Discourse Processes, 45:211–236, 2008

Copyright © Taylor & Francis Group, LLC

ISSN: 0163–853X print/1532–6950 online
DOI: 10.1080/01638530802073712

Gender Differences in Language Use:

An Analysis of 14,000 Text Samples
Matthew L. Newman
Department of Social and Behavioral Sciences
Arizona State University

Carla J. Groom*
Department of Psychology
The University of Texas at Austin

Lori D. Handelman
Oxford University Press
New York

James W. Pennebaker
Department of Psychology
The University of Texas at Austin

Differences in the ways that men and women use language have long been of interest
in the study of discourse. Despite extensive theorizing, actual empirical investiga-
tions have yet to converge on a coherent picture of gender differences in language. A
significant reason is the lack of agreement over the best way to analyze language. In
this research, gender differences in language use were examined using standardized
categories to analyze a database of over 14,000 text files from 70 separate studies.
Women used more words related to psychological and social processes. Men referred
more to object properties and impersonal topics. Although these effects were largely

*Carla J. Groom is now employed by the UK Department for Work and Pensions.
Correspondence concerning this article should be addressed to Matthew L. Newman, Department
of Social and Behavioral Sciences, Arizona State University, P.O. Box 37100, Phoenix, AZ 85069–
7100. E-mail: matt.newman@asu.edu
212 NEWMAN, GROOM, HANDELMAN, PENNEBAKER

consistent across different contexts, the pattern of variation suggests that gender
differences are larger on tasks that place fewer constraints on language use.

GENDER DIFFERENCES IN LANGUAGE USE:

AN ANALYSIS OF 14,000 TEXT SAMPLES

The last several decades have seen an explosion of research on the nature and exis-
tence of differences between men and women. One particularly popular question
has been the extent to which men and women use language differently. This popu-
larity stems, in part, from the fact that language is an inherently social phenome-
non and can provide insight into how men and women approach their social
worlds. Within the social sciences, an increasing consensus of findings suggests
that men, relative to women, tend to use language more for the instrumental pur-
pose of conveying information; women are more likely to use verbal interaction for
social purposes with verbal communication serving as an end in itself (e.g.,
Brownlow, Rosamon, & Parker, 2003; Colley et al., 2004; Herring, 1993).
At the same time, a number of theorists have argued against the existence of any
meaningful differences in men’s and women’s language (e.g., Bradley, 1981;
Weatherall, 2002). One contributor to this doubt may be the lack of a commonly
accepted metric of analysis among empirical studies of language. Multiple studies,
for example, have analyzed a small number of text samples and then made broad
generalizations about the differences between women and men. In this project, we
explored gender differences in language use in a very large data set of written and
spoken text samples using a computerized text analysis tool. Through this explora-
tion, we hope to provide some empirical resolution to the questions of whether,
how, and why men and women use language differently.

PREVIOUS RESEARCH ON GENDER DIFFERENCES

IN LANGUAGE USE

The empirical literature has been thoroughly reviewed elsewhere (e.g., Mulac,
Bradac, & Gibbons, 2001). What follows is a brief overview of previous research on
men’s and women’s language use. In addition to the overall message goals men-
tioned earlier, men and women may also have different semantic goals in mind when
they construct sentences. Some researchers (e.g., Mulac, Weimann, Widenmann, &
Gibson, 1988) found that questions are more common in women’s contributions to
dyadic interactions (e.g., “Does anyone want to get some food?”), whereas direc-
tives that tell the audience to do something (e.g., “Let’s go get some food”) are more
likely to be found in men’s conversational contributions. In a study of 96 schoolchil-
dren taken from the 4th, 8th, and 12th grades, Mulac, Studley, and Blau (1990) found
 GENDER DIFFERENCES IN LANGUAGE USE 213

that boys in all three age groups were more likely than girls to offer opinions (e.g.,
“This idea is Puritanical.”). When mean sentence length is calculated, women come
out as the wordier gender both in writing (e.g., Mulac & Lundell, 1994; Warshay,
1972) and speaking (Mulac & Lundell, 1986; Mulac et al., 1988; Poole, 1979).
However, men use more words overall and take more “turns” in conversation (e.g.,
Dovidio, Brown, Heltman, Ellyson, & Keating, 1988).
Some recent studies have failed to replicate these findings. Thomson and
Murachver’s (2001) study of e-mail communication found that men and women
were equally likely to ask questions; offer compliments, apologies, and opinions;
and to hurl insults at their “net pal.” Other studies have reported significant differ-
ences in the opposite direction. In a comparison of 36 female and 50 male manag-
ers giving professional criticism in a role play, it was the men who used signifi-
cantly more negations and asked more questions, and the women who used more
directives (Mulac, Seibold, & Farris, 2000). However, the study did confirm that
men used more words overall, whereas women used longer sentences. One possi-
ble explanation for these contradictory reports is that the different contexts in
which the language samples were generated influenced the size and direction of
the gender differences.
Beginning with Robin Lakoff’s (1975) pioneering work, gender differences
have also been investigated at the level of specific phrases. Lakoff identified in
women’s language two specific types of phrases—hedges (e.g., “it seems like,”)
and tag questions (e.g., “ … aren’t you?”)—that can be inserted into a wide variety
of sentences. A number of studies have reported greater female use of tag ques-
tions (e.g., McMillan, Clifton, McGrath, & Gale, 1977; Mulac & Lundell, 1986),
although others have found the opposite (e.g., Dubois & Crouch, 1975). Other re-
searchers have found further evidence that women use phrases that may communi-
cate relative uncertainty. Uncertainty verb phrases, especially those combining
first-person singular pronouns with perceptual or cognitive verbs (e.g., “I wonder
if”), have been found more often in women’s writing (Mulac & Lundell, 1994) and
speech (Hartman, 1976; Poole, 1979). A related interpretation of women’s use of
hedge phrases is that women are more reluctant to force their views on another per-
son. Consistent with this idea, Lakoff claimed that women were more likely than
men in the same situation to use extra-polite forms (e.g., “Would you mind … ”), a
claim that was supported by subsequent empirical work (Holmes, 1995; McMillan
et al., 1977).
Gender differences have also been examined by studying the actual words peo-
ple use. Mirroring phrase-level findings of tentativeness in female language,
women have been found to use more intensive adverbs, more conjunctions such as
but, and more modal auxiliary verbs such as could that place question marks of
some kind over a statement (Biber, Conrad, & Reppen, 1998; McMillan et al.,
1977; Mehl & Pennebaker, 2003; Mulac et al., 2001). Men have been found to
swear more, use longer words, use more articles, and use more references to loca-
214 NEWMAN, GROOM, HANDELMAN, PENNEBAKER

tion (e.g., Gleser, Gottschalk, & John, 1959; Mehl & Pennebaker, 2003; Mulac &
Lundell, 1986).
One striking result reported by Mehl and Pennebaker (2003) was that women
were more likely to use first-person singular. This is consistent with repeated find-
ings that depressed people use more first-person singular (e.g., Bucci & Freedman,
1981; Rude, Gortner, & Pennebaker, 2004; Weintraub, 1981), given that depres-
sion is more common among women (Diagnostic and Statistical Manual of Mental
Disorders [4th ed., text revision], American Psychiatric Association, 2000). How-
ever, the word “I” intuitively connotes individualism or selfishness, which fits the
male stereotype better than the female stereotype. The result is also at odds with a
review by Mulac et al. (2001), which cited findings that men used first-person sin-
gular more often. However, their conclusion was based on only two studies: one
representing analyses of 32 essays (4th-grader sample; Mulac et al., 1990), one
representing 148 essays (Mulac & Lundell, 1994), and both using relatively imper-
sonal writing tasks (essays and descriptions of photographs). Certainly, if the en-
tire category of personal pronouns is considered, women frequently are the higher
users (e.g., Gleser et al., 1959; Mulac & Lundell, 1986). Based on the existing data
alone, therefore, it is not possible to either confirm or disconfirm the stereotype
that men use I more than women.
Emotion words appear to be another area of conflicting findings, despite the ex-
istence of a fairly clear stereotype. Several studies have reported that women refer
to emotion more often than do men (Mulac et al., 1990; Thomas & Murachver,
2001). Yet, Mulac et al.’s (2000) study of managers providing criticism in a role
play found precisely the reverse. Mehl and Pennebaker (2003) offered a potential
reconciliation: Women used more references to positive emotion, but men referred
more to anger—a finding that is perfectly consistent with gender stereotypes.

