Niche as a Determinant of Word Fate in Online Groups

Patterns of word use both reflect and influence a myriad of human activities and interactions. Like other entities that are reproduced and evolve, words rise or decline depending upon a complex interplay between their intrinsic properties and the environments in which they function. Using Internet discussion communities as model systems, we define the concept of a word niche as the relationship between the word and the characteristic features of the environments in which it is used. We develop a method to quantify two important aspects of the size of the word niche: the range of individuals using the word and the range of topics it is used to discuss. Controlling for word frequency, we show that these aspects of the word niche are strong determinants of changes in word frequency. Previous studies have already indicated that word frequency itself is a correlate of word success at historical time scales. Our analysis of changes in word frequencies over time reveals that the relative sizes of word niches are far more important than word frequencies in the dynamics of the entire vocabulary at shorter time scales, as the language adapts to new concepts and social groupings. We also distinguish endogenous versus exogenous factors as additional contributors to the fates of words, and demonstrate the force of this distinction in the rise of novel words. Our results indicate that short-term nonstationarity in word statistics is strongly driven by individual proclivities, including inclinations to provide novel information and to project a distinctive social identity.


Introduction
Much information about the fabric of modern human society has been gleaned from large-scale records of human communications activities, such as time stamps and network structures for email exchanges, mobile phone calls, and Internet activity [1][2][3][4]. But the flow of words has the potential to be even more informative. Words characterize both external events and otherwise unobservable mental states. They tap into the variety of experience, knowledge, and goals of different interacting individuals. The word stream is information-dense, because the number of distinct words and expressions is so great. The lexicon of a literate adult is estimated to contain over 100,000 distinct items [5], and it continues to grow as new words are encountered [6].
Records of the linguistic transactions within a community provide an ongoing statistical sampling of the vocabulary of a language. The sample at any time reflects both the social context (who is speaking, and to whom) and the topical context (what they are speaking about). But the language dynamics does not just passively mirror the context. Language adapts to new circumstances and needs through lexical innovation [7]. Large datasets available from the Internet provide an unprecedented opportunity to study the dynamics of words, as well as phrases and tags [8][9][10][11]. Here, we explore lexical fluctuations in relation to both individuals and topics by analyzing records of Usenet groups. Created over one decade before the World Wide Web, the Usenet groups were amongst the first systems for world-wide exchange of messages on the Internet. Usenet archives reveal the rise of "Netspeak", the language nowadays widely used on the Internet and in telephone text messages [12]. The groups we studied, rec.music.hip-hop and comp.os.linux.misc, were selected for their great lexical creativity. In these datasets, users serve as proxies for individuals, and threads as proxies for topics (see Methods). Our study goes beyond the analysis of user activity in Usenet groups [13], and focuses instead on the content of the messages.
It is known that word frequency is a factor in frequency dynamics on historical time scales [14,15], a finding that is expected from models of language learning across human generations [16]. Here, we identify two new factors-the dissemination of words across individuals (users) and the dissemination of words across topics (threads)-and we develop a method to quantify dissemination that controls for word frequency. Because words are acquired and reproduced by users as they communicate with each other about different topics, these two dissemination measures serve to characterize two important dimensions of the word niche. We apply these measures to demonstrate that dissemination is a much more powerful determinant of word fate than word frequency is; poorly disseminated words are more likely to experience a frequency reduction than widely disseminated words.
These results suggest analogies between word fates and the fates of biological species. In population biology, the term niche refers to the relationship between a species and the aspects of its environment that enable it to live and reproduce. Quantifying the breadth and versatility of a species' niche, as distinct from the species' sheer abundance, is key to understanding its competitive position within an ecosystem [17]. The geographic size of the niche is a statistical correlate of species duration, as species with large ranges are less likely to become extinct [18,19]. Analogies between language and population biology have proved fruitful in understanding the dynamics of entire languages, in particular the relationship of community size to overall rates of linguistic change [20,21] and to properties of the syntactic and morphological systems [22]. Here, we work at a more fine-grained level, quantifying the impact at short (two-year) time scales of the heterogeneous usage of language inside a community. Because we consider the role of heterogeneity amongst people within the community, the results also support comparisons between the dynamics of the linguistic system and other social dynamics, such as the spread of opinions or the popularity of news items, videos, and music [23,24].
