SEGMENTING TWITTER HASHTAGS

International Journal on Natural Language Computing (IJNLC) Vol. 5, No.4, August 2016
DOI: 10.5121/ijnlc.2016.5402 23
SEGMENTING TWITTER HASHTAGS
Jack Reuter 1
, Jhonata Pereira-Martins1
and Jugal Kalita2
1
Department of Computer Science, Wesleyan College Middletown CT 06459
2
Department of Computer Science, University of Colorado Colorado Springs CO 80918
ABSTRACT
Social Media Posts On Platforms Such As Twitter Or Instagram Use Hashtags,Which Are Author-Created
Labels Representing Topics Or Themes, Toassist In Categorization Of Posts And Searches For Posts Of
Interest. The Structural Analysis Of Hashtags Is Necessary As Precursor To Understandingtheir Meanings.
This Paper Describes Our Work On Segmenting Nondelimited Strings Of Hashtag-Type English Text. We
Adapt And Extend Methods Used Mostly In Non-English Context For Word Segmentation, To Handle
Hashtags In English, Obtaining Effective Results.
KEYWORDS
Hashtags, segmentation, Twitter
1. INTRODUCTION
Social media sites such as Twitter use hashtags, which are key phrases preceded by the pound (#)
symbol to mark posts with similar themes. A hashtag can be placed anywhere within a tweet, and
it is usually used to highlight the tweet’s keywords or main subject matter. Popular hashtags are
presented on Twitter as trending topics. The discussion of celebrities, TV shows and movies,
natural disasters like fires and earthquakes, and political hot topics all use hashtags so that, when
searching a site like Twitter, posts on the topic of interest can be easily found. From these
hashtags, Twitter also compiles lists of the most noteworthy topics of the day, month, year.
On Twitter, a post may contain none, one, or as many hashtags as the user may desire (or may fit
within the 140 character limit). Research has shown that tweets with hashtags receive two times
the engagement than those without, and that tweets with one or two hashtags have 21% higher
engagement than those with three or more1. In addition to Twitter, hashtags are used copiously in
other social media sites such as Instagram, Pinterest, Google+, Tumblr and Orkut. Starting in
2013, the social media giant, Facebook, has supported Twitter-like hashtags2. Consulting about
how to use hashtags properly as a marketing tool, and analysis of hashtags to track and measure
hashtag engagement and influence have spawned many businesses, including websites such as
hashtags.org.
A hashtag is simply a key phrase, composed of one or more tokens, written without whitespace.
Although the tokens are frequently words of a language like English, they do not have to be. One
can use abbreviations, wrongly spelled words, made-up words containing arbitrary character
sequences, numeric digits, names of people, places or events; in short, anything possible on an
everyday keyboard. One needs to be mindful about keeping a hashtag short, for both readability
and for actual length limitations imposed by previously mentioned sites. In Table 1 we see some
“English” hashtags that are composed of English or English-like tokens. Of course, it is possible
to mix languages, such as English and Spanish, or English and German in the same hashtag.

24
Table 1: A few hashtags, showing examples of ambiguous tokenizations, along with preferred
tokenizations.
Word segmentation is an important first step in natural language processing. It is difficult to
derive meaning from a piece of text without first having a good understanding of the words that it
comprises. We believe that hashtag tokenization should be the first step in automatic
understanding, clustering or classification not only of individual hashtags, but of social media
posts containing them. In this paper, we discuss methods for, experiments in, and results of
automatic hashtag segmentation.
2. PROBLEM DEFINITION
Formally, the problem is as follows: Given a string a1a2:::ak, where each ai is a meaningful
substring (i.e., a word, name, abbreviation, etc.), determine where the boundaries lie between
each ai and ai+1. For example, given a string such as “randompicoftheday” return “random pic of
the day”. Initially it may seem like a simple problem. Simply loop through all substrings of the
input, looking for matches in a dictionary. Once all matches have been found, segment the string
accordingly.
The real problem, however, lies in the phrase “then segment accordingly.” “The key to accurate
automatic word identification...lies in the successful resolution of these ambiguities and a proper
way to handle out of- vocabulary words” [1]. Although referring to the segmentation of Chinese
characters, the sentiment is still very much appropriate. In our case, ambiguities occur when a
string has more than one meaningful segmentation, or has out-of-vocabulary words when our
dictionary fails us. Both scenarios occur frequently, and the success of our methods depends on
the handling of such situations.
Consider, for example, the string “brainstorm”. Using a table lookup, a machine could read this as
either “bra in storm”, “brain storm”, or the correct, untouched “brainstorm”. Following Webster
and Kit [2], we will refer to this as conjunctive ambiguity, i.e., when a meaningful string can be
broken up into n meaningful substrings. The natural solution of course is to take the segmentation
with the largest matched word. This maximum matching approach handles conjuctive ambiguity
very well, for it is unlikely that a syntactically sound clause happens to merge into a larger word,
yet it is quite common for a larger word to break up into dictionary- match able pieces.
Maximum matching fails, however, in cases of disjunctive ambiguity—the situation when a
string “ABC” can be broken up meaningfully as either “AB C” or “A BC”. “doubledown”, for
example, could be interpreted as either “doubled own” or the correct “double down” (or even “do
u bled own” which exhibits both conjunctive and disjunctive ambiguity). Taking the maximum
matching here would result in the incorrect segmentation “doubled own”. Disjunctive ambiguity,
it appears, requires a bit more syntactic knowledge to resolve.