Limitations of Previous Research

A frustration of studying natural language is that people use words in a variety of
ways that change as a function of context. To draw broad conclusions about how
men and women differ in their language use across settings, nontraditional meth-
ods with large samples are often required. In previous studies examining age and
individual differences, mean effects sizes for gender have typically ranged be-
tween .08 and .20. For example, Pennebaker and Stone (2003) reported a signifi-
cant effect size (d) of .12 between gender and sentence length with a sample of
over 2,000 people. However, the mean words-per-sentence was 23.4 for men and
19.1 for women, with a standard deviation of 35.1, across written and spoken text.
A difference of 4 words compared to a standard deviation of 35 words may not be
apparent in a small sample with low statistical power. Thus, large samples may be
required to detect the smaller differences between men’s and women’s language.
 GENDER DIFFERENCES IN LANGUAGE USE 215

Unfortunately, many previous studies have had fewer than 50 participants per
cell. Larger samples are often difficult to collect when each sample must be hand
coded. The need to conserve coder time also reduces the number of features that
can be coded in a single study. This reality has focused attention toward features of
language that can be easily related to gender stereotypes (e.g., hedges), potentially
missing differences in less obvious language categories (e.g., pronouns). Thus, a
strategy that allowed for the efficient analysis of large samples of text could help to
create a more complete picture of gender differences in language use.
A related limitation is that coding schemes are not always consistent across
studies. Even where the name of the language category is shared by two or more
studies, the actual features coded for may be different. One researcher’s uncer-
tainty verb phrase is another’s hedge. This problem is exacerbated by multivariate
approaches that compare men and women on a set of language features, rather than
reporting mean differences on individual features. The simplest form of this ap-
proach was used by Crosby and Nyquist (1977) in which they created a composite
“female register” index. The more complex multivariate approaches (e.g., Mulac
& Lundell, 1986) use multivariate analyses of variance (MANOVAs) in which lan-
guage features are weighted differently to achieve maximum discrimination be-
tween the genders. Mulac and his associates (Mulac et al., 1986; Mulac et al.,
1990; Mulac et al., 2000) argue that language is produced and comprehended as a
gestalt, and should be analyzed accordingly. However, such an approach makes it
difficult to compare results of one study with results of a second study that uses dif-
ferent combinations of features. A standardized set of language categories, com-
posed of a standardized set of features used in coding for that category, would shed
new light on the ways in which men and women communicate differently.

Function Words
Because of these limitations, empirical studies of language itself have yet to pro-
vide a coherent picture of gender differences in language use. Perhaps the greatest
stumbling block has been in deciding how to analyze language as it relates to
women and men. Language is inherently complex, and can be analyzed at several
levels of analysis. As discussed earlier, explorations of gender differences in lan-
guage have ranged from the overall structure of men’s and women’s narratives
(e.g., Herring, 1993; Tannen, 1990), down to the level of specific phrases (e.g.,
Holmes, 1995; McMillan et al., 1977; Thomas & Murachver, 2001) and words
(e.g., Biber et al., 1998; Danner, Snowdon, & Friesen, 2001; Gleser et al., 1959;
Mehl & Pennebaker, 2003; Mulac et al., 2001). Which dimensions of language
should we examine to capture differences in how men and women approach the
world?
A growing body of research suggests that we can learn a great deal about peo-
ple’s underlying thoughts, emotions, and motives by counting and categorizing the
216 NEWMAN, GROOM, HANDELMAN, PENNEBAKER

words they use to communicate (Pennebaker & King, 1999; Pennebaker, Mehl, &
Niederhoffer, 2003; cf. Shapiro, 1989). This approach has proved particularly
fruitful with respect to “function words,” which include pronouns, articles, prepo-
sitions, conjunctions, and auxiliary verbs. These words are distinct from content
words (nouns, verbs, and adjectives), and are used to “glue” other words together.
In the English language, there are fewer than 200 commonly used function words,
yet they account for over one half of the words we use.
Differences in the use of function words reflect differences in the ways that in-
dividuals think about and relate to the world. For example, using “you and I” in-
stead of “we” reflects a different perspective on the relationship between the
speaker and the referent. Using more pronouns in general (rather than nouns) re-
fers to a shared reality, in that both parties have to understand who “he” is. Em-
pirically, the use of first-person singular has been associated with age; depression;
illness; and, more broadly, self-focus (Pennebaker et al., 2003). First-person plural
can variously be a marker of group identity and, on occasion, a sign of emotional
distancing (Brewer & Gardner, 1996; Pennebaker & Lay, 2002). Function words
can also reflect psychological state independent of content. For example, people
telling the truth use more first-person singular and more qualifying conjunctions
(e.g., but) than those instructed to lie—although they are discussing the same top-
ics (Newman, Pennebaker, Berry, & Richards, 2003). This approach to language
suggests that differences in how individuals communicate can sometimes be as
meaningful as what they communicate.
An examination of gender differences in function word use might shed new
light on the psychology of men and women. Apart from personal pronouns, how-
ever, function words have been relatively neglected in previous gender difference
work. In one notable exception, Koppel, Argamon, and Shimoni (2003) discrimi-
nated between male and female authors in a sample of fiction and nonfiction from
the British National Corpus. Their goal in that study was to predict author gender
without regard to the psychological meaning of the words. These authors used a set
of training documents to create a prediction equation, which was used to classify
writing by gender, at an accuracy rate near 80%. Empirically, those words that best
discriminated between men and women were function words. In a second notable
exception, Biber et al. (1998) used parts of speech to create an index of whether a
text sample was “involved” (e.g., more pronouns, present-tense verbs) or “infor-
mative” (e.g., more nouns, long words). Consistent with prior research, females’
language was more involved than males’ language.

PRESENT STUDY

In this study, we conducted a comprehensive survey of gender differences in lan-

guage. This survey was brought within reach by a combination of two method-
 GENDER DIFFERENCES IN LANGUAGE USE 217

ological developments. The first was our text analysis program, Linguistic In-
quiry and Word Count (LIWC; Pennebaker, Francis, & Booth, 2001), which
allowed us to perform an extensive linguistic analysis on each individual text in
our archive. LIWC analyzes text samples on a word-by-word basis and com-
pares each to a dictionary of over 2,000 words divided into 74 linguistic catego-
ries. Output is expressed as a percentage of the total words in the text sample.
Some of the categories are defined purely grammatically. For example, the “arti-
cles” category searches for instances of a, an, and the. Other categories, such as
positive emotion words, were formed initially by having independent judges de-
cide which words should go into each category. Thus, an element of qualitative
human judgment is incorporated into an automated and perfectly consistent cod-
ing system. LIWC usually recognizes about 80% of the words in a given text
sample—proper nouns and very low-frequency words comprise the other 20%.
Once the text samples are assembled, thousands of samples can be analyzed on
dozens of dimensions in a matter of seconds.
A word count strategy such as LIWC is an admittedly crude way by which to
study language use. It cannot detect the context or underlying meaning of words. It
fails to appreciate sarcasm or irony. Because of these problems, words are often in-
correctly categorized. For example, the word “mad” is currently categorized as an
anger word. Phrases such as “I’m mad about you” (suggesting positive emotion) or
“mad as a hatter” (indicative of mental health problems) will be miscoded by the
computer. It is best to think of word count strategies as probabilistic ways of study-
ing language use. Statistically, we have found that mad is correctly coded about
90% of the time. Fortunately, in any given text, someone who is angry will use sev-
eral other anger-relevant words. In those cases where the person uses mad and is
actually happy, other positive emotions words will surface. In short, word count
approaches are prone to errors; but, with large data sets, the likely error rate is ex-
tremely low.
The second methodological development has been the creation of a text archive
itself. Over the last decade, we have collected a large corpus of over 500,000 text
files in the development of LIWC. Labs from all over the world have provided us
with language samples based on written and transcribed spoken language. In addi-
tion, we have accrued samples of books, poems, song lyrics, and other art
forms—many of which had never been subjected to linguistic coding. As a result,
we have the opportunity to observe gender differences on a much larger scale than
has been attempted in the past (e.g., Biber et al., 1998). This corpus is described in
more detail in the Method section.
This research strategy shares certain features with a traditional meta-analysis,
but is also distinct in several important respects. As in the case of a meta-analysis,
we drew on data collected in a variety of labs on a variety of populations, enabling
us to take advantage of the statistical power, increased external validity, and oppor-
218 NEWMAN, GROOM, HANDELMAN, PENNEBAKER

