# Corpus: web-zh) # Query: word,(meet (meet (meet 1:[tag="VV"& word!= | | M& | ^ | MM | ^^ | ^ | MS |
^mmmmmnmmmti^mmmMinmmim" &
word="(?i).{1,2}"][tag="AD" & word="^"]-1 -1) 2:[tag="VV | VA"& word!= | | | ^ | MM | 2)[word!="^"]-3 -3) & 1.word!=2.word
word	Frequency
	111457
T liäo	65697
	59793
&	36774
# dé	33857
&	31377
±	29568
	22326
a	19412
m	12632
^ zhao	10366
	10114
f	9877
16	The query above contains only two of these verbs, the others are present here, e.g.	^ |
etc. The limit for
the length of the tokens is set to 1 or 2 by the expression: "word="(?i).{1,2}".
17	The regular expression means that the corpus is searched for the following pattern:
token (not BU) + verb2 (not g nor + adverb BU + mono- or disyllabic verbl (which is not g nor and verbl * verb2. Only the verbl is KWIC in the concordance and other tokens are used as contextual filters. See more at https://www.sketchengine.eu/documentation/cql-meet-union/
110 Lubos GAJDOS
The result of the affirmative form might be achieved by the same query with only minor modification:18
(meet (meet (meet 1:[tag="W"& word!= "g | M" "& word="(?i).{1,2}"]
[tag="DER"]-1 -1) 2:[tag="VV|VA"& word!= "g| M"]-2 -2)[word!="^"]-3 -3)
& 1.word!=2.word
The total frequency of 167822 (225.35 per million) clearly shows that the occurrence of the affirmative form is far less frequent. This fact only validates the previous assumption mentioned in the literature.19 The following list contains a sample of the most frequent verbs: ± 11598, M 10084, glj 9769, ft 7 614, ft 7607, ft^, 3736, % 2977 etc.
If we move back to the calculation of the overall frequency of BU, the value of the negative form of potential complements (1112.13 per million) should be subtracted from the total frequency, i.e. 8785.72 per million. Needless to say, these are only approximate numbers and further research is required.
3 Verb collocates
The first chapter discusses the collocability of the negative BU and MEI. In this chapter, I further explore this topic. When comparing the total frequency of BU vs. MEI, some considerations should be taken into account, i.e. some verbs/adjectives collocate with BU only, some registers use only a limited number of MEI etc.
After saving the results as a text file (from the NoSketchEngine UI), I proceed to test the 2 lists20 for the duplication21 and calculate the average value of co-occurrence. When comparing two lists for duplication in the spreadsheet program, there are many tokens in the MEI list which are marked as they have no counterpart in the BU list. This might cause surprise at first since one would expect only tokens from the BU list not having a counterpart. The explanation is rather simple: (1) most of these tokens have a disyllabic morphological structure (V+X), e.g.	and cannot be paired with
their monosyllabic counterpart in the BU list by the spreadsheet program (e.g. or (2) the frequency of the BU counterpart is below the lowest frequency of samples (see footnote 13).
