Part of Speech and the Semantic Contribution of
Reduplication
Authors
Alex Duchnowski
Arushi Kalpande
Ford Rowe
For the full code and data used in this paper, please visit
our
GitHub repository.
1 Background and Context
The repetition of a morpheme, word, or phrase, also
known as reduplication, has a variety of uses in the
world’s languages. When looking at the behavior of
English reduplication, we see that its scope is fairly
limited in terms of its semantic effects, lacking a
productivity that is found in partial or total
reduplication found in other languages (Harada 2007).
This does not mean that analyzing English reduplication
is not worthwhile, and as will be discussed later in
this paper, there are still unanswered questions
relating to the mechanical properties of reduplication.
In fact, the impetus for the research seen here was
derived from a casual in-the-wild observation made by
this paper’s authors, in which they observed that while
both of the following tokens involve the apparently
identical morphological process of reduplication, the
semantic behavior of “run run” (as opposed to telling
someone just to “run”) and “blackmail blackmail” (as opposed to emotional blackmail) contained a
subtle difference. In the former, the reduplication
seemed to be contributing some form of emphasis or
intensity, while in the latter the reduplication seemed
to be pointing towards the blackmail being real or
normal, as opposed to a more specific or non-standard
type of blackmail. This paper, in addition to the goals
outlined in Section 1.3, seeks to identify whether or
not the differing part-of-speech of the base forms is
predictive of the differing semantic behavior of their
reduplicated forms.
1.1 Contrastive Reduplication
The specific phenomenon that will be examined in
this paper is Contrastive Focus Reduplication, also
simply called Contrastive Reduplication (CR), which
is a type of morphological process in English that
is used to contrast the meaning of the reduplicated
form with the meaning of its non-reduplicated base,
which may be explicit or implied within the context
of an utterance.
As a result of limiting our scope in this way, our
research will not be a comprehensive analysis of all
semantic behavior of all types of English
reduplication. This means that excluded from our
analysis will be reduplicated forms such as baby
talk (eg. goo-goo, ga-ga); deprecative
reduplication, which is used in the depreciation or
devaluation of the base form (eg. party-schmarty,
vegetables-schmegetables); intensive reduplication,
which is used to non-contrastively intensify the
meaning of the base form (eg. triplication in “It’s
all about money, money, money!”); and others as
outlined in the typological work of Ghomeshi et
al. (2004).
Additionally, there are some forms of reduplication
which may fall under the scope of CR, but whose
semantic content is not variable in such a way as to
be relevant to the research conducted here. Drawing
upon the work of Widlitzki (2016), our investigation
into CR will be limited to context-dependent
reduplication, in which the meaning of the
reduplicated form is somewhat dependent upon the
context in which it is uttered. By this measure, we
will be excluding CR which falls under the umbrella
of having a single default meaning or an
emerging default meaning. The former refers
to CR which has a broadly recognized meaning that is
not dependent upon context, such as “really-really”
or “very-very.” The latter refers to CR forms which
are seemingly in the process of developing a
conventionalized meaning, such as the increasingly
salient romantic connotation of the phrase
“like-like.”
1.2 Semantic Contribution
Having established the scope of CR that we will be
dealing with in this paper, we’ll now turn to
elaborating on what we mean when we refer to the
‘semantic contribution’ of reduplication. Put
simply, we are referring to the way in which the
meaning of the reduplicated word or phrase changes
for having undergone the morphological process of
reduplication. The following is a three-way typology
that we have developed to facilitate the distinction
of the different semantic contributions of CR, based
primarily on prior work by Ghomeshi et al. (2004)
and Widlitzki (2016). Most relevantly, from Ghomeshi
et al. (2004:343) we take the categories
Prototypical, Extreme, and
Salient to be the three major types of
meanings conveyed by CR.
Prototypical
CR which is prototypical is used to contrast a
real, typical, or standard example with an
example that is not understood to be so. This is
what is most commonly understood to be the
effect of English contrastive reduplication,
and, as we will see later in the paper, is the
most commonly occurring type. Examples of
Prototypical CR (PCR) would be the following:
I’ll make the tuna salad, and you make
the SALAD-salad.
Here’s the skinny. We dressmakers have a
very strict code, so I need to know, do
you deserve to wear virginal white?
Because if you don’t, you’ll have to
wear an off-white, what we call a hussy
white. So which will it be, WHITE-white?