The relation with social dynamics is strengthened by a case study of novel words with rising frequency, in which we compare a set of words for products and public figures to a set of slang words. The rise in use of words in the first set is mainly driven exogenously by events that are external to the Usenet group, such as product releases, political crises, and public performances. Because the use of slang words is strongly influenced by the social values and patterns of communication within any given linguistic group [25,26], the use of the (slang) words in the second set should be more influenced by factors endogenous to the Usenet community. The force of this distinction in word dynamics mirrors its force in other social behaviors, ranging from YouTube viewing to scientific discoveries, marketing successes, financial crashes, and civil wars [27,28]. Finally, we explore the correlations between individuals and topics as dimensions of word dissemination. The two dimensions are shown to be separable, and individual choices prove to be more important than topic in determining patterns of word usage. These results highlight the importance of individuality in the use of language, and imply limits on the role of social influence and social conformity.

Dissemination of words across users and threads
If everyone knew the same words, and chose to use them at random with their given frequencies, the dissemination of words across users would be the result of a Poisson process. We are interested in the extent to which the actual number of users of each specific word deviates from this baseline model. We define the measure of dissemination of each word w across users as where N w is the number of occurrences of the word in the dataset, U w is the actual number of users whose posts include word w at least once, andŨ is the expected number of users predicted by the baseline model. The latter is determined fromŨ = N U i=1Ũ i , where N U is the number of users andŨ i is the probability that user i used w at least once when all the words in the text are shuffled randomly (see Methods). Dissemination across threads is analogously defined as where T w is the number of threads in which the word appears, andT is the corresponding expected value from the baseline model. The word frequency is defined as f = N w /N A , where N A = w N w is the total number of words in the dataset; N w is a count, and the frequency f normalizes this count to a probability. In the rest of the paper, we focus on the properties of the dissemination measures D U w and D T w , or D U and D T for notational simplicity.
The expected value of D U is 1 for a word of any frequency that is distributed randomly across users. D U > 1 indicates over-disseminated words and D U < 1 indicates concentrated or clumped words. For example, in a half-year window centered on 1998-01-01 in the comp.os.linux.misc group, the words thanks and redhat have almost identical frequencies, but contrast in their dissemination (thanks: N w = 4, 121, D U = 1.19; redhat: N w = 4, 146, D U = 0.75). A similar contrast is provided for the same time window in the rec.music.hip-hop group by the words please (N w = 2, 336, D U = 1.17) and article (N w = 2, 366, D U = 0.59). The measure D U exhibits a lower bound determined by the number of occurrences of the word: 1 Nw ≤ D U . For any given set of posts, there is also an upper bound determined by the relationship of N w and N U toŨ : D U ≤ min{N w , N U }/Ũ . Due to the discreteness close to the lower bound, we set a threshold N w > 5 for the computation of D U,T . The few dozen most frequent words (mainly common function words) are also omitted from our analysis, because D U is not informative when N w is too large compared to the number of users. Figure 1 shows results on the expected statistical fluctuation around D U = 1 for randomly distributed words in a representative window of each Usenet group, as determined by a Monte Carlo simulation. The upper and lower extremes of the fluctuation depend on frequency, but only slightly.
The dissemination across threads D T is closely related to the residual inverse document frequency (r-IDF ), a measure used in text processing to characterize the extent to which a word is associated with particular documents [29,30]. IDF , defined as the reciprocal of the number of documents in which the word occurs, is strongly influenced by word frequency. Residual IDF addresses this artifact by taking the difference r-IDF = log(T ) − log(T ), whereT is approximated using a Poisson baseline model with equal document lengths. When this condition holds, − log(D T ) = r-IDF . The measure D T is a generalization of r-IDF that remains valid when the lengths of the documents are very unequal, as for the present datasets (see Supporting Information S1, Figure S1).

D U and D T as predictors of word fate
To explore the changes over time in the statistical attributes of words, we begin by partitioning each dataset into non-overlapping half-year windows. Figure 1 displays the behavior of D U within a representative half-year window for both groups. Most words are significantly clumped. At all word frequencies, the median D U falls below the 10th percentile for random fluctuation of the expected value under the baseline model. For words with log 10 f < −3.5, D U varies considerably and is not correlated with frequency f . Words with log 10 f > −3.5 are extremely high-frequency words, and comprise less than 0.5% of all distinct words in this window. But even these words are somewhat clumped. These findings are reproduced in all half-year windows for both Usenet groups, as summarized in Figure 2AB. They provide the user counterpart to prior observations of clustering of words in documents and in time [8,[29][30][31][32].