tunities to examine the effects of study-level moderators that are yielded by such
pooling of data. However, we did not pool analyses conducted by other research-
ers; instead, we used LIWC to code and categorize all of the original raw data, and
performed our own primary analyses on the resulting measures. A traditional
meta-analytic strategy would have been to seek out all the language studies that
had ever coded for a particular language feature (e.g., use of first-person singular);
use reported means for men and women to obtain study-level effect sizes; and cal-
culate an overall, weighted, meta-analytic effect size for this language feature.
Such a traditional strategy would yield no more and no less than a quantitative syn-
thesis of existing research, constrained by the limited number of studies that had
coded for each feature using varying definitions.
Our first goal in this study was to ask a rather simple question: Do men and
women use language differently? To answer this question, a large corpus of text
samples was subjected to LIWC analysis, and these linguistic data were analyzed
for main effects of gender. In analyzing this large body of primary data, our goal
was to resolve some of the discrepancies in previous studies. We expected results
consistent with the overall picture from previous research—that is, men’s language
should focus relatively more on conveying information, and women’s language
should focus relatively more on social connections. However, because of the issues
with previous research mentioned earlier, and because of the large and diverse cor-
pus of text at our disposal, we refrained from making predictions about specific
language categories. Instead, we took an exploratory, bottom-up approach to
men’s and women’s language use (see Oberlander & Gill, 2006, for a discussion
about the merits of a bottom-up approach). We expected the largest differences to
be found on function words because these words appear to be particularly good
markers of how individuals relate to the world. However, we also examined a range
of social and psychological process words, including references to friends, family,
and emotions, to better understand how men and women differ in their language
use (see Table 1 for a list of language categories).
Our second goal in this study was to examine whether the context in which text
samples were produced affected gender differences in language use. Few previous
attempts have been made to systematically study how context influences the size
and direction of gender differences in language use. As described later, the corpus
contained text samples from seven different context categories (see Table 2). We
expected that language would differ dramatically across these contexts. However,
our primary interest was in the interactions between gender and communication
task. We predicted that the overall picture of gender differences would persist
across context; but, in keeping with our bottom-up approach to this project, we re-
frained from making predictions about specific language categories. In addition,
previous research has identified differences between speaking and writing, such
that the latter is thought to involve more planning and complexity (e.g., Biber,
 TABLE 1
Main Effects of Gender on Language Use

Female Male

LIWC Dimension Examples M SD M SD Effect Size (d)

Linguistic dimensions
Word count 1,420 5,403 1,314 6,016 ns
Words per sentence 21.26 31.22 23.90 48.12 –0.07
Question marks 3.21 7.33 3.07 7.86 ns
Words $ six letters 13.99 4.42 15.25 5.91 –0.24
Numbers 1.37 1.31 1.59 1.55 –0.15
Negations no, never, not 1.85 1.10 1.72 1.17 0.11
Articles a, an, the 6.00 2.73 6.70 2.94 –0.24
Prepositions on, to, from 12.46 2.44 12.88 2.64 –0.17
Inclusive words with, and, include 6.42 1.88 6.34 2.03 ns
Exclusive words but, except, without 3.82 1.54 3.77 1.64 ns
Psychological processes
Emotions 4.57 1.99 4.35 2.07 0.11
Positive emotions happy, pretty, good 2.49 1.34 2.41 1.40 ns
Optimism certainty, pride, win 0.56 0.58 0.58 0.61 ns
Positive feelings happy, joy 0.61 0.61 0.51 0.65 0.15
Negative emotions 2.05 1.65 1.89 1.56 0.10
Anxiety nervous, afraid, 0.48 0.68 0.38 0.64 0.16
tense
Sadness grief, cry, sad 0.55 0.76 0.47 0.70 0.10
Anger hate, kill 0.61 0.81 0.65 0.92 ns
Swear words damn, ass, bitch 0.09 0.25 0.17 0.44 –0.22
Sensations 2.22 1.27 2.06 1.30 0.12
Feeling touch, hold, feel 0.58 0.67 0.47 0.66 0.17
Hearing heard, listen, sound 0.78 0.74 0.71 0.72 0.10
Seeing view, saw, look 0.72 0.78 0.74 0.83 ns
Cognitive processes 7.35 2.57 7.17 2.82 0.07
Causation effect, hence 1.02 0.76 1.02 0.88 ns
Insight think, know 2.40 1.28 2.28 1.38 0.09
Discrepancy should, would, 2.32 1.31 2.23 1.46 0.07
could
Tentative perhaps, guess 2.54 1.43 2.54 1.57 ns
Certaintyns always, never 1.35 0.94 1.21 0.96 0.14
Hedge verb phrases I + guess, I + reckon 0.57 0.67 0.50 0.67 0.11
Social processes
Social words 9.54 4.92 8.51 4.72 0.21
Communication talk, share, converse 1.26 0.95 1.20 0.95 ns
Friends pal, buddy, 0.37 0.51 0.33 0.53 0.09
coworker
Family mom, brother, 0.77 1.04 0.64 1.01 0.12
cousin
Humans boy, woman, group 1.22 1.33 1.15 1.33 ns
(continued)

219
220 NEWMAN, GROOM, HANDELMAN, PENNEBAKER

TABLE 1 (Continued)

Female Male

LIWC Dimension Examples M SD M SD Effect Size (d)

Pronouns 14.24 4.06 12.69 4.63 0.36

First-person singular I, me, my 7.15 4.66 6.37 4.66 0.17
First-person plural we, us, our 1.17 2.15 1.07 2.12 ns
Second person you, you’re 0.59 1.05 0.65 1.15 –0.06
Third person she, their, them 3.41 3.45 2.74 3.01 0.20
Time and space
Time till, started, hour 4.09 1.94 4.03 2.14 ns
Past-tense verb walked, were, had 4.36 2.97 4.02 2.84 0.12
Present-tense verb walk, is, be 11.71 4.00 10.98 4.10 0.18
Future-tense verb will, might, shall 1.27 1.03 1.33 1.18 ns
Space here, up, around 2.40 1.18 2.47 1.31 ns
Motion verbs walk, move, go 1.22 0.89 1.15 0.93 0.07
Current concerns
Occupation work, class, boss 2.34 1.88 2.59 2.10 –0.12
Money cash, taxes, income 0.25 0.39 0.29 0.49 –0.10
Metaphysical death, god 0.41 0.88 0.47 0.97 –0.06
Leisure house, TV, music 1.11 0.92 1.07 1.08 ns
Home house, kitchen, lawn 0.80 0.76 0.68 0.79 0.15
Sports football, play, game 0.13 0.30 0.19 0.50 –0.15
Physical functions ache, breast, sleep 1.33 1.22 1.28 1.29 ns
Sex lust, pregnant, gay 0.30 0.53 0.27 0.60 ns

Note. Means (except for word count and words per sentence) refer to percentages of the total words
in a sample. Effect size (Cohen’s d) was calculated by dividing the mean difference by the pooled stan-
dard deviation. Positive effect sizes mean women used the category more; negative effect sizes mean
men used it more. All mean differences except those labeled “ns” were significant at p < .001, based on
univariate statistics from a multivariate analysis of variance.