In (1), the salad being described as the
“salad-salad” is most likely a leafy green salad
that may be referred to as the ‘prototypical’
version of a salad. In being brought to contrast
with a tuna salad, a less ‘prototypical’ type of
salad than a leafy green salad, (1) demonstrates
an example of PCR. In (2), the color
“white-white” refers to the former “virginal”
white. The speaker is using reduplication to
draw a distinction between pure white and
off-white, which is not white-white, thus (2) is
also representative of the use of PCR.
Extreme
CR which is extreme is used to indicate an
intensified meaning, most straightforwardly
comparable to saying that something is
“very/extremely/super X.” Examples of Extreme CR
(ECR) would be the following:
A [to B, who is about to give a
recital]: Are you nervous? B:
Yeah, but, you know, not
NERVOUS-nervous.
First of all, I’m not like a
CELEBRITY-celebrity; I don’t get
recognized when I’m walking down the
street.
In (3), Speaker B is contrasting their level of
nervousness on a gradable scale, wherein they
are asserting that they are just nervous (i.e.,
a ‘little’ nervous) as opposed to
“nervous-nervous” (i.e., ‘very’ nervous), thus
demonstrating a use of ECR. In (4), the speaker
is drawing a contrast between different levels
of celebrity, arguing that they are only a
celebrity to a certain degree, as opposed to
celebrities who are of a higher caliber. In this
way (4) is an example of ECR by way of the
speaker’s assertion that they are not an extreme
version of their occupation.
Salient
CR which is salient is used to indicate that
there is a contextually salient or accessible
sense or reference for the meaning of the
reduplicated form. This is different from the
above two types of CR, which are both
contextually-dependent in their own way, by
being inclusive of grammatical categories such
as proper nouns, and word types such as
indexicals, whose referents change at the level
of the utterance rather than the sentence.
Examples of Salient CR (SCR) would be the
following:
A: Did you check out the leak in the
bathroom? B: What leak? A:
The LEAK-leak. [drags her into the
bathroom]
We have a wonderful bed. I think you’ll
enjoy it. Not ENJOY-enjoy. Sleep.
Enjoyable sleep.
They just call him psycho-Marcus because
they’re trying to distinguish between
him and MARCUS-Marcus and normal-Marcus.
In (5), Speaker A is employing contrastive
reduplication in order to assert that the leak
they are referring to is not some generalized
leak out in the world, but is rather a specific
leak whose referent should be accessible to
Speaker B in the conversational context. Because
we are not participating in the conversation
between Speakers A and B, unlike the meanings of
the CR present in (1-4), the meaning of (5) is
left slightly unclear due to the highly
contextually salient nature of SCR.
(6) is representative of SCR where it applies to
the contrast between a metaphorical and literal
reading of the base form. The speaker is using
“enjoy-enjoy” to refer to a meaning that
contains innuendo, which is a nonstandard
reading of “enjoy” that is only accessible in
the utterance as a result of their mentioning
the bed in the prior sentence. In this way, the
meaning of “enjoy-enjoy” arises due to its
salience in the context of its utterance, and
thus we can classify (6) as an instance of SCR.
Finally we have chosen to classify tokens like
the one that appears in (7) as SCR for a reason
much similar to (5), in that proper nouns behave
in such a way as to ‘pick out’ specific
contextually salient individuals, and thus CR
may be used to differentiate between two or more
similarly-named people. This is the case in (7),
where the speaker is using “Marcus-Marcus” to
mean something along the lines of “the Marcus
that we know” or “the Marcus that isn’t the
other two Marcuses that we know,” and so we can
refer to this type of reduplication as SCR.
1.3 The Task at Hand
In this paper, we will ask whether the part of
speech (POS) of a reduplicated form influences its
semantic contribution in the context of its
utterance. We investigate this question by
annotating and analyzing two English-language
corpora. It is our hypothesis that we will see a
distinct trend in certain parts of speech being more
commonly Prototypical/Extreme/Salient than others,
such as scalar adjectives trending extreme or nouns
being prototypical.
This question is of value in that English
reduplication is rare and often taken to be a purely
prototypical process, but as we will see both in our
discussion of existing research and in our analysis
of our own data, this is not necessarily the case.