We now examine D U as a predictor of frequency change for words over two-year periods. We first note that D U is strongly related to the likelihood that a word with N w > 5 in a window t 1 falls below this threshold in a window t 2 taken two years later. This is illustrated for both Usenet groups in Figure 3AD, where t 1 and t 2 mark the centers of the half-year windows. The finding is so statistically robust that it is reproduced for every choice t 1 and t 2 = t 1 + 2 years, in both groups. The same pattern is also mirrored in the frequency changes of words that are above the N w > 5 threshold at both t 1 and t 2 . Within this group of words in the selected window of comp.os.linux.misc, D U is a strong predictor of whether the word rose or fell in frequency ( Figure 3B). In the selected window of rec.music.hip-hop, D U is likewise a strong predictor of the changes in word frequencies ( Figure 3E). The consistency of this pattern over all windows may be seen by comparing ∆ log 10 f for words with D U = 0.4 and with D U = 1.0, values that span the well-populated portion of the range in D U . Words with the former value tend to decline in frequency (∆ log 10 f is negative), while words with the latter value tend to maintain or increase their frequencies (∆ log 10 f is near zero or positive). There is no t 1 , t 2 pair for either dataset in which the effect is reversed ( Figure 3CF). This far, our analysis has focused on D U . In sociolinguistic parlance, we have considered the "indexicality" of words, that is the extent to which words are associated with individuals or types of people. Now, let us also consider D T , our measure of "topicality" (dissemination across topics). As shown in Figure 2CD and in Figure 4, the results just described for D U also hold for D T . The connection between D T and frequency change agrees with Ref. [33]'s study of foreign borrowings in news articles. What is the relative importance of these factors in predicting frequency change? As Table 1 shows, D U is more important than D T . Moreover, both are more important than log 10 f , whose importance is comparatively slight, as shown in Figure 5.
Words change over time not just in their frequency, but also in their dissemination. A signal aspect of changes in D U,T is a strong negative correlation with frequency change (∆ log 10 f ). For comp.os.linux.misc, the correlations of ∆ log 10 f with ∆D U and ∆D T are −0.54 and −0.40, respectively; for rec.music.hip-hop, −0.55 and −0.39, respectively. These negative correlations can be understood by comparing two scenarios. In one scenario, a word rises in frequency because it becomes more widely used; it is used by more individuals and/or in the discussion of more topics. In this scenario, the increase in frequency is accompanied by steady or increasing values of the dissemination measures D U,T . In a contrasting scenario, a word rises in frequency without a concomitant increase in the number of users and/or topics, because it is used more repetitively by the same few people and/or in discussing the same topics. In this scenario, the increase in frequency is accompanied by decreasing values of D U,T , because the use of the word becomes more and more concentrated in comparison to what the random baseline would predict. In this case, it follows from Figure 3 that the resulting low D U,T puts the word at risk of declining in frequency thereafter. Just as a population that explodes in a narrow ecological niche may well crash later, it appears that repetitive communications are more discounted than emulated by others. This picture broadly resembles recent observations about buzzwords in the blogosphere, which are reported in Ref. [11] to exhibit great fluctuations in their frequencies, as well as an apparent association between a fast rise and subsequent obsolescence. The fact that the correlations of frequency change (∆ log 10 f ) with dissemination change (∆D U and ∆D T ) are strongly negative means that the second scenario is the dominant one in our datasets. Overall, fluctuations in frequency driven by variability in user behavior and topic dominate the statistical behavior, with the result that patterns similar to those in Figures 3 and 4 are also observed by making the same calculations in the reversed time direction (that is, by relating D U,T at t = t 2 to −∆ log 10 f ). These large, short-term fluctuations add an important new dimension to the study of the long-term dynamics of language, as any novel expression must survive in the short term to survive in the long term.