1991). Based on this literature, we also predicted that the size of the gender differ-
ence would be largest in spoken language (i.e., the conversation category) because
it is more natural and spontaneous (e.g., Biber, 1991).

Method
Text corpus.1 Our archive of electronic text samples represented 70 studies
from 22 laboratories. These laboratories included 14 universities in the United
States (63 studies), 1 university in New Zealand (4 studies), and 3 universities in
England (3 studies). Forty-four of the studies (63%) were conducted by at least one
of the authors. The studies were conducted over a 22-year period (1980–2002), and
included samples of fiction going back as far as the 17th century. All the files

1We are happy to make this corpus available to other researchers. Interested parties can contact

James W. Pennebaker for more information at pennebaker@mail.utexas.edu

 GENDER DIFFERENCES IN LANGUAGE USE 221

TABLE 2
Characteristics of Text Files, by Context and Participant Gender

Context No. Text Files Mean Word Count (SD) % Aged 18–22 % Written

Emotion
Men 3,603 1,566 (5,159) 62.7 96.7
Women 5,263 1,574 (4,724) 57.9 93.0
Time management
Men 520 299 (213) 82.3 96.9
Women 723 295 (201) 86.4 97.2
Stream of consciousness
Men 793 481 (246) 89.8 95.2
Women 1,033 561 (260) 88.5 96.2
Fiction
Men 37 21,593 (52,366) 0.0 100.0
Women 29 24,302 (55,593) 0.0 100.0
TAT–inkblot
Men 680 266 (175) 98.9 100.0
Women 996 311 (211) 97.3 100.0
Exam
Men 170 631 (636) 100 96.5
Women 90 706 (455) 93.3 97.8
Conversation
Men 168 3,466 (4,370) 54.1 0.0
Women 219 7,808 (7,575) 39.3 0.0

Note. N = 14,324 text files (5,971 men; 8,353 women). Age percentage refers only to the subset of
the sample (70.7%) for which age information was available. TAT = Thematic Apperception Test.

contained primary data from individual participants, either written (93%) or tran-
scribed from speech (7%). It is noteworthy that only about two thirds of the text
files were from college-age participants, in contrast to most psychological studies
of language to date. There was also a good mix of spoken and written samples,
with the entirely spoken conversation category being balanced by predominantly
written samples in the other categories.
After excluding files for which no gender information was available, and stud-
ies including only men or women, there remained text samples from 11,609 partic-
ipants, consisting of approximately 45,700,000 words. In many of the studies, par-
ticipants had provided multiple samples within a particular context, to aid in the
reliability of linguistic style (M = 3.5 samples per person, SD = 2.4). When this was
the case, samples were aggregated such that there was only 1 text file per person,
per context. Of the 11,609 participants, 2,130 contributed text samples in two or
more contexts. The aggregation process yielded 14,324 final text files, with 5,971
written by men and 8,353 written by women.
The corpus contained text samples from seven different context categories:
emotion, time management, stream of consciousness, fiction, Thematic Apper-
ception Test (TAT)–inkblot, exams, and conversation. Table 2 summarizes for
222 NEWMAN, GROOM, HANDELMAN, PENNEBAKER

each category the number of text files, the mean word count, the percentage of
college-age participants, and the percentage of written (vs. spoken) samples. The
emotion category contained language samples in which participants addressed
emotional (usually traumatic) life events. These included writing studies con-
ducted in a traditional laboratory setting and interviews in which participants
discussed such topics as their family histories or their thoughts and feelings
about traumatic events. The time management category contained language sam-
ples from the control conditions in writing studies in which participants wrote or
spoke about time management. The stream of consciousness category contained
language samples from college student participants who were asked simply to
track their thoughts and feelings as they occurred as part of class assignments in
at least 10 Introductory Psychology classes. The fiction category contained the
full text of fictional novels, mostly consisting of the top-selling fiction books
from the year 1996 (e.g., The Alienist by Caleb Carr, and K is for Killer by Sue
Grafton). The TAT–inkblot category consisted of participants’ free responses to
describing drawings of specific scenes (TAT) or standard inkblots typically used
in the Rorschach test. The exam category consisted of essays written for class
exams in psychology courses. Finally, the conversation category contained spo-
ken natural language samples, talk-show transcripts, and other nontraumatic
face-to-face interviews.

Text analysis. By default, LIWC analyzes each text on 74 language dimen-

sions. Two categories unique to conversation, nonfluencies (e.g., er, hmm, um) and
fillers (e.g., y’know), were excluded. We also excluded categories with very low
base rates (< 0.2%), such as TV and sleep, although we made an exception for two
low base rate categories, swear words and sports, because of their close relations to
gender stereotypes.
In addition, we created a new dictionary category of “hedges.” Previous re-
search (e.g., Mulac et al., 2000) suggested that women might be more likely to use
verb phrases such as “I figure” and “I guess” to qualify the impact of a sentence. To
capture hedges with LIWC, 12 graduate student raters were asked to decide
whether each of 43 candidate phrases was commonly used as a hedge, intended to
qualify a statement and reduce its force. Only those items with high interrater reli-
ability (agreement by at least 10 of the 12 raters) were selected for the final dictio-
nary. Our final data set represented 14,324 text files, analyzed along 54 language
dimensions.

Results
Main effects of gender on language. To answer the most basic question of
whether men and women in our sample used language differently, we performed a
MANOVA on all of the LIWC categories that met our inclusion criteria, with gender
 GENDER DIFFERENCES IN LANGUAGE USE 223

as the independent variable. As expected, this analysis yielded a significant

multivariate effect, F(53, 14,270) = 30.66, p < .001. Individual univariate analyses of
variance allowed us to identify the language dimensions that men and women used
differently. Table 1 presents the results of each univariate test, including effect sizes
(Cohen’s d) for significant analyses. Due to the large number of analyses being con-
ducted, as well as the exceptionally large sample size, only highly significant p val-
ues < .001 were considered significant. In addition, many of the significant effect
sizes are in or below the range generally considered small; Cohen (1992) recom-
mended a cutoff of .20 for a “small” effect. Given the bottom-up nature of this pro-
ject, we took a more liberal approach and confined our interpretations to effect sizes
of d = .10 or greater (see the Discussion section for more on this issue).
The table is organized in generally the same way as LIWC itself was con-
structed. The overarching structure distinguishes between (a) linguistic properties,
(b) words connoting psychological processes, (c) words connoting social pro-
cesses, (d) words that clarify temporal and spatial context, and (e) current concerns
that are wholly content rather than style oriented. The overall picture painted by
Table 2 is of a multitude of differences combined with a good deal of overlap
between the language of men and women.
Females’ language was more likely than men’s to include pronouns and social
words, a wide variety of other psychological process references, and verbs. Nega-
tions and references to the home were also features of the female profile. Men ex-
ceeded women on a number of linguistic dimensions including word length, num-
bers, articles, and prepositions. Men also discussed various current concerns more
frequently, and swore more often. It is worth noting that some of these word fre-
quencies are inextricably linked. For example, using more prepositions requires
using more articles (“to the store,” makes sense; “to store” does not). Consistent
with this, the Pearson correlation between gender and article use slightly decreased
when we controlled for prepositions (from r = .12 to r = .10).
Null effects were found for word count, question marks, first-person plural, and
a variety of relatively narrow categories including anger, gender, time, and space.
Table 1 highlights the fact that men and women did not merely focus on different
word categories (e.g., anxiety vs. anger, or emotion words vs. cognitive words), but
actually used different superordinate language dimensions. Female language em-
phasized psychological processes, social processes, and verbs. Male language em-
phasized current concerns. Thus, the results are consistent with the idea that men
and women employ language for different reasons.
To illustrate the main effects of gender on language, we selected representative
examples from the “stream of consciousness” communication context. This partic-
ular context was selected for two reasons: (a) The cell sizes were reasonably large
(793 men; 1,033 women), and (b) it imposed the fewest constraints on the content
of participants’ language. Excerpts from 4 representative women and 4 representa-
tive men are presented below.
224 NEWMAN, GROOM, HANDELMAN, PENNEBAKER

Examples of females’ language. In the following excerpts, female college

students tracked their thoughts as they occurred. Free from constraints, these
women made references to psychological processes (e.g., mad, uneasy, remember,
nervous), social processes (e.g., sister, friends), verbs (e.g., watching, taking, talk-
ing, thinking), and negations (e.g., can’t, not):

Female #1: Okay, well, I am watching this movie. I’m not really watching it because
I’m typing, but I’m listening to it. I really can’t type that well, so there are probably
going to be a few misspelled words. My sister made me mad a while ago because I
asked her to call me when her husband got home and she didn’t.