In related research (Ghomeshi et al., 2004;
Widlitzki, 2016) we find acknowledgements of CR
outside of the scope of purely prototypical
contrast, but the underlying cause for this observed
semantic variation in CR has not been elaborated
upon to the degree that we will be exploring it in
this paper. Insofar as the existing literature is
concerned, the motivating factor in the semantic
variation of CR forms is currently unclear.
Furthermore, since reduplication has multiple subtly
distinct effects, our hypothesis is based on the
intuition that looking at the type of content being
reduplicated might shed light on why the resulting
semantic effects differ.
2 Data
2.1 Salad-Salad Corpus
The majority of our confirmed instances of CR are
taken from what we will be referring to in this
paper as the Salad-Salad Corpus, which consists of
the examples and data used in Ghomeshi et
al. (2004). All listed entries in this corpus were
annotated as instances of contrastive reduplication
as a part of their work, thus providing us with a
robust base of data upon which to test our
hypothesis. The Salad-Salad Corpus was last updated
on May 30, 2014, and contains 334 entries along with
their source, context, and full quotation. The
corpus also contains an index of POS and source
medium, as well as other notations such as whether
or not the reduplicated form involved repeating the
inflectional morphology of a given token.
2.2 OpenSubtitles Corpus
In addition to the Salad-Salad Corpus, we drew upon
a sample of a 2018 OpenSubtitles Corpus containing
about 222,000 sentences (1.38 million tokens).
OpenSubtitles is an online database of subtitles in
various languages for movies and TV shows; users
upload their transcriptions, which are then shared
publicly and downloadable for free. At time of
writing, the site contains 7,124,730 subtitle
entries. We accessed this corpus through
OPUS, an online database of 1,210 corpora across 744
languages.
The OpenSubtitles Corpus that we selected is
constrained to English-language subtitles only, is
drawn from a variety of genres and periods, and
includes both live-action and animated content. Our
sample of this corpus, which will be elaborated upon
in the following section, consisted of 613 examples
which were then manually checked for CR.
3 Methodology
3.1 Preprocessing
When processing both corpora, we limited our focus
to the four parts of speech which most often undergo
reduplication in English: Nouns, Verbs, Adjectives,
and Adverbs (Widlitzki, 2016). The Salad-Salad
Corpus required minimal filtering beyond this, as
these entries were already confirmed to be CR by
Ghomeshi et al. (2004) and were already annotated
for their POS in the corpus’s index. Apart from
filtering for the four POS’s of interest, we also
removed instances of reduplicated “like”, as this
has an emerging default meaning (as noted in section
1.1).
Our pre-processing work on the OpenSubtitles corpus
was significantly more intensive, as we needed to
filter for repetition of words while retaining
relevant context for each example and filtering out
repetition which was unlikely to be CR. We started
by transforming the initial corpus so that every
line contained copies of the lines surrounding it
for context. The initial filtering of this content
was then achieved with the following Linux script
(for full context, see
this file):
```{bash}cat data/$name/contextualized_$name.txt |grep-vP'^[^\t]*\b([a-z]{3,})[ ,-]+\1[ ,-]+\1\b'|# remove rows with triples in the first columnsed-re's/^([^\t]*)([[\(].*[]\)])(.*)\2/\1\3/g'|# remove bracketed content in the first columnsed-re's/^([^\t]*)/\L\1/'|# convert text in the first column to lowercaseegrep"^[^$tab]*\b([a-z]{3,})[ ,-]\1\b"|# filter for reduplications in the first columnsed-re's/^[^\t]*\b([a-z]{3,})[ ,-]\1\b[^\t]*\t/\1\t/g'`# extract reduplicated words in the first column`\> data/$name/${name}_redups.txt```
As can be seen in the code above, we removed
bracketed content since this was typically used to
denote scene descriptions, sound effects, or other
non-dialogue information. We also converted the
column of interest to lowercase to facilitate
searching for reduplication without
case-sensitivity. Additionally, we removed instances
of repetition where a word or morpheme was used more
than twice in a row, as this was often singing or
laughter, rather than actually reduplication. After
these filters were applied, we searched for and
extracted examples where a single word at least
three letters in length was repeated (with a space,
comma, or hyphen separating the two copies). The
length requirement was to filter for more contentful
words and to preemptively remove words in other POS
categories such as prepositions and pronouns. Part
of speech was then labeled using the POS tokenizer
of the spaCy Python package, and these
POS annotations were then used to filter for the
four most commonly reduplicated parts of speech, in
keeping with the filtering described for our
Salad-Salad Corpus data. Applying this filter to
about 222,000 lines of subtitles resulted in 613
candidate examples of CR.