Case study: Rising slang and product words
A new word must establish itself in a niche to survive in the language. The survival rate of lexical innovations is not known, but any successful innovation must have overcome short-term fluctuations in f that risked driving it to an early extinction. We now present a case study of successful innovations. First we identify all words that were not used during the first years of the group, and that were consistently used for at least some years thereafter (for precise thresholds, see Supporting Information S1, Text S2). From this collection of rising words, we selected two sets of words for each group. The first set is designated as P-words because they refer to products (such as gnome, a desktop environment introduced in 1998) and public figures (such as eminem, a rapper popular from the late 1990's). Exogenous factors contribute strongly to their use. The second set, designated as S-words, exemplifies slang words and other novel vernacular language. These novel words were selected with the aid of on-line dictionaries of Internet and Usenet terms (see Supporting Information S1, Text S2). We consider the dynamics of these words to be more dominated by factors endogenous to the linguistic systems and social networks of the Usenet groups. Although many of the S-words may have been learned from people outside of a Usenet group, such as celebrities seen on television, the group itself is the locus of the the social values and conventions that lead to some celebrities being imitated and others ignored. Paired lists of P-words and S-words were frequency matched to the extent possible. The words and their statistics are listed in Supporting Information S1, Tables S1-S4. Figure 6 compares the dynamics of example P-words and S-words. Temporal fluctuations in the total activity of the group ( Figure 6CD) provide a backdrop for considering the different fluctuations in the number of occurrences of some typical P-words and S-words ( Figure 6AB). Our Usenet database also allows us to go beyond the frequency dynamics of words over time, as explored in Ref. [34]'s recent study of words in books, and look at the roles of topics and individuals in determining this dynamics. In Figure 7, we show the behavior of the words in a frequency-D U,T space. As indicated by the horizontal boxplots, the P-words and S-words are located in the frequency region below log 10 f = −3.5, in which the frequency is not correlated with D U,T . Trajectories over time for two example words are superimposed, beginning when the words first reach N w > 5. In contrast to the example S-words, the example P-words begin with very low D U values, and rise greatly in frequency before becoming widely disseminated. The vertical boxplots show that P-words have overall lower D U,T than S-words (though both fall below the median of all words). The contrast in D U,T over the entire period is replicated if we consider just the early rising period of each of the words in both groups (see the aggregated statistics displayed in Figure 7, and further details in Supporting Information S1, Tables S1-S4).
Significant clumping in D U is expected for S-words, because choices of vernacular language such as lol (laughing out loud) and prolly (probably) reflect the individual's construction of social identity [35,36]. How can we construe the finding that P-words are even more clumped in D U than the S-words are? Recalling that all of the words in the case study were preselected to exemplify rising trends, it seems possible that the highly clumped P-words reflect the distinctive information access of their users. For example, gnome, which has a D U value of 0.46 in its early rising period, refers to a graphical desktop environment that was originally created by two Mexican programmers, Miguel de Icaza and Federico Mena. By discussing their experience with this interface, its early adopters bring information to the comp.os.linux.misc group that other users do not yet have. In short, by contributing posts about experiences and activities external to the Usenet group, a small number of users can be the vehicle for exogenous factors to come to influence the vocabulary of the group more generally.
The low D U of the P-words and S-words would tend to predict a decline in frequency (see above), but instead the frequencies of these particular words rose. For P-words, the rise is driven by events external to the Usenet community. For example, the P-word ssh (from comp.os.linux.misc) refers to the secure shell network protocol. The invention of ssh allowed people to carry out remote file transfers without compromising sensitive in-formation such as passwords. The immediate adoption of this technological improvement is clearly one reason for the rise in use of the word ssh. In rec.music.hip-hop, the use of the P-words bush, saddam, and iraq reflects discussion about the war in Iraq. Both the war, and the political events leading up to it, took place outside of the Usenet community. In Figure 6B, the 2005 rise in the frequency of eminem reflects heavy media coverage of his possible retirement. The use of the P-words also reflects endogenous factors to some extent. The fact that bush, saddam, and iraq met the inclusion criteria in rec.music.hiphop, but not in comp.os.linux.misc, suggests that a shared interest in politics is more important within the Usenet hip-hop community than in the Usenet linux community.
However, for the S-words, we consider that the endogenous factors were even more important. For these words, there are alternative ways of referring to the same general concept. In both groups, lol competes with rofl (rolling on the floor laughing), ha-ha, and other expressions. In rec.music.hip-hop, addy competes with address. In comp.os.linux.misc, y2k competes with year 2000, and boxen (as a plural of box, generalizing the jocular plural of Vaxen for the Vax brand of computers) competes with boxes, servers, computers, etc. The choice of one such word over an alternative expression with the same referent reflects the social value associated with the word, which is a nonreferential component of its meaning. By their nature, slang words stand out from other words through being used to "establish or reinforce social identity or cohesiveness within a group, or with a trend or fashion in society at large" [25]. In African-American Vernacular English (the original language of hip-hop), the transitory slang expressions of various subgroups of speakers, such as teenagers and musicians, serves to differentiate them within a larger African-American community sharing a rather stable lexicon and grammar [26]. Reference [12] suggests that on-line groups are especially likely to use jargon and slang as a means of constructing and affirming group solidarity, since the group has no identity outside of its on-line communications. But the use of some S-words also reflects exogenous factors to some extent, which may help explain their success despite the relatively low dissemination. The invention of cell-phone texting probably contributed to the availability of acronyms as slang expressions, the rise of server farms probably contributed to the need for a way to refer to computers as fungible units, and the linguistic influence of a particular rapper might have increased after a successful performance. However, these factors seem weaker than for the P-words, because they do not appear to dictate the particular choice of word out of all the alternatives. Related cases of social dynamics for which a combination of exogenous and endogenous factors has been considered include music downloads [23] and popularity patterns for YouTube videos and for stories on the news portal Digg [37,38].