Female #2: Palms are sweaty, my stomach is uneasy, and my head just feels in pain.
I’m sick, I’m not supposed to get sick. I’m pre-med I’m supposed to be taking good
care, promoting health. I need to get better, it is essential that I get better.

Female #3: okay, so I’m sitting here talking to my friends. I miss them so much. they
live back home in houston. I wish I could see them just like old times. I remember
when we would all hang out at school together. it was great.

Female #4: Right now, I am thinking about my chemistry homework and test. I am
very nervous about it and I am worried that I may not succeed to my fullest potential.

Examples of males’ language. In the following excerpts, male college stu-

dents tracked their thoughts as they occurred. Free from constraints, these men
used longer words and more articles (e.g., “a bit,” “the music,” “a journal”). Men
also made references to current concerns (e.g., assignment, apartment), including
multiple references to the “stream of consciousness” instructions that are absent
from females’ language:

Male #1: Sorry for any grammar mistakes in this timed writing. There’s a bit of pres-
sure writing every thought you have within 20 minutes and try and make it com-
pletely coherent. The music in the back ground plays that of falling falling falling.

Male #2: I find it amusing that in writing a stream of consciousness about what I am
thinking, my mind is completely focused on what I am going to write in the stream of
consciousness paper. Thus, my stream of consciousness is about my stream of con-
sciousness about my stream of consciousness, etc.

Male #3: Stream of Consciousness? How do you start something so vague. I keep a
journal which I write in occasionally, but I can not remember the last time an assign-
ment consisted solely of write your thoughts.
 GENDER DIFFERENCES IN LANGUAGE USE 225

Male #4: Cool. I’m currently sitting in the campus library completing this assign-
ment because the Time Warner Cable people consistently refuse to show up to our
apartment to install our internet. The two guys next to me keep talking about web
sites and how they can improve the overall aesthetic beauty of the page by writing
some of the code in java script.

Age as a Moderator Recent research has suggested that language use also
varies according to an individual’s age (Pennebaker & Stone, 2003), and that gen-
der differences vary across children of different ages (Mulac et al., 1990). There-
fore, we repeated the previous analysis using linguistic categories that had been
adjusted for age. Age information was available for 70.7% of the sample, either
specific to each individual or, in relatively homogenous populations where the
group mean was known (primarily Introductory Psychology classes), by using the
mean to replace missing values. This resulted in a total of 10,131 text files with as-
sociated age information.
A significant multivariate effect of gender was again found for the age-adjusted
means, F(51, 10,079) = 17.73, p < .001. The pattern of univariate results was
nearly identical when the effects of age were controlled. Not a single effect
switched directions from female to male advantage or vice versa. Six previously
significant effects now failed to meet our alpha level (p < .001): second-person
pronouns, total cognitive words, discrepancies, hedge verb phrases, motion verbs,
and metaphysical references. This may be partly attributable to the loss of 29.3%
of the sample for which age information was unavailable. Nevertheless, the overall
picture was of gender differences in language use that remained unchanged when
age was controlled for.

Function Words Versus Other Categories We hypothesized that the larg-

est differences between males’ and females’ language would be found with func-
tion words. Function words, which tend to be processed in the brain differently
from more traditionally studied nouns and regular verbs, have been found to be
linked to emotional state such as depression (Rude et al., 2004), social connections
with others (Chung & Pennebaker, 2007), and how individuals think about the
world around them (Pennebaker et al., 2003). As a beginning way to examine the
nature of function words and gender differences, we compared the average effect
size for the function word categories to the average effect size for the content cate-
gories of nouns and verbs.
Because our interest was in the size rather than the direction of this analysis, we
used the absolute value of effect sizes that fit into each category (see Table 1). We
also limited this comparison to effect sizes that met our alpha criterion (p < .001).
The average effect size for verbs (feeling, hearing, insight, past-tense verbs, pres-
ent-tense verbs, and future-tense verbs) was d = .10, and the average effect size for
nouns (friends, family, occupation, money, metaphysical, home, and sports) was d
226 NEWMAN, GROOM, HANDELMAN, PENNEBAKER

= .11. The average effect size for function words (articles, prepositions, discrep-
ancy words, and total pronouns) was d = .21. Consistent with our predictions, the
size of the gender difference was roughly twice as large for function words as for
either nouns or verbs. However, it is also clear from looking at Table 1 that men and
women differed in their use of both content and function words.

Communication Context To examine the effects of communication context,

a 2 (Gender) × 7 (Context) MANOVA was performed on the set of 54 linguistic
categories. A significant multivariate main effect of context was found, F(432,
85,464) = 119.84, p < .001; in addition to a main effect for gender, F(72, 14,239) =
5.72, p < .001. These were qualified by a significant Gender × Context interaction,
F(432, 85,464) = 4.64, p < .001. Table 3 lists the 37 categories for which signifi-
cant interactions were found, together with effect sizes for the gender differences
and the form of the interaction.
In about one half of the cases (19 out of 37), the interactions were simply addi-
tive—that is, the gender differences differed in magnitude across the conditions,
but were in the direction reported in Table 1. The remaining 18 interactions were
crossovers—that is, the overall gender difference was reversed in at least one con-
dition; or, in the cases where no overall effect had been found, different conditions
showed effects in opposite directions. In nearly all of the crossover cases, the effect
was driven by reversals in conversation contexts. In conversation, men and women
still differed quite dramatically from each other, but not in the same ways that they
did in writing or more formal speech. In conversations, for example, men used
many more negations, negative emotion words, present-tense verbs, and refer-
ences to leisure activities, whereas women predominated in their references to
numbers. A closer examination of these findings suggests that men’s speech is
characterized by more negative emotion and references to the past relative to men’s
writing.
The bottom section of Table 3 presents average effect sizes for each condition,
computed using the absolute value of each language dimension. The average effect
sizes for the five superordinate categories show a complex pattern. Men’s and
women’s use of both linguistic categories and time and space references differed
most in the conversation context. In contrast, the use of psychological and social
processes both showed the largest gender differences in the fiction context. Finally,
gender differences in reference to current concerns were equally large in both the
fiction and conversation contexts.
Collapsing across all LIWC categories, the average overall effect sizes were
largest for fiction writing (d = .31), natural conversation (d = .26), and exam essays
(d = .22). Overall effect sizes were considerably smaller for stream of
consciousness (d = .11), TAT–inkblot (d = .09), emotional writing (d = .08), and
time management (d = .08). It is also noteworthy that these latter two conditions,
 TABLE 3
Significant Gender 3 Context Interactions

Effect Sizes

Crossover Time Stream of

LIWC Category Effect? Emotion Management Consciousness Fiction TAT–Inkblot Exams Conversation