3.2 Manual Annotation
Salad-Salad
Our resulting list was 286 entries long, down
from the original 334, and was not entirely free
of non-CR tokens. For example, an entry in which
Jay Leno makes a joke about going to prison
within a prison — “prison-prison” — was notated
by the original authors as CR, but we agreed
that this did not seem to be an accurate
notation. In addition to verifying the presence
of reduplication, we manually annotated whether
or not the semantic contribution of the
reduplication was Prototypical, Extreme, or
Salient. We divided our annotation into two
sessions, starting with the first 100 examples,
then meeting to go over our answers and discuss
any discrepancies before proceeding to annotate
the rest. When we disagreed on an example, we
would discuss the reasoning behind our
annotations and then revise our judgements to
produce a single label for each example. This
was done both to improve our understanding of
the typology and to ensure we had a large enough
set of labeled examples to perform an
informative analysis. Subsequently, we annotated
the rest in the same manner, annotating
individually and then agreeing on a unified
ground truth as a group afterwards. Our total
number of verified reduplications in the
Salad-Salad corpus was 280.
OpenSubtitles
Our list of 613 samples, while filtered for
examples that fit a pattern of reduplication,
contained a significant amoun of non-CR data. An
example of this is repetition that was not
reduplication, such as the use of the phrase
“now I’m a black widow widow” in
A Bug’s Life (the first “widow” is part
of “black widow”, the type of spider, and the
second is the normal usage of “widow”). To
address this, we added the label N to our
annotations to indicate that an example was not
an instance of CR. By including the surrounding
context of each example, we were able to more
easily judge whether a repetition was meaningful
CR or not. We split the samples into the first
200 entries and the remaining 413 and annotated
the sets as was done with the Salad-Salad
corpus, with meetings after annotating each set
to resolve any disagreements.
Agreement
Using the krippendorff python
package (Castro, 2017), we calculated
Krippendorff’s Alpha values to see how well we
agreed on these labels for each set (we used a
few N/U annotations for Non-RC/Unclear for the
Salad-Salad dataset, but we treat such labels as
missing annotations for the purposes of
calculating Krippendorff’s Alpha). The results
are as follows:
Salad-Salad:
Examples 1-100: 0.2132
Examples 101-286: 0.5204
OpenSubtitles:
Examples 1-200: 0.6873
Examples 201-613: 0.5542
Overall, this agreement was relatively week, as
social scientists typically discard data with an
alpha value below 0.667. However, we were able
to come to a consensus on the majority of our
annotations, and the disagreements we had were
typically over edge cases or examples that were
difficult to classify. Thus, while we
acknowledge that our agreement was not perfect,
we believe that our annotations are still valid
and can be used for tentative analysis.
3.3 Word Length & Frequency
In addition to the above processing, we will include
in our analysis in the next section some discussion
of word length and word frequency as potentially
influential factors in our results. This is
partially derived from the acknowledgement that POS
may not exclusively have a part to play in the
semantic contribution of reduplication (for
instance, a word’s frequency may impact whether or
not it can be understood with a prototypical versus
salient interpretation), and are interested in
substantiating more strongly the specific link
between POS and semantic contribution, should there
be one. Word length could be calculated
straightforwardly in R.
```{r}salad <-read_csv("data/salad_consolidated_refined.csv") # Reading the Filesalad$Length <-nchar(salad$Word) # Creating a Length Columnfreq_list <-read_csv("data/unigram_freq.csv") # Reading the Frequency Counter Filefreq_list$Word <- freq_list$word # Making it so that both tables have a "Word" columnfreq_list <- freq_list[, -1] # Preventing the "word" column from sticking aroundsalad_counts <-inner_join(salad, freq_list, by="Word") # Adding the frequency list to the main tibble```
```{r}full_222k <-read_delim("data/en_v1_222k/en_v1_222k_full.csv", delim="\t") # Reading the File, adjusting for it being .csvfull_222k$Length <-nchar(full_222k$Word) # Doing the Character Countsfull_222k_counts <-inner_join(full_222k, freq_list, by="Word") # Doing Freqswrite_csv(full_222k_counts, "data/en_v1_222k/full_222k_counted.csv") # Writing the File```
Our calculations for word frequency are based on a
dataset
of the top 333,333 most frequent English words as
they appear in the Google Web Trillion Word Corpus.