By having the lowest overall distribution of D U values, the P-words contrast with all other rising words, including both the S-words and typical words whose frequencies increased (as exemplified in Figure 3BE by data points in the upper-right quadrant of each panel). This suggests that exogenous forcing is more efficient than other kinds of forcing. The fact that S-words had higher D U values overall than the P-words did, with no S-word rising from as low a D U value as the lowest P-words, makes the S-words appear more similar to words in general. In the absence of strong forcing by external events, the social dynamics within the group dominates the word dynamics, with reinforcement by peers providing a natural mechanism for the words to rise. The results support our understanding of D U as a determinant of frequency change; high D U values provide an index of the fact that relatively many different users provide examples of use of a specific word that others may imitate. The D U values for S-words are somewhat low compared to the distribution for all words. We can speculate about the mechanisms for this outcome. Exogenous factors in the use of S-words, mentioned just above, may play a greater role than is typical for words in general. Moreover, the force and emotions associated with the social value of the S-words may provide an additional factor driving the dynamics.
Most of our principal observations about the dissemination across users (D U ) of Pwords and S-words are also true for the dissemination of the same words across topics (D T ), as shown by comparing Figure 7AB to Figure 7CD. Given that the measures D U and D T both quantify the relative extent of the word niche, these detailed parallels in the behavior of the two measures raise the question of how many dimensions we are really dealing with. Since people form social groupings around shared interests [39,40], and choose words that express solidarity with these same groupings, do the two dimensions of indexicality and topicality reduce to just one underlying dimension? Or are the two dimensions separable, even if related through complex interactions? We take up these questions rigorously in the next section.

Factoring the relative contributions of individuals and topics
We have shown that most words, including both highly indexical words such as slang words and highly topical words such as products, are significantly concentrated in both D U and D T . We have sketched some reasons for these dimensions to be positively correlated. How can we rigorously evaluate their separability and relative importance? To address this issue, we consider new measures that effectively factor indexicality and topicality as contributors to D U,T , and we standardize the datasets to eliminate distributional artifacts.
We first introduceD U as a modification of D U in whichŨ in Eq. (1) is calculated from a baseline model that shuffles the words only within threads, rather than across all users and all threads. Analogously, we introduceD T as a modification of D T in which T in Eq. (2) is calculated from a baseline model that shuffles the words only within posts of the same user. These new quantities provide a direct measure of the extent to which individuals and topics contribute to the concentration of words observed above. While D U reveals whether the word is clumped or over-disseminated by comparing the actual dissemination with that obtained by "erasing" all the structure,D U maintains the structure of the threads and considers randomization of words across users within them. IfD U is significantly closer to 1 than D U is, then topics must strongly influence the individuals' choice of words. Analogously, the role of individuals can be confirmed by comparing the extent to whichD T is closer to 1 than D T is.
To ensure that users and threads serve as comparable proxies of individuals and topics, we randomly trim the datasets to eliminate the differences in their distributions that are visible in Supporting Information S1, Figure S1. For each window, the trimming scheme standardizes the user contribution per thread and the size of all posts, matches the number of users and threads, and approximately matches the distribution of posts per user and per thread (see Supporting Information S1, Text S3 and Figure S2). The trimmed comp.os.linux.misc (rec.music.hip-hop) dataset remains large enough for our statistical analysis, with an average of 4, 593 (1, 503) posts and 2, 383 (585) users and threads per half-year window, and an overall average of 77.6 (51.2) words per post.