25
The other main issue is the handling of strings outside of out dictionary. The string “votedems”,
for instance, should be returned as “vote dems” and not “voted ems”, while our dictionary may
only contain the abbreviation “ems” and not “dems”. With typos, abbreviations, online slang, and
just a general abundance of linguistic rule-breaking in hashtags, these situations are bound to
occur, and occur frequently. For our effort to succeed, such unknowns must be recognized and
handled appropriately.
3. RELATED WORK
Twitter posts are voluminous in number, publicly available and easily accessed, and as a result,
they have become a prime source of “modern” Internet-mediated language for studies of all
stripes, with hundreds of papers published, including a recent survey [3]. For example, studies
have attempted to normalize the informality present in such language [4, 5, 6], tag parts-of-speech
[7, 8], summarize a collection of posts [9, 10, 11], identify topics [12, 13] and detect expressed
sentiments [14, 15]. To the best of our knowledge, however, most of these studies simply ignore
hashtags, which are crucial for indexing and categorizing tweets.
In languages such as German and Chinese, due to frequent use of large compounds and lack of
delimitation, respectively, effective segmentation is important. Nie et al. [16] segment Chinese
text heuristically, analyze unknown strings statistically for likelihood of being real words. Xue et
al. [1] segment Chinese words by tagging them with a MaxEnt model trained using features
involving neighboring characters. Peng et al. [17] use Conditional Random Fields (CRFs) to tag
characters as either START or NONSTART, using POS tags and neighboring characters as
features, as well as an N-best system to process and accept probable unknown words. Koehn and
Knight [18] learn splitting rules for German compounds and choose most likely segmentations
based on how frequently the segmented tokens appear in a corpus, as well as whether or not the
translation of a segmentation contains the same words as the translation of the original string.
Part-of-speech (POS) tags are also used to help avoid splitting off suffixes and prefixes from root
lemmas.
There is not much prior work in hashtag segmentation. Berardi et al. [19], who took part in TREC
Microblog Track 2011 [20] for searching a set of 16 million tweets for relevant tweets at a certain
time, discuss approaching hashtag segmentation using a Viterbi-type algorithm [21] to improve
search performance, although they do not provide any results. Srinivasan et al. [22] discuss
several approaches primarily devoted to segmentation of web domain names, although they
discuss the segmentation of hashtags briefly. Their contention is that domain names and hashtags
share certain similar properties in how they are constructed and therefore, similar approaches can
be used. They start with several corpora, and develop scoring formulas for segmentation using
unigrams and bigrams found in the corpora and segmentation length. Models obtained are then
weighed using a maximum margin learning problem solved using the SVM- struct framework
[23]. They do not provide any direct results for hashtag segmentation, but do show that it can
improve recall in Twitter searches. Bansal et al. [24] show that hashtag segmentation can improve
linking of entities extracted from tweets. They assume that an unknown or OOV segment inside a
hashtag is surrounded by known segments on its two sides. If there are ambiguities, they
hypothesize various lengths for the intervening segment, scoring the alternatives using unigram
based features and probable lengths of the unknown segment. Probable lengths are determined
assuming its length follows the distribution of lengths of known tokens found in tweets.
Segmentation of a hashtag involves several steps, as we will soon see. The current wisdom is that
when there is a pipeline involving several steps of processing in an NLP task, the results can
often be improved by dispensing with pipelining and performing the steps jointly. For example,