To normalize the values, the log of the frequency
was taken.
4.0 Analysis & Discussion
4.1 Findings
As mentioned in the previous section, we annotated
both the Salad-Salad and OpenSubtitles datasets.
Although Salad-Salad’s contents were all determined
to be CR, as became apparent during the annotating
phase, the vast majority of the examples in our
OpenSubtitles subset were still not true CR even
after several rounds of filtering. The following pie
chart illustrates the percentages of different
labels, with ‘N’ representing ‘not CR.’
library(tidyr)library(dplyr)
Attaching package: 'dplyr'
The following objects are masked from 'package:stats':
filter, lag
The following objects are masked from 'package:base':
intersect, setdiff, setequal, union
Word Label POS Word_Length Freq_Count Log_Freq_Count
1 adam S NN 4 21694386 16.89256
2 aggressive P JJ 10 7345918 15.80966
3 aggressive P JJ 10 7345918 15.80966
4 alone E JJ 5 39326417 17.48741
5 amherst P NN 7 2846723 14.86168
6 anywhere E RB 8 23947583 16.99138
Warning in readLines("data/annotations/opensubs_agreed_2.txt"): incomplete
final line found on 'data/annotations/opensubs_agreed_2.txt'
all.labels <-c(labels.1, labels.2)label.freq <-table(all.labels)label.df <-data.frame(Label =names(label.freq), Frequency =as.numeric(label.freq))label.df$Percentage <- label.df$Frequency /sum(label.df$Frequency) *100ggplot(label.df, aes(x ="", y = Frequency, fill = Label)) +geom_bar(stat ="identity", width =1) +coord_polar("y", start =0) +labs(title ="Distribution of Semantic Contribution Labels in OpenSubs Dataset",fill ="Label") +theme_void() +scale_fill_discrete(labels =paste(label.df$Label, sprintf("(%0.1f%%)", label.df$Percentage))) +theme(legend.position ="bottom")
Thus, we decided to select only the positive CR
examples from the OpenSubtitles dataset and append
these annotated examples to the Salad-Salad dataset
and perform statistical analysis on this combined
dataset. First, we looked at the distributions of
both POS and Semantic Contribution Labels to get a
sense of the makeup of our dataset, as seen in the
following bar graphs. Nouns were by far the most
common POS to be contrastively reduplicated, and the
most common semantic contribution assigned across
the board was the prototypical label.
ggplot(combined.data, aes(x = POS, fill = POS)) +geom_bar(show.legend =FALSE) +geom_text(stat='count', aes(label=after_stat(count)), vjust=-0.5) +labs(title ="Distribution of Parts of Speech",x ="Part of Speech",y ="Count")
ggplot(combined.data, aes(x = Label, fill = Label)) +geom_bar(show.legend =FALSE) +geom_text(stat='count', aes(label=after_stat(count)), vjust=-0.5) +labs(title ="Distribution of Semantic Contributions",x ="Semantic Contribution",y ="Count")
Looking at the stacked graph below, we can see
certain trends beginning to surface. Although
prototypical labels still make up a large percentage
of the examples for each POS, there is a more even
split between prototypical and extreme readings for
adjectives, and prototypical and salient readings
for adverbs. The exact percentages of semantic
contribution by POS is given in the table below the
graph.
library(nnet)summary(combined.data)
Word Label POS Word_Length
Length:294 Length:294 Length:294 Min. : 2.000
Class :character Class :character Class :character 1st Qu.: 4.000
Mode :character Mode :character Mode :character Median : 5.000
Mean : 5.194
3rd Qu.: 6.000
Max. :13.000
Freq_Count Log_Freq_Count
Min. :2.957e+05 Min. :12.60
1st Qu.:2.173e+07 1st Qu.:16.89
Median :4.965e+07 Median :17.72
Mean :1.430e+08 Mean :17.70
3rd Qu.:1.447e+08 3rd Qu.:18.79
Max. :2.399e+09 Max. :21.60
ggplot(combined.data, aes(x = POS, fill = Label)) +geom_bar() +labs(title ="Distribution of Semantic Contributions by POS",x ="POS",y ="Count")
`summarise()` has grouped output by 'POS'. You can override using the `.groups`
argument.