The exact distributions of values of D U and D T change with the trimming. Trimming generally increases D U and D T for the words that survive, but the trends and all conclusions from previous sections still stand. For example, the overall median D U changes from 0.71 to 0.87, and the overall median D T changes from 0.73 to 0.89, for the comp.os.linux.misc group. The relative differences in both groups remain essentially unchanged, which means that the measures D U,T provide meaningful comparisons even when the distributions are not streamlined. However, the trimmed set offers the advantage of providing exact and non-artifactual information about the correlations between the measures. Table 2 displays the important correlations amongst the original and modified measures. The correlation between D U and D T is positive, confirming the expectation that indexicality and topicality are related. But it is far less than 1, suggesting that D U and D T contribute substantially different information. The measures D U andD U , as well as D T andD T are positively correlated, as expected because these are related measures by definition. Finally, the negative correlation betweenD U andD T is a confirmation that these quantities partially factor D U and D T and hence provide the information they are designed to provide. Notice that this negative correlation is possible, despite the positive correlation of the other pairs of variables, because the positive correlations are not all close to one.
We now use the trimmed datasets and modified measures to further test the relative importance of indexicality and topicality. As shown in Figure 8AC,D U andD T are statistically larger than D U and D T , respectively, but they remain smaller than 1. This confirms that most words are clumped with respect to both users and threads. Overall, D U is smaller than D T , indicating that words are generally more concentrated with respect to users than to threads. This observation is rigorously confirmed by the fact thatD U is smaller thanD T to a comparable extent as D U is smaller than D T . Figure 8BD shows that also for individual words,D U andD T are typically larger than D U and D T , respectively. Furthermore, we can elucidate the effect of threads on users by considering the magnitude of the differenceD U − D U , and similarly, the effect of users on threads by considerinĝ D T − D T . These comparisons reveal that the effect of threads on users is statistically smaller than the effect of users on threads, both in the aggregate ( Figure 8AC) and for individual words ( Figure 8BD).
The most striking effect shown in Figure 8AC is the large number of words with small D U in comparison to D T . After trimming, over all windows, the comp.os.linux.misc (rec.music.hip-hop) dataset has 5, 356 (1, 808) words with D U < 0.4, versus 1, 657 (337) words with D T < 0.4. The list of words with D U < 0.4 but D T > 0.4 includes both very common words and highly topical words. In comp.os.linux.misc, example words include imagination, coffee, angst-ridden, and saukrates (a rapper); in rec.music.hip-hop, examples include regards, baptized and tauri (a Hungarian Warcraft server). It is interesting that such words are even more distinctive to individuals than to topics. A contributing factor to this clumpiness is the use of formulaic expressions. Such expressions, which are found in signature blocks, as well as in other conventionalized communications like greetings and insults, often have quite idiosyncratic lexical choices.
Altogether, we have strong evidence that the lexical make-up of the threads is strongly determined by the individual users. This speaks against the possibility that the topic dictates the vocabulary, and equally against the possibility that mutual imitation causes strong convergence in lexical choices as people interact in the discussion. This is a striking result. It contrasts with the major thrust of research on modeling the evolution of lexical systems, which is to explain convergence in the community [41,42]. This suggests that individuals may be more autonomous in their choices of words than in a wide range of other behaviors, from yawning and gait [43] to complex conscious decisions like the decision to purchase a product or to vote [44]. Given that individuals use different words to talk about the same topic, that word concentration over users is more extreme than over threads, and that D U is the strongest predictor of frequency change, the heterogeneity of people emerges as the single strongest factor in lexical diversity, both at any particular time and over time.

Discussion
We have introduced two new quantities, D U and D T , as measures of the dissemination of words across individuals and topics, and used them to characterize the vocabulary of two online discussion groups over a period of more than a decade. We found that almost all words are concentrated with respect to both individuals and topics, and that at shortterm (two-year) time scales, the word's concentration in the space of users and topics, as revealed by D U,T , is a strong determinant of word fate. D U and D T are separable components, and both trump word frequency. However, D U trumps D T . Word frequencies over time reflect a replicator dynamic, that is, a dynamic in which the words are reproduced by being copied through imitation [20,41,42,45]. Including both learning and use, this dynamic reflects an interaction of social and cognitive factors [46]. Word learning is facilitated by variety in the context of use [47], and rates of word use are in turn subject to great fluctuations over time, as a reflex of shifting user behavior and shifting topics. For a lexical innovation to survive in the language, it must avoid an absorbing boundary near f = 0, at which it is used so rarely that no one can learn it. Our investigation of the relationship between frequency change and dissemination change shows that a key to success beyond short-term fluctuations is increasing frequency (f ) hand-in-hand with increasing dissemination (D U,T ). The success of the P-words in our case study can be understood by considering that exogenous forcing by external events allowed them to overcome the handicap of low dissemination values. S-words, selected to exemplify more endogenous dynamics, behaved more like words in general by displaying higher dissemination values when rising.