Linguistic
Word Count .00 –.01 .23* .04 .16* .10* .50
Words per sentence Yes –.03 .03 –.12* –.14* –.01 .28* .00
Question marks Yes –.02 –.03 .04 –.04 .04 –.07* .49*
Words $ six letters –.16* –.04 –.09* –.32* –.17* –.45* –.44*
Numbers Yes –.09 –.15* –.13* –.37* –.10* –.09* .14*
Negations Yes .08 .04 .17* .58* .05 .13* –.22*
Articles –.21* –.07 –.33* –.70* –.22* –.05 –.77*
Prepositions –.11* –.09* –.12* –.26* –.09* –.11* –.74*
Inclusive words .04 .05 .03 –.38* .07 –.17* –.35*
Exclusive words Yes .03 .05 .03 .06 –.04 .13* –.51*
Psychological
Positive emotion .03 .03 –.03 .31* .15 –.06 .12*
Negative emotion Yes .09* .05 .05 .01 .03 –.26* –.30*
Anxiety Yes .13* .05 .09* .04 .12* –.30* .16*
Anger –.02 –.05 –.09* –.10* .04 –.18* –.43*
Swear words –.14* –.10* –.24* –.28* –.12* .06 –.43*
Senses Yes .10* .09* .12* .56* .03 .25* –.12*
See Yes .00 .02 –.08 .24* –.03 .14* –.32*
Hear .05 .21* .13* .45* .02 .22* .00
Discrepancies Yes .02 .03 .05 .57* .16* .34* –.13*
Tentative .01 .03 –.05 .25* –.06 .35* –.19*
Social
Social words .11* .20* .21* .74* .28* .56* .14*
Communication .02 .20* .13* .55* .02 –.08 .05
227

Humans Yes .06 .05 –.06 –.01 .01 .47* –.17*

continued)
228 TABLE 3 (Continued)

Effect Sizes

Crossover Time Stream of

LIWC Category Effect? Emotion Management Consciousness Fiction TAT–Inkblot Exams Conversation

Total pronouns .25* .20* .33* .79* .30* .46* .03

First-person singular .16* .00 .17* .27* –.01 .29* .10*
Second person –.07 .00 –.05 .32* –.08 .24* .33*
Third person .16* .18* .20* .66* .32* .51* –.08
Time and space
Space Yes –.04 .09* –.07 –.19* –.04 –.06 –.70*
Motion verbs Yes .06 .00 .10* .11* .05 .13* –.27*
Time Yes .05 –.06 –.04 .23* –.03 .21* –.23*
Past Yes .09* .04 .01 .43* .11* .05 –.10*
Present .14* .04 .17* .28* .09* .41* –.21*
Future Yes –.05 –.12* .02 .10* –.03 .32* –.16*
Current concerns
Money Yes –.07 –.07 –.04 –.18* –.05 .11* –.39*
Leisure Yes .06 –.14* –.03 .44* .09* .16* –.32*
Home .12* –.11* .20* .46* .15* .25* .02
Sports –.08 –.13* –.19* –.05 –.06 –.10* –.39*
Overall average: 0.08 0.08 0.11* 0.31* 0.09* 0.22* 0.26*
(absolute value)
Linguistic 0.08 0.06 0.13* 0.29* 0.10* 0.16* 0.42*
Psychological 0.06 0.07 0.09* 0.28* 0.08 0.22* 0.22*
Social 0.12* 0.12* 0.16* 0.48* 0.15* 0.37* 0.13*
Time and space 0.07 0.06 0.07 0.22* 0.06 0.20* 0.28*
Current concerns 0.08 0.11* 0.12* 0.28* 0.09* 0.20* 0.28*

Note. All interactions reported in this table are significant at p < .001 based on univariate analyses of variance. Effect size (Cohen’s d) was calculated by di-
viding the mean difference by the pooled standard deviation. Positive effect sizes mean women used the category more; negative effect sizes mean men used it
more. “Crossover effect” means that significant effect sizes were found in opposite directions across conditions. “Average” effect sizes at the bottom of the table
represent the mean absolute value across all language dimensions. LIWC = Linguistic Inquiry and Word Count.
*p < .001.
 GENDER DIFFERENCES IN LANGUAGE USE 229

both from studies of emotional writing, showed the largest differences in the use of
social process words.

DISCUSSION

Comparing the language of men and women in a large, heterogeneous sample of

written and spoken text reveals small but consistent gender differences in language
use. For the women who contributed 8,353 text files to the study, the English lan-
guage was more likely to be used for discussing people and what they were doing,
as well as communicating internal processes to others, including doubts.
Thoughts, emotions, senses, other peoples, negations, and verbs in present and
past tense figured high on the list of words that women used more than men. For
the men who contributed 5,970 files, language was more likely to serve as a reposi-
tory of labels for external events, objects, and processes. Along with discussion of
occupation, money, and sports were technical linguistic features such as numbers,
articles, prepositions, and long words. Swear words added emphasis to male lan-
guage. Contrary to popular stereotypes, men and women were indistinguishable in
their references to sexuality, anger, time, their use of first-person plural, the num-
ber of words and question marks employed, and the insertion of qualifiers in the
form of exclusion words (e.g., but, although).
It mattered little whether age was controlled for, although the map of gender
differences did vary somewhat across communication context. The primary differ-
ence was that men’s speech was characterized by more negative emotion and more
references to the past relative to men’s writing. It is interesting to note that the ef-
fect sizes were almost all larger in natural conversation, perhaps because spoken
language is more “natural” than writing (e.g., Biber, 1991). Similarly, natural lan-
guage, by its very nature, tends to be more informal and less constrained. Function
words, such as pronouns, are used at much higher rates in conversation—
especially by women. In addition, when given the freedom to talk about any topic,
men (but not women) elected to talk about concrete objects, which require nouns
and, of course, articles. Taken together, the general pattern of variation across con-
texts suggests that gender differences are larger on tasks that place fewer con-
straints on language use.
The effect sizes on all language dimensions were in the range generally consid-
ered small. In fact, only five dimensions met Cohen’s (1992) criterion for a small
effect when we collapsed across communication context—long words, articles,
swear words, social words, and pronouns. Although larger effects were seen in the
conversation and stream of consciousness contexts, these five dimensions appear
to be the most meaningful differences in males’ and females’ language. In inter-
preting the size of these effects, two considerations are worth noting.
230 NEWMAN, GROOM, HANDELMAN, PENNEBAKER

First, Eagly (1995) pointed out that a sizable portion of other gender differences
have effect sizes ranging from small to moderate. These results are particularly
compelling because of the diverse content of the text samples (i.e., some of the
samples came from experimental studies—ranging from writing about trauma to
describing a picture—whereas others came from fiction writing and natural con-
versations). Despite this, men and women used language in reliably and systemati-
cally different ways. Writing about a traumatic experience is very different from
writing a class exam, but men and women wrote differently across both contexts.
This mirrors the substantial intraindividual consistency in language use reported in
earlier work (Pennebaker & King, 1999).
Second, it is important to note the context in which samples were collected. On
the surface, the difference between using 14% pronouns and using 12% pronouns
seems rather subtle. However, these differences are based on an average of 15 min
of communication, comprising an average of 1,000 words. This means that women
used 140 pronouns compared to men’s 120 pronouns. These numbers translate into
a difference of roughly 2 to 3 pronouns every minute. Thus, gender differences in
written and spoken language appear to be subtle, but reliable. The fact that we are
confronted with these differences every day yet fail to notice them highlights the
degree to which they are a part of everyday life. At the same time, it is important to
keep in mind that these differences are averages at the population level. The impli-
cation of this fact is that predictions about language use by individuals should be
made cautiously, if at all.