pos.totals <- combined.data %>%group_by(POS) %>%summarize(total.count =n())label.percentages <- label.counts %>%left_join(pos.totals, by ="POS") %>%mutate(percentage = (count / total.count) *100)label.percentages.df <-as.data.frame(label.percentages)kable(label.percentages.df, caption ="Percentage of Semantic Contributions by POS")
Percentage of Semantic Contributions by
POS
POS
Label
count
total.count
percentage
JJ
E
32
72
44.444444
JJ
P
35
72
48.611111
JJ
S
5
72
6.944444
NN
E
13
154
8.441558
NN
P
107
154
69.480519
NN
S
34
154
22.077922
RB
E
7
32
21.875000
RB
P
12
32
37.500000
RB
S
13
32
40.625000
VB
E
6
36
16.666667
VB
P
21
36
58.333333
VB
S
9
36
25.000000
The regression model we ran on the dataset
formalized these trends, as discussed in the next
section.
4.2 Analysis
We ran a multinomial logistic regression model on
our dataset, as our output variable (semantic
contribution) was categorical. We ran a total of
three models, each looking at a different set of
predictor variables. The first model takes POS, word
length, and log word frequency into account. The
summary results of this model are shown below.
# weights: 21 (12 variable)
initial value 322.992013
iter 10 value 253.237645
iter 20 value 252.191106
iter 20 value 252.191105
iter 20 value 252.191105
final value 252.191105
converged
As the distributions showed that nouns and
prototypical were the most common in their
respective variables, we used these categories as
our baselines when calculating the log odds in the
model.
The results can be interpreted as follows:
The log odds of having a salient reading rather
than a prototypical reading:
decreases by 0.80 as you switch from
noun to adjective.
increases by 1.42 as you switch from
noun to adverb.
increases by 0.29 as you switch from
noun to verb. (not significant result)
increases by 0.084 for every 1 letter
increase in word length. (not
significant result)
decreases by 0.097 for every 1 unit
increase in log frequency. (not
significant result)
The log odds of having an extreme reading rather
than a prototypical reading:
increases by 2.01 as you switch from
noun to adjective.
increases by 1.86 as you switch from
noun to adverb.
increases by 0.85 as you switch from
noun to verb. (not significant result)
increases by 0.147 for every 1 letter
increase in word length. (not
significant result)
decreases by 0.149 for every 1 unit
increase in log frequency. (not
significant result)
The significance of the coefficient results was
determined by running a 2-tailed z test, and the
resulting z and p scores are given below.
z <-summary(model)$coefficients/summary(model)$standard.errorsz
(Intercept) POSJJ POSRB POSVB Word_Length Log_Freq_Count
S 0.06164851 -1.547434 2.849142 0.6546104 0.944361 -0.9193509
E -0.12347188 5.185728 3.012900 1.5273242 1.563639 -1.2351447
p <- (1-pnorm(abs(z), 0, 1)) *2p
(Intercept) POSJJ POSRB POSVB Word_Length Log_Freq_Count
S 0.9508427 1.217586e-01 0.004383735 0.5127186 0.3449852 0.3579121
E 0.9017334 2.151730e-07 0.002587638 0.1266804 0.1179023 0.2167767
Using a 0.05 threshold, we can see that the log odd
differences for noun to verb, word length, and log
frequency are not significant. For the case of noun
to verb, this result makes sense as the original
distributions also show that these two categories
show a similar split between the three semantic
categories. For the continuous variables, although
the trends might not be significant, and a larger
dataset might disprove trends observed here, we
still ran models on just these variables to get a
sense of the overall trends. The results of the word
length model are plotted below. Based on the
coefficients calculated by the model, the graph
plotted the predicted likelihood of being assigned a
particular semantic contribution label given word
length.