18	There is a dedicated tag for the de-marker, i.e. DER.
19	See e.g. Liu (2004, p. 583).
20	Each list contains the 1000 most frequent verbs that collocate with BU and MEI.
21	This might be done in MS Excel, LibreOffice Calc or any spreadsheet program.
Retrieving Linguistic Information from a Corpus on the Example
111
Table 4: The 10 most frequent verbs collocating with BU and MEI (# Corpus: web-zh)
#	Query: word,(meet [tag="VV"]2:[tag="AD" & word="^"]-1 -1)
#	Query: word,(meet 1:[tag="VV" & word="(?i).{1,2}"][tag="AD | VV" & word="S | SW"]-1 -1)
	*		mm
word	Frequency	word	Frequency
m	377442	M	56260
A	311129	*	14539
M	288756	m	13999
	227691	m	9843
	125328	u	9215
^	124875	m	9141
m	122930	*m	7189
rn	91458	m	6086
m	85108	£	5602
	78854	*	5371
The results indicate that:
•	From the list of the 1000 most frequent tokens (verbs) with the negative BU, 619 tokens collocate with MEI too, yet from the 100 most frequent tokens, there are 69 of them; the rest are e.g. the following tokens: £n, ^T, M, IS,
M, etc. that co-occur with BU only;
•	From the list of the 100 most frequent tokens (verbs) with the negative MEI, a few preferably collocate with MEI, e.g. Mi'J, MS, etc.;22
•	The lower the frequency of a token in the BU list, the less frequent it collocates with both negatives;
•	Generally speaking, the co-occurrence of the negative MEI with the same verb is about 2.5-time less frequent as with the BU negative, however, statistical data reveals great disparities between tokens (see table 5). That is to say that verbs on the left side of the table collocate almost always with the negative BU, on the other hand, verbs on the right side almost exclusively collocate with the negative MEI.
22 This may be seen from the following comparison: the query [word="^|^^" & tag="W|AD"][word=";^il"] with the frequency of 4542 (6.10 per million) and the query [word=" & tag="AD"][word=";il"] 62 (0.08 per million).
112 Lubos GAJDOS
Table 5: Collocability of verbs (# Corpus: web-zh)			
Preference for BU		Preference for MEI	
word	ratio	word	ratio
M A ft « n Mi 7	1511,8 858,4 781,7 408,7 325.2 307,7 301.4 278.5 236.3 235.4	Sa mv MA mn sa mm	0,005 0,182 0,201 0,277 0,290 0,314 0,323 0,340 0,342 0,385
4 Adjective and adverbs collocates
This chapter focuses on the collocability of adjectives and adverbs and the same searching methods are used.
As for the adjectives, a brief look at the given statistical data (827140 or 1110.67 per million vs. 8486 or 11.39 per million; see table 6) demonstrates that adjectives (almost) always collocate with the negative BU. The exceptions here may be considered as phrases.
Table 6: Collocability of adjectives (# Corpus: web-zh) # Query: word,(meet [tag="VA"]2:[tag="AD" & word="*"]-1 -1)
# Query: word,(meet [tag="VA"]2:[tag="AD|W" & word="j%|	-1)
	*		
	No. of results: 827140		No. of results: 8486
word	Frequency	word	Frequency
»	71999		1978
	58946		884
£	46448		812
%%	34917	£	374
	28451		329
x	23685		257
M	18118	»	251
l	14844	mm	205
	14831		195
*	13191	»»	180
Retrieving Linguistic Information from a Corpus on the Example
113
The situation with regard to adverbs is a little different. While the results indicate a strong tendency to the negative BU, yet both negatives may be used.
Table 7: Collocability of adverbs (# Corpus: web-zh)
#	Query: word,(meet [tag="AD"]2:[tag="AD" & word="*"]-1 -1)
#	Query: word,(meet [tag="AD"]2:[tag="AD | VV" & word="% | SW']-1 -1)
	*		SIM
	No. of results: 444055		No. of results: 52008
word	frequency	word	frequency
S	76724		6480
X	39336	S	4637
	22521		4062
M	19298		2989
H	17114	£	2855
Èâ H	13088		2316
	12286		1889
£	10410	*	1541
£	8623	X	988
	8621		920
5 Conclusion
To begin with, statistical data given in this study should only be taken as exhibiting a general tendency and not as a fully accurate description of "real" language. It should also be pointed out that this paper only examines the occurrence of negatives at the first position to the left of collocates. In this respect, new methods should be devised for solving issues addressed here, e.g. the problem with the POS annotation and its error rate which may significantly affect statistical data or the problem with identifying the difference between the negative MEI and the verb you ^ (with the tag VE) etc. This leads us to the questions how to interpret the results in light of these points and what valuable results this study brings.
Firstly, when comparing results of both negatives, it seems that some verbs described as "auxiliary" or "modal" tend to collocate with the negative MEI more often than stated by language prescription. On the other hand, empirical data support the claim that adjectives only collocate with the negative BU. As for adverbs, there is still a strong preference for BU, but because I do not consider adverbs as a "true" collocate to negatives (rather as part of a bigger structure), this question should be explored in future research.