Comparison With Previous Research

Many of the findings of our analyses echo those reported by earlier studies using
more traditional text analysis methods; other findings clarify disputed questions in
the literature. Table 4 summarizes the relations between previous data and these re-
sults. Past research has identified differences at four main levels of analy-
sis—words, phrases, sentences, and overall messages. Because LIWC is a word-
based text analysis system, the most direct correspondence to previous literature is
at the word level, and these results are presented first. However, many phrase-level,
sentence-level, and message-level features are associated with particular word
choices; and several of LIWC’s word categories serve as effective proxies. Human
coders and LIWC may have used slightly different coding strategies, but the basic
conclusions overlap.

Different words. Women’s greater use of pronouns mirrored previous work,

and the finding that women used more certainty words parallels earlier discoveries
that women used more intensive adverbs (Biber et al., 1998; McMillan et al., 1977;
Mehl & Pennebaker, 2003; Mulac et al., 2000). Successful replications for men’s
language included substantially greater use of numbers, articles, long words, and
 TABLE 4
Comparisons Between Previous Findings and Present Results

Previous Findings Relevant LIWC Category Present Results

Different words
F>M
Pronouns Total pronouns 3 Replication
Intensive adverbs Certainty words 3 Replication
M>F
Numbers Numbers 3 Replication
Articles Articles 3 Replication
Long words Words $ six letters 3 Replication
Swearing Swear words 3 Replication
Mixed
First-person singular First-person singular F>M
Emotion words Total emotion F>M
Negative emotion F>M
Positive feelings F>M
Anger F=M
Different phrases
F>M
Polite forms Discrepancies (e.g., could) 3 Replication, small effect
Hedging phrases Tentative (e.g., maybe) 7F=M
Hedges (e.g., suppose) 3 Replication
M>F
Locatives Space (e.g., above) 7M=F
Prepositions 3 Replication
Mixed
Oppositions Exclusive (e.g., but) F=M
Justifiers Causation (e.g., because) M=F
Different sentences
F>M
Long sentences Words per sentence M > F, small effect
Negations Negations (e.g., never) 3Replication
M>F
Word count Word count 7F=M
Mixed
Directives Second-person pronoun M > F, small effect
Questions Question marks M=F
Different messages
F>M
Personal concerns and Psychological words 3 Replication
interpersonal queries Social process words 3 Replication
M>F
Information exchange Numbers 3 Replication
Prepositions 3 Replication
Articles 3 Replication
Current concerns 3 Replication

Note. “Small effect” refers to effect sizes of d < .10. LIWC = Linguistic Inquiry and Word Count; M
= male; F = female.

231
232 NEWMAN, GROOM, HANDELMAN, PENNEBAKER

swearing (e.g., Gleser et al., 1959; Mehl & Pennebaker, 2003; Mulac & Lundell,
1986). Reflecting the mixed bag of earlier work on emotional references, women
use more affect words, but this was not restricted to positive emotions, as one ear-
lier study suggested (Danner et al., 2001). Women were more likely than men to re-
fer both to positive feelings and to negative emotions—specifically, sadness and
anxiety (cf. Thomas & Murachver, 2001; Mulac et al., 1990). The previous finding
of a male advantage in anger words was not replicated (Mehl & Pennebaker, 2003).
The most striking discovery was that women, not men, were the more prolific users
of first-person singular pronouns (i.e., I, me, and my).

Different phrases. LIWC’s discrepancy category approximates the category

of polite forms (e.g., “Would you mind if … ,” “Should I get the door?”). Fitting
with earlier work, we found a small but reliable tendency for such words to appear
more often in women’s texts (e.g., Holmes, 1995; McMillan et al., 1977). LIWC
has two categories relevant to the critical linguistic area of hedging, but these
yielded surprisingly different results. LIWC’s new hedging category, combining
“I” with a variety of verbs such as “guess,” found the expected gender difference:
Women were more likely to hedge (cf. Hartman, 1976; Mulac & Lundell, 1994;
Poole, 1979). However, women were no more likely to use words from the tenta-
tive category (e.g., maybe, perhaps). The use of phrases, such as “I guess,” may re-
flect previous findings that women use more polite forms (Holmes, 1995;
McMillan et al., 1977), and are reluctant to force their views on other people.

Different sentences. It is difficult to construct a negation—a sentence of

what something is not—without one of the words from LIWC’s negations category
(e.g., no, not, never). Fitting with the prior observation that women were more apt
to make such statements (Mulac & Lundell, 1986; Mulac, Lundell, & Bradac,
1986), we found a small difference favoring women in this domain. Previous liter-
ature has found that in terms of words used, it was actually men who consumed
more “airtime” (e.g., Dovidio et al., 1988). Our data found no evidence of any dif-
ferences in overall word count. Finally, we failed to find any tendency for women
to use question marks, contrasting with earlier reports that women asked more
questions and inserted more tag questions into their sentence (e.g., McMillan et al.,
1977; Mulac & Lundell, 1986).

Different messages. This is the most difficult level of analysis to relate to

LIWC’s word-count output because, by definition, it concerns what is implicit in
language rather than what is contained in language’s manifest features. However, it
is informative to consider the types of topics that men and women use their words
to talk about. This study provides strong evidence that women seem to have more
of a “rapport” style, discussing social topics and expressing internal thoughts and
feelings more often, whereas men “report” more often, describing the quantity and
 GENDER DIFFERENCES IN LANGUAGE USE 233

location of objects (e.g., Herring, 1993; Tannen, 1990). The absence of a differ-
ence in first-person plural may indicate that the word “we” is not a simple marker
of a communal, interdependent mindset (cf. Brewer & Gardner, 1996), rather than
indicating doubts about whether women really are rapport oriented.

CONCLUSION

Text analyses based on word count cannot, by their very nature, capture the context
in which words are used. Interpreting the gender differences is clearly a nuanced
matter. Part of our aim was to use LIWC technology to get a broader sample than
any hand-coded study could ever manage. A qualitative investigation of the gender
differences we have reported would be one useful avenue for future research. Such
an investigation would allow for a more complete explanation of the ways in which
social roles and relationships between speakers contribute to differences in lan-
guage use. It must also be acknowledged that our data came from a pre-existing ar-
chive of texts that had either been collected in our own laboratory or had been vol-
unteered by outside labs. However, the size and diversity of the dataset suggest that
a more extensive sample would not have altered the overall findings.
Coates and Johnson (2001) pointed out that the study of language provides a
uniquely “social” perspective on the study of gender differences. Given that our
understanding of other human beings is heavily dependent on language, the aver-
age differences in communication style that we report are likely to play a central
role in the maintenance of gender stereotypes and may perpetuate the perception of
a “kernel of truth” that underlies those stereotypes. However, it is important to note
that our analyses merely identify how men and women communicate differently,
without addressing the issue of why these differences exist. Gender differences in
language use likely reflect a complex combination of social goals, situational de-
mands, and socialization—just to name a few—but these data do not identify these
origins. Rather, our goal was to provide a clear map of the differences in men’s and
women’s language, and to offer a starting point for future research into the nature
and origin of gender differences.
Our analyses demonstrate small but systematic differences in the way that men
and women use language, both in terms of what they say and how they choose to
say it. Although our focus was more on function words than content words, it is
clear that both types offer numerous opportunities for future research. By using a
very large and diverse data corpus combined with a computerized text analysis
program we were able to put the controversial topic of language-based gender dif-
ferences on firmer empirical ground. Furthermore, our data support and clarify,
rather than contradict, previous research, suggesting that word-count strategies are
a viable, highly efficient alternative to linguistic analysis based on human coders.
Computerized text analysis offers the statistical power and coding consistency that
234 NEWMAN, GROOM, HANDELMAN, PENNEBAKER

are ultimately essential for a complete answer to the questions that have captured
the imagination of laypeople and scientists alike: when, where, why, and how do
men and women talk differently?

ACKNOWLEDGMENTS

Preparation of this project was supported by grants from the National Institutes of
Health (MH52391) and the Army Research Institute (W91WAW-07-C-0029). We
thank Matthias Mehl and Barbara Luka for their helpful comments.