length.model <-multinom(Label ~ Word_Length, data = combined.data)
# weights: 9 (4 variable)
initial value 322.992013
iter 10 value 277.820475
final value 277.820469
converged
summary(length.model)
Call:
multinom(formula = Label ~ Word_Length, data = combined.data)
Coefficients:
(Intercept) Word_Length
S -1.587050 0.1035639
E -2.158591 0.1986573
Std. Errors:
(Intercept) Word_Length
S 0.4656235 0.08452959
E 0.4734252 0.08237858
Residual Deviance: 555.6409
AIC: 563.6409
new.data <-data.frame(Word_Length =seq(min(combined.data$Word_Length), max(combined.data$Word_Length), length.out =100))length.predictions <-predict(length.model, newdata = new.data, type ="probs")length.predictions.df <-data.frame(new.data, length.predictions)length.predictions.df <-gather(length.predictions.df, Label, Probability, -Word_Length)ggplot(length.predictions.df, aes(x = Word_Length, y = Probability, color = Label)) +geom_line() +labs(title ="Likelihood of Semantic Contribution Label by Word Length",x ="Word Length",y ="Predicted Probability",color ="Semantic Contribution Label") +theme_minimal()
As word length increases, the probability of getting
a prototypical reading decreases, while the
probability of getting an extreme reading increases.
There is also a slight increase in the salient
category, but this is a much slighter trend than
seen in the other two. Initial explanations for why
these trends could have arisen are discussed in the
next section. Similarly, we ran a model with log
frequency as the predicting variable.
freq.model <-multinom(Label ~ Log_Freq_Count, data = combined.data)
# weights: 9 (4 variable)
initial value 322.992013
final value 280.015169
converged
summary(freq.model)
Call:
multinom(formula = Label ~ Log_Freq_Count, data = combined.data)
Coefficients:
(Intercept) Log_Freq_Count
S -1.0582332 0.0002438423
E 0.9773974 -0.1182008828
Std. Errors:
(Intercept) Log_Freq_Count
S 1.670483 0.09366132
E 1.649430 0.09365141
Residual Deviance: 560.0303
AIC: 568.0303
new_data <-data.frame(Log_Freq_Count =seq(min(combined.data$Log_Freq_Count), max(combined.data$Log_Freq_Count), length.out =100))freq.predictions <-predict(freq.model, newdata = new_data, type ="probs")freq.predictions.df <-data.frame(new_data, freq.predictions)freq.predictions.df <-gather(freq.predictions.df, Label, Probability, -Log_Freq_Count)ggplot(freq.predictions.df, aes(x = Log_Freq_Count, y = Probability, color = Label)) +geom_line() +labs(title ="Likelihood of Semantic Contribution Label by Frequency",x ="Log Normalized Frequency",y ="Predicted Probability",color ="Semantic Contribution Label") +theme_minimal()
This model also revealed some general trends.
Namely, prototypical readings (and to a lesser
extent, salient readings) increased as frequency
increased, while extreme readings decreased as
frequency increased. In case the trends seen in
these graphs were a result of POS being a
confounding variable influencing word length or
frequency (e.g., if adjectives in our dataset were
longer or less frequent overall than other POS,
trends seen in these models’ results could just be a
result of POS and not the variable at hand), we
calculated simple mean and median stats for word
length and frequency by POS.
These stats show that across the different parts of
speech, the word length and frequency were roughly
the same (the only notable divergence is adverbs
being slightly more frequent on average), and so we
can conclude that the trends observed in the word
length and frequency models are a result of those
variables directly and not an underlying POS effect.
4.3 Discussion
Thus, based on the models and the various trends and
results, we can conclude that POS does have a
statistically significant impact on the semantic
contribution of contrastive reduplication.
Meanwhile, word length and frequency also seem to
correlate to certain trends in semantic contribution
of CR, but based on our data, these trends are not
statistically significant.
There were many trends relating to parts of speech.
First, nouns are most likely out of the parts of
speech to get prototypical readings, as we expected.
Verbs are similar to nouns in their overall
distribution of semantic contributions, with the
highest portion of them having prototypical
readings. Adjectives, on the other hand, are more
likely to get extreme readings than other parts of
speech (a trend which we had also anticipated).
Finally, adverbs are most likely out of the parts of
speech to get salient readings, and get them more
often than other readings. These trends have several
potential explanations. For instance, as supported
by our annotating experiences as well, it was much
easier to conceptualize a prototypical definition
for nouns and verbs than it was for modifiers. For
example, it is difficult to describe a prototypical
definition for the adjective “favorite,” and indeed,
both examples of this word got an extreme reading.
Building off of this, as adjectives can have a
scalar nature, they lend themselves well to extreme
CR readings. Finally, in our dataset, indexicals
such as “here” were classed as adverbs, and as
indexicals require context to resolve their
meanings, we believe this is a reason why adverbs
got the salient reading most often.