Let us now move on to the negative MEI. There are many verbs that preferably collocate with MEI rather than with BU. A closer look at the results reveals that their
114 Lubos GAJDOS
morphological structure is disyllabic and the left morpheme is often a so-called "resultative complement" (jieguo buyu ^n^^Ki). This finding may imply that the category of verbal aspect and tense23 deserves closer attention. That means if MEI is regarded as past time marker, these verbs are commonly used in past tense and the present tense (with BU) may describe the situation as a condition or future tense. A similar phenomenon is also observed in some Slavic languages, where the present and preterite of perfective verbs fulfil these functions too (e.g. compare the present perfective form "urobim" vs. the past perfective form "urobil" in Slovak). This suggests that these verbs in Chinese might be treated as perfective. In order to fully explore this topic, the marker le T, as a counterpart to the negative MEI, should be included in an comparative analysis. There is a very detailed, corpus-based study conducted on this subject by Petrovčič (2009), Operator Le in Chinese worth noting here.
To conclude, the article shows how to use a corpus when searching for evidence of some language phenomena. As for negation in Chinese, the paper only suggests a different approach to this subject and additional research is needed.
References
Baker P., Hardie A., & McEnery, T. (2006). A Glossary of Corpus Linguistics. Edinburgh: Edinburgh University Press.
Fei, X. (2000). The Part-Of-Speech Tagging Guidelines for the Penn Chinese Treebank (3.0). Available at https://www.cs.brandeis.edu/~clp/ctb/posguide.3rd.ch.pdf
Gajdoš, L., Garabik, R., & Benicka, J. (2016). The New Chinese Webcorpus Hanku - Origin, Parameters, Usage. Studia Orientalia Slovaca, 15(1), 53-65.
Gajdoš, L. (2018). Korpusova analyza adverbiale v pravnej činštine [Corpus Analysis of adverbial complements in legal Chinese]. In T. Guldanova (Ed.), Kontexty sudneho prekladu VII (pp. 27-39). Bratislava: Univerzita Komenskeho.
Liu, Y.	et al. (2004). Practical Chinese Grammar [^^RinÜ]. Beijing: Shangwu
yishuguan.
Petrovčič, M. (2009). Operator Le in Chinese. Saarbrücken: VDM Verlag.
Petrovčič, M. (2017). Traditional and Contemporary Approaches to Chinese Particles. In S. Bračič & M. Petrovčič (Eds.), Partikeln überall: Deutsch - Slowenisch - Chinesisch, (pp. 103122). Ljubljana: Znanstvena založba Filozofske fakultete.
Sketch Engine. Available at https://www.sketchengine.eu
Švarny, O., & Uher, D. (2014). Prozodicka gramatika činštiny [Prosodic Grammar of Chinese Language]. Olomouc: Univerzita Palackeho.
Yip, P., & Rimmington, D. (2009). Basic Chinese. Abingdon: Routledge.
23 See also Petrovčič (2017, pp. 108-109).
Retrieving Linguistic Information from a Corpus on the Example
115
Appendix: The Hanku tagset24
Tag	English	Example
AD	adverb	fa
AS	aspect particle	m
BA	preposition BA in ba-construction	^
CC	coordinating conjunction	M
CD	cardinal number	
CS	subordinating conjunction	M
DEC	markers - nominalizer	
DEG	genitive marker	m
DER	resultative DE	
DEV	manner DE ^	
DT	determiner	a
ETC	et cetera	
FW	foreign word	ISBN
IJ	interjection	
JJ	other noun-modifier	£
LB	preposition BEI in long bei-construction	$
LC	localizer	±
M	measure word	
MSP	other particle	m
NN	noun	
NR	proper noun	^in
NT	temporal noun	
OD	ordinal number	IH
ON	onomatopoeia	
P	preposition	
PN	pronoun	m
PU	punctuation	o
SB	preposition BEI in short bei-construction	$
SP	sentence-final particle	7
VA	predicative adjective	*
VC	copula	
VE	verb as the main verb	
VV	verb	
24 For details see Fei (2000, pp. 4-35).