Word frequency affects word fate at historical time scales when different forms compete to express the same meaning [14,15,34]. Why did frequency not prove to be important in the dynamics of the whole vocabulary, as studied here? The language system has strong functional pressures for words to be distinct from each other, in both form and meaning [6,41,42,45,48]. Although dictionaries use words to explain the meanings of other words, and thesauri group together words with related meanings, true synonymy is very rare [49,50]. For words which might seem to be synonyms, such as soda vs. pop, or yes vs. yup, there is normally a difference in dialect, formality, or other contextual factors governing the use of the word. Because almost every word is learned with a distinctive meaning (or set of meanings), and replication has low error rates, it follows that most words do not have a direct competitor for exactly the same meaning and contexts of use. If an active competition between two forms develops historically, then both can survive if they develop distinctive roles within the space of the lexical, syntactic, and pragmatic components of the linguistic system. For example, the English future auxiliary gonna is a new competitor for the older future will, but both survive because gonna is preferentially used in some constructions (such as questions), whereas will is preferentially used in others (such as the main clauses of conditionals) [51]. Reference [51] indeed uses the term niche to characterize these distinctive components in the usage of different future expressions, suggesting that differentiated niches are critical to their ongoing use in the language. These results complement those presented here by analyzing dimensions of the word niche that are internal to the linguistic system. The picture presents strong parallels to the exclusion principle in evolutionary biology, which states that occupying distinct niches protects species from competition [52]. Similar reasoning can also be applied to explore the competition between entire languages. In a model of language competition that assumes the speakers to be monolingual, distinct languages are similarly predicted to survive only if they are spoken by distinct, partially unmixed populations [53]. This prediction is attenuated if bilingualism in itself has high value or status as a human capability [54], permitting bilinguals to occupy a social position that is not available to monolinguals.
Diversity therefore depends on the diversity and viability of the individual niches. For biological species, the size of the geographical range and the species duration are correlated [18,19]. In studies of the lexicon, the individual words assume the role of species, and we have shown that the relative extent of the word niche is associated with the likelihood of a favorable or unfavorable fate. But we have also shown that the relative extent of the word niche does not provide the whole story about viability. In population biology, exogenous events such as asteroid impacts can overcome the general statistical trends associated with dissemination. The same thing is true here, where exogenous events such as inventions and wars can overcome general statistical trends associated with the dissemination of words. This generalization is further illustrated by the recent finding that censorship can induce large and distinctive deviations from typical frequency trajectories for the names of people [34].
We found that D U and D T are positively correlated, but still provide distinct information. A positive correlation is expected because individuals have characteristic interests. Further mechanisms contributing towards this correlation result from the participation of individuals in social and geographical structures. For example, these can cause clumping in product use, as shown by profiling the Internet for software products [55], which entails clumping of the words used to discuss those products. Structures in the social network can even contribute directly to product adoption, because the usefulness of many products (such as high-tech innovations) can depend on the number of neighbors who already use the product [23,56]. These same mechanisms pertain to other words, insofar as concepts and opinions resemble products.
We suggest, however, that other mechanisms limit the correlation between D U and D T , and explain the striking degree to which individuals were found to use different words in discussing the same topic. The variety in human social identities is thought to provide an impetus for innovation in modes of expression, as discussed in classic works of sociolinguistics [35,36,57]. Because people tend to associate with people like themselves, the variety in social identities can also give rise to clusters within social networks [58], and these clusters can in turn hinder lexical convergence [46,57,59]. The fundamental principles of discourse call for one to strike a balance between anchoring contributions in what the listener already knows, and providing novel and relevant information [60]. Online discourse can be viewed as a collective exploration of the conceptual world [61]. It follows from this study that the most engaging and fruitful discourse is discourse in which people cooperate in differentiating themselves and what they say.