26
Zhang et al. [25] improve performance of a greedy algorithm that jointly predicts segmentation,
POS tags and dependency parses, improving all of them iteratively, using hill-climbing. The
contention is that in a pipeline architecture, the errors cascade from one step to the next, but joint
prediction reduces error propagation. Joint prediction [26, 27, 28] has performed better compared
to pipeline architectures, especially for morphologically rich languages like Arabic and Chinese.
4. OUR APPROACH
In the rest of the paper, we describe the several approaches we take to hashtag segmentation and
present our results. We first present the methods we use to generate and assign scores to possible
segmentations for a given input. We then present the more general algorithm we use to select the
best answer from the generated set of possibilities. After implementing the algorithm’s steps in
pipeline, we also describe a hill-climbing version of the algorithm, because we want to
investigate if joint prediction performs better for the hashtag segmentation task, as claimed by
several recent papers in similar tasks. We finally present segmentation results produced by our
approaches, compare them, and identify the best performer.
5. SEGMENTATION SCORING METHODS
Each of the following methods defines a scoring function whose aim is to assign top scores to
correct segmentations, thus turning the process of segmentation into a search for the highest
score. We have developed these scoring functions based on our understanding of what ideal
segmentations should look like. The goal is to give high scores to good segmentations and low
scores to bad ones.
5.1 Maximum Known Matching (MKM)
To begin, we try a simple maximum matching approach, i.e., given a string s, get all possible
segmentations of s into dictionary words, then return the “longest” segmentation. The question
then becomes how to define the length of a segmentation. Should we prefer the segmentation
containing the longest words? That which has the largest average word length? We want to
consider both. We first want to consider just the segmentations with the largest average word
length, then take that which contains the longest word (if the longest words are equal then
compare the second longest, third, etc). In other words we need a function f that fulfills the
following two conditions:
1. size(s1) > size(s2) =) f(s1) < f(s2)
2. size(s1) = size(s2) ^ s1 > s2 =) f(s1) > f(s2)
where s1 and s2 are segmentations, size a function that returns the number of words in a
segmentation, and s1 > s2 meaning s1 contains longer words than s2. Note that average word
length is inversely proportional to the number of words in a segmentation. We could of course
write out the logic above, i.e. define a function that takes the segmentation of shortest length
containing the largest individual words, but it may be useful later on to be able to assign a
numeric score which models the same choice. Thus, we define the length score of a segmentation
as follows:

27
where len(w) returns the length of a word w, and s is a segmentation into i words.
5.2 Maximum Unknown Matching (MUM)
The problem with the previous aptly named approach, however, is that it accepts no unknown
words. To amend this, we expand our algorithm in the following way: Given a string s, rather
than looking at all segmentations into known words, consider all segmentations of s where each
division point borders at least one known word. Then return the segmentation with the highest
length score.
But now we must amend our definition of length score, for as it stands it will simply return s
itself (or, if we exclude s, a segmentation with a very large unknown word). The previous
definition of length score has no way of weighing known versus unknown words, and places high
value on the average word length of a segmentation. To adjust we redefine the score as follows:
where s is a segmentation into i known words and j unknown words.
5.3 Two grams (2GM)
These simple methods produce fairly effective results (see evaluation section), but, as discussed
earlier, successful disambiguation cannot rely merely on length—some syntactic data must be
incorporated. Using a database of 2-gram occurrences in a corpus, we define the 2-gram score of
a segmentation s, simply as the number of recognized 2-grams in s, divided by the total number
of 2-grams in s. A segmentation of length 1, thus containing no possible 2-grams, receives a score
of -1. The final scores are then normalized to fall between 0 and 1, and a score of -1 is set by
default to 0.5.
We ignore the actual occurence count of the recognized 2-grams in order to avoid over-
segmentation. The string “d at a mining”, for example, would return a much higher score than the
correct “data mining”, due to the frequency of the 2-gram “at a”, were occurence count taken into
consideration. Without it, the latter outscores the former.
To account for numbers and ordinals, which are not included in our 2-gram datasets, as well as
acronyms and contractions, we translate each unrecognized word into a set of possible words, via
either a translation dictionary—the contents of which are detailed below—or, in the case of
numbers and ordinals, a predefined ruleset. Out of these possibile translations, that with the
highest 2-gram score is assumed to be correct.