The trends in the word length and frequency results
were described in the previous section, and they
also have preliminary explanations. The decrease in
prototypical readings as word length increases makes
sense, as longer words are likely less frequent and
thus have less readily accessible prototypical
versions. This is confirmed by the frequency model’s
results, as prototypical readings increase as
frequency increases. Extreme readings, on the other
hand, increase with word length, and this
corresponds to its decrease as frequency increases.
Why exactly longer or less frequent words tend to
get extreme readings is not immediately clear. As
mentioned above, since there is no influence of POS
on the average length or frequency of any given
word, we cannot assume this trend in the extreme
readings is correlated with adjectives and their
length/frequency. Since adjectives have this
tendency for getting extreme readings, we had
predicted that this could be a factor in the trend
for extreme readings, but since this might not be
the case, an explanation for this trend is left for
future research. Finally, though there is a slight
slope in the lines for salient readings, it is
overall much more stable than the other factors.
This could be because given a rich enough discourse
context, any word can take on a salient CR meaning,
and thus the likelihood of getting this reading is
not influenced by word length or frequency. However,
it is worth noting that as mentioned above, the word
length and frequency results were not statistically
significant according to our dataset. Thus, further
research on a larger dataset might confirm or deny
the significance of these trends.
4.4 Limitations
There are, of course, aspects of the research we
have described here which could be improved through
further investigation or a reworking of our
methodological process. A particular factor that
limits the extent of our findings is that some of
the entries in the Salad-Salad Corpus are drawn from
written sources, and as acknowledged in Widlitzki
(2016), the informal register of reduplication as a
whole can be a biasing influence on what is written
versus spoken when it comes to this linguistic
phenomenon. Our data is not exclusively spoken
language, and thus may be somewhat skewed relative
to a data that is entirely based on transcribed
utterances.
A major caveat to our findings is that our typology
was fleshed out both using examples from our
dataset, and overall was designed by the three of us
in anticipation of annotating the rest of our data,
resulting in us having some theory as to how our
annotations ‘should’ align with our data when we did
sit down and begin sorting. For example, it is
possible that our classification of many adjectives
as Extreme may have been to some degree influenced
by our expectation that this would be the case upon
aggregating our annotated data. There is certainly
some bias in our research as a result of us acting
as both annotators and analysts.
We would also like to acknowledge that the typology
that we constructed at the outset, though
deliberated and agreed upon between the three of us,
could be fundamentally flawed on a theoretical
level; there could be categories missing whose
presence would create a better model for the trends
that we observe, for instance. Moreover, our low
agreement levels suggest that this typology may not
be very well-defined, and so we might wish to refine
these categories before performing any follow-up
research. There are also other typologies of CR not
based on a P/E/S paradigm of semantic contribution,
such as the scale of conventionalization of meaning
discussed in Section 1 that was proposed in
Widlitzki (2016), that may be more illuminating with
regards to the reason for variation within CR.
5.0 Conclusion
Based on our dataset, a few statistically significant
trends between POS and semantic contribution of
reduplication have been identified. Additionally, some
trends between word length/frequency and semantic
contribution have also been observed, but more research
into the strength of these trends is needed.
Supplemental data from a larger corpus can provide more
evidence to confirm or challenge these trends, and
potentially bring to light new insights regarding the
nature of contrastive reduplication.
Ghomeshi, Jila & Jackendoff, Ray & Rosen, Nicole
& Russell, Kevin. (2004).
Contrastive Focus Reduplication in English (The
Salad-Salad Paper). Natural Language and Linguistic Theory. 22. 307-357.
10.1023/B:NALA.0000015789.98638.f9.
Bianca, Widlitzki. (2016).
Talk talk, not just small talk. Exploring English
contrastive focus reduplication with the help of
corpora. ICAME Journal. 40. 10.1515/icame-2016-0008.
Inkelas, Sharon & Downing, Laura. (2015).
What is Reduplication? Typology and Analysis Part
1/2: The Typology of Reduplication. Language and
Linguistics Compass. 9. 502-515. 10.1111/lnc3.12166.
Harada, Ryuji. (2007).
A Note on Language Typology by the Status of
Reduplication. CORE. oai:otsuma.repo.nii.ac.jp:00001776