Methods
Datasets. Usenet group archives are available at http://groups.google.com. The smallest unit of text is the post. Each post is attributed to a user and belongs to a thread (as defined by an initial post and all replies to it). We focus on two Usenet groups from their first post through 2008-03-31: (i) comp.os.linux.misc, which concerns Linux operating systems, includes 128, 903 users and 140, 517 threads beginning 1993-08-12; (ii) rec.music.hip-hop, which is devoted to hip-hop music, has 37, 779 users and 94, 074 threads beginning 1995-02-08. The activity of users in Usenet groups is bursty [32] and heterogeneous [13]. In the comp.os.linux.misc group, for example, the average user contributes 5.4 posts and remains active for 249.3 days, but the most persistent users have more than 1, 000 posts over more than 10 years. The average thread has 4.9 posts and is active for 4.5 days, but the longest threads have more than 1, 000 posts over 3 years. See Supporting Information S1, Text S1 for information about preprocessing of the text, and Figures S1 and S3 for information about the fat-tailed distributions that characterize these groups. Baseline model. The expected number of usersŨ in Eq. (1) is calculated by assuming that all words are randomly shuffled, while holding constant the number of users and the number of words per user. Let N w be the number of occurrences of the word w, m i be the total number of words contributed by user i, and N A ≡ i m i = w N w . The probability that the j + 1 th occurrence of w does not belong to user i is given by (1 − m i /(N A − j)). The probabilityŨ i that user i used word w at least once is calculated as the complement of the probability of not using it: where the approximation is valid for m i /N A 1 and f w ≡ N w /N A 1. This corresponds to a Poissonian baseline model with a fixed probability of using w given by the observed word frequency f w . The error in the approximation is smaller than 0.1% for the datasets we consider. This approximation was used in all calculations involving the untrimmed datasets, while the exact relation was used for the trimmed datasets. An analogous procedure is used for the calculation of the expected number of threadsT .   . Dissemination across users D U as a predictor of falling below threshold and of frequency decay. The analysis is performed over half-year window pairs t 1 and t 2 separated by two years for the comp.os.linux.misc and rec.music.hip-hop groups. A,D, Fraction of words with N w > 5 in t 1 that fall to N w ≤ 5 in t 2 . Histogram in gray: results from selected window pairs centered on t 1 = 1998-01-01 and t 2 = 2000-01-01. Red line: average over different non-overlapping window pairs with t 1 ranging from the (rounded off) beginning of the group through 2006-01-01, and t 2 = t 1 + 2 years. The probability of falling below threshold goes down as D U increases. B,E, Scatter plots of all words with N w > 5 in both windows (12,883 words for comp.os.linux.misc, 12,237 words for rec.music.hip-hop). Values on y-axis: log-frequency change ∆ log 10 f ≡ log 10 f (t 2 ) − log 10 f (t 1 ). Red lines: running median, 10th percentile, and 90th percentile. Words with rising frequency appear above and words with falling frequency appear below ∆ log 10 f = 0. Examples of words with large frequency changes are highlighted. The probability of frequency decay is greater for words with low D U . C,F, Summary of the dominant pattern in panels B,E over all non-overlapping windows with t 1 ranging from the beginning of the group to 2006, and t 2 = t 1 + 2. Median values of ∆ log 10 f at D U = 0.4 and D U = 1 are shown for each pair of windows.   Figure 3. The dashed green lines in panels B,E indicate the minimum possible ∆ log 10 f for a given log 10 f (t 1 ), due to the threshold N w > 5 imposed at t 2 . The analysis in Table 1 includes only the range log 10 f min < log 10 f < log 10 f max , where f min and f max are the limits of the range considered. The range is truncated at log 10 f max = −2.52 because, for words above this frequency, N w is so large compared to the number of users or threads that D is not informative. The range is truncated at log 10 f min = −4.61 for comp.os.linux.misc (log 10 f min = −4.52 for rec.music.hip-hop) because below these cutoffs the exclusion of words falling under the threshold (i.e., N w ≤ 5) introduces artifacts in the relationship to ∆ log 10 f (c.f. the relationship of the dashed green lines to the 10th percentile line). Specifically, f min was chosen for each dataset so that the percentage of words falling below the threshold at t 2 would be less than 5% of the words with log 10 f min < log 10 f < log 10 f max .    Table 1. Relative importance of dissemination across users, dissemination across threads, and frequency in word dynamics.

Group
D U D T log 10 f comp.os.linux.misc 9.9% 3.5% 0.2% rec.music.hip-hop 22.0% 5.0% 0.4% Relative importance of the three factors as predictors of frequency change (∆ log 10 f ), calculated using the method of Ref. [62]. Importance is based on the fraction of the variance of ∆ log 10 f explained by each factor. This method conservatively estimates the relative importance of the independent variables in a multiple regression setting. The data are combined over all window pairs t 1 , t 2 = t 1 + 2 considered in Figure 3. To avoid artifactual correlations for small and large f , the range of words is restricted in f , as indicated in the caption of Figure 5.  To obtain the correlations, first we calculateD U , D U ,D T , D T for all words with N w > 5 in the half-year windows of the trimmed datasets. The Pearson correlation coefficient, for each pair of variables, is then calculated over all words. The values reported in the table correspond to the averages ± standard deviations calculated over all non-overlapping half-year windows.