REFERENCES

American Psychiatric Association. (2000). Diagnostic and statistical manual of mental disorders (4th
ed., text revision). Washington, DC: Author.
Biber, D. (1991). Variation across speech and writing. Cambridge, England: Cambridge University
Press.
Biber, D., Conrad, S., & Reppen, R. (1998). Corpus linguistics: Investigating language structure and
use. Cambridge, England: Cambridge University Press.
Bradley, P. H. (1981). The folk linguistics of women’s speech: An empirical examination. Communica-
tion Monographs, 48, 73–90.
Brewer, M. B., & Gardner, W. (1996). Who is this “we”? Levels of collective identity and self represen-
tations. Journal of Personality & Social Psychology, 71, 83–93.
Brownlow, S., Rosamon, J. A., & Parker, J. A. (2003). Gender-linked linguistic behavior in television
interviews. Sex Roles, 49, 121–132.
Bucci, W., & Freedman, N. (1981). The language of depression. Bulletin of the Menninger Clinic, 45,
334–358.
Carr, C. (1996). The Alienist. New York: Random House.
Chung, C. K., & Pennebaker, J. W. (2007). The psychological functions of function words. In K. Fiedler
(Ed.), Social communication (pp. 343–359). New York: Psychology Press.
Coates, L., & Johnson, T. (2001). Towards a social theory of gender. In W. P. Robinson & H. Giles
(Eds.), The new handbook of language and social psychology (pp. 451–464). New York: Wiley.
Cohen, J. (1992). A power primer. Psychological Bulletin, 112, 155–159.
Colley, A., Todd, Z., Bland, M., Holmes, M., Khanom, M., & Pike, H. (2004). Style and content in
emails and letters to male and female friends. Journal of Language and Social Psychology, 23,
369–378.
Crosby, F., & Nyquist, L. (1977). The female register: An empirical study of Lakoff’s hypothesis. Lan-
guage in Society, 6, 313–322.
Danner, D. D., Snowdon, D. A., & Friesen, W. V. (2001). Positive emotions in early life and longevity:
Findings from the nun study. Journal of Personality & Social Psychology, 80, 804–813.
Dovidio, J. F., Brown, C. E., Heltman, K., Ellyson, S. L., & Keating, C. F. (1988). Power displays be-
tween women and men in discussions of gender-linked tasks: A multichannel study. Journal of Per-
sonality & Social Psychology, 55, 580–587.
Dubois, B. L., & Crouch, I. (1975). The question of tag questions in women’s speech: They don’t really
use more of them, do they? Language in Society, 4(3), 289–294.
 GENDER DIFFERENCES IN LANGUAGE USE 235

Eagly, A.H. (1995). The science and politics of comparing women and men. American Psychologist,
50, 145–158.
Gleser, G. C., Gottschalk, L. A., & John, W. (1959). The relationship of sex and intelligence to choice of
words: A normative study of verbal behavior. Journal of Clinical Psychology, 15, 183–191.
Grafton, S. (1996). K is for Killer. New York: Ballantine Books.
Hartman, M. (1976). A descriptive study of the language of men and women born in Maine around
1900 as it reflects the Lakoff hypotheses in language and woman’s place. In B. L. Dubois & I. Crouch
(Eds.), The sociology of the languages of American women (pp. 81–90). San Antonio, TX: Trinity
University Press.
Herring, S. C. (1993). Gender and democracy in computer-mediated communication. Electronic Journal of
Communication, 3(2). Retrieved June 3, 2003, from http://www.cios.org/getfile/ HERRING_V3N293
Holmes, J. (1995). Women, men and politeness. Harlow: Longman.
Koppel, M., Argamon, S., & Shimoni, A. (2003). Automatically catagorizing written texts by author
gender.. Literary and Linguistic Computing, 17, 101–108.
Lakoff, R. (1975). Language and woman’s place. New York: Harper Colophon Books.
McMillan, J. R., Clifton, A. K., McGrath, D., & Gale, W. S. (1977). Women’s language: Uncertainty or
interpersonal sensitivity and emotionality? Sex Roles, 3, 545–559.
Mehl, M. R., & Pennebaker, J. W. (2003). The sounds of social life: A psychometric analysis of stu-
dents’ daily social environments and natural conversations. Journal of Personality & Social Psychol-
ogy, 84, 857–870.
Mulac, A., Bradac, J. J., & Gibbons, P. (2001). Empirical support for the gender-as-culture hypothesis:
An intercultural analysis of male/female language differences. Human Communication Research,
27, 121–152.
Mulac, A., & Lundell, T. L. (1986). Linguistic contributors to the gender-linked language effect. Jour-
nal of Language & Social Psychology, 5, 81–101.
Mulac, A., & Lundell, T. L. (1994). Effects of gender-linked language differences in adults’ written dis-
course: Multivariate tests of language effects. Language & Communication, 14, 299–309.
Mulac, A., Lundell, T. L., & Bradac, J. J. (1986). Male/female language differences and attributional
consequences in a public speaking situation: Toward an explanation of the gender-linked language
effect. Communication Monographs, 53, 115–129.
Mulac, A., Seibold, D. R., & Farris, J. L. (2000). Female and male managers’ and professionals’ criti-
cism giving: Differences in language use and effects. Journal of Language & Social Psychology,
19(4), 389–415.
Mulac, A., Studley, L. B., & Blau, S. (1990). The gender-linked effect in primary and secondary stu-
dents’ impromptu essays. Sex Roles, 23, 439–469.
Mulac, A., Wiemann, J. M., Widenmann, S. J., & Gibson, T. W. (1988). Male/female language differ-
ences and effects in same-sex and mixed-sex dyads: The gender-linked language effect. Communica-
tion Monographs, 55, 315–335.
Newman, M. L., Pennebaker, J. W., Berry, D. S., & Richards, J. M. (2003). Lying words: Predicting de-
ception from linguistic styles. Personality and Social Psychology Bulletin, 29, 665–675.
Oberlander, J., & Gill, A. (2006). Language with character: A stratified corpus comparison of individ-
ual differences in e-mail communication. Discourse Processes, 42, 239–270.
Pennebaker, J. W., Francis, M. E., & Booth, R. J. (2001). Linguistic Inquiry and Word Count (LIWC):
LIWC2001. Mahwah, NJ: Lawrence Erlbaum Associates, Inc.
Pennebaker, J. W., & King, L. A. (1999). Linguistic styles: Language use as an individual difference.
Journal of Personality & Social Psychology, 6, 1296–1312.
Pennebaker, J. W., & Lay, T. C. (2002). Language use and personality during crises: Analyses of Mayor
Rudolph Giuliani’s press conferences. Journal of Research in Personality, 36(3), 271–282.
Pennebaker, J. W., Mehl, M. R., & Niederhoffer, K. (2003). Psychological aspects of natural language
use: Our words, our selves. Annual Review of Psychology, 54, 547–577.
236 NEWMAN, GROOM, HANDELMAN, PENNEBAKER

Pennebaker, J. W., & Stone, L. D. (2003). Words of wisdom: Language use over the lifespan. Journal of
Personality & Social Psychology, 85, 291–301.
Poole, M. E. (1979). Social class, sex, and linguistic coding. Language and Speech, 22, 49–67.
Rude, S. S., Gortner, E. M., & Pennebaker, J. W. (2004). Language use of depressed and depres-
sion-vulnerable college students. Cognition and Emotion, 18, 1121–1133.
Shapiro, D. (1989). Psychotherapy of neurotic character. New York: Basic Books.
Strainchamps, E. (1971). Our sexist language. In V. Gornick & B. K. Moran (Eds.), Woman in sexist so-
ciety (pp. 240–250). New York: Basic Books.
Thomson, R., & Murachver, T. (2001). Predicting gender from electronic discourse. British Journal of
Social Psychology, 40, 193–208.
Weatherall, A. (2002). Gender, language, and discourse. London: Routledge.
Weintraub, W. (1981). Verbal behavior: Adaptation and psychopathology. New York: Springer.