28
5.4 POS tagging
2GM is flawed, however, in the same way that MKM is flawed; it lacks strength in handling
unknown words. 2-grams which lie outside of the database return a count of 0.
As an attempt to correct this, we approach the problem as a POS tagging problem. I.e. given a
segmentation, tag each word with the appropriate POS, then assign it a score based on the
probability of a given sequence of POS tags.
This approach poses two new problems: a) Tag appropriately, and b) Score tag sequences. And
for each we pose two solutions, one using the ARK POS tagger for Twitter [8], and one using
Hidden Markov Models (HMM) implemented with the MALLET toolkit [29]. Using ARK, we
can tag segmenations by POS, then, using their POS n-gram data, repeat 2GM using POS tags
rather than words themselves. Similarly, with a trained HMM, we can tag segmentations by POS,
then score a tag sequence by taking its average edge weight in the model. This leads us to four
new possible strategies:
1. Tag with ARK, assign probabilities with ARK (AA),
2. Tag with ARK, assign probabilities with HMM (AH),
3. Tag with HMM, assign probabilities with ARK (HA), and
4. Tag with HMM, assign probabilities with HMM (HH).
6. ALGORITHMS
We now have seven scoring methods for disambiguation—MKM, MUM, 2GM, plus the four
listed above. When used in overall segmentation algorithms, these scores are normalized on each
new input to fall between 0 and 1, based on the highest scoring segmentation for the current
input.
6.1 Pipeline
These scores, plus some simple heuristics, leave us with the pipeline-based segmentation
algorithm as detailed in Algorithm 1.
In the pipeline algorithm, highest Scoring Seg () searches by brute force using a scoring function
that is some convex combination of previously mentioned scoring functions (MUM, 2GM, AH,
etc.); prune() heuristically filters out unlikely segmentations; all Possible Segmentations ()
returns every possible segmentation of s—unless s exceeds a length threshold, in which case at
least one of the k-longest words in s is made to be present in every segmentation (to limit search

29
space size); and segment By Delimeters () segments based on punctuation and capitalization,
returning, for example “Club Rio - June 19 th - 8 - 12 am” when given “ClubRio-June19th-8-
12am”. Note that this will fail on confusing inputs. “LadiesoftheDMV”, for example, looks
delimited but is in fact only partially, and would thus be under-segmented. Only those rule-based
segmentations which meet a certain length score threshold, or are composed entirely of known
words, therefore, are returned. The rest have their unknown sections fed back into the segmenter
to be treated by the normal algorithm.
6.2 Hill-climbing
Based on the work of Zhang et al. [25], we also consider a greedy hill-climbing algorithm. I.e.,
rather than pruning down to a set of probable segmentations and then searching by brute force,
start with a random segmentation, calculate the scores of all “nearby” segmentations, then climb
to the one with the highest score and recurse. The algorithm terminates once no further upward
steps are possible. k random restarts are allowed to minimize the chance of getting stuck in local
maxima. In our implementation, segmentations of a string of length n string are considered as
binary strings of length n -1, where a “1” indicates that the corresponding character in the original
string is followed by a splitting point. “Nearby” segmentations are then simply defined as the set
of binary strings obtained by flipping a bit in the original, i.e. either adding or removing a
splitting point.
In its initial implementations, hill-climbing, although more elegant than the pipeline approach,
proved to be slightly less successful (at best yielding an f-score of 76.5%, whereas, at the time,
the pipeline approach peaked at 82.2%). Likely, the definition of nearby segmentations was too
narrow, creating too many local maxima to avoid.
7. UNKNOWN HANDLING
Nie et al. [16] define the following process for unknown handling:
1. Perform a maximum matching.
2. Gather remaining unknowns.
3. Remove unlikely candidates based on a predefined ruleset.
4. Add those which occur most frequently to the dictionary and repeat.
This method enables the segmenter to improve with repeated applications, and it is, generally, the
method we use to accomodate unknown words. Rather than defining a ruleset for how words
should look, however, we train a Markov chain on a list of English words. The states of the chain
correspond to letters of the
Table 2: POS tags used

30
alphabet, and transitions between states model letter sequence probabilities. The average
transition probability of a given string should, then, theoretically mirror its likelihood of being a
real word. The removal of unlikely candidates then comes down to choosing a threshold value.
As with Nie’s method, the frequency of a given unknown is also factored into its estimated
probability of being a real word. Because of this, the size of the test set will directly affect the
appropriate threshold value.
In addition to this algorithm, unknown handling has also been attempted via the use of spell-
checking resources. Spelling errors, intentional or unintentional, are the root cause of many
unknowns, and handling them effectively would have significant effects on performance. Several
strategies were tested—treating unknown words within some edit distance threshold of known
words as known words themselves; treating such words as “semi-known” words, and adjusting
the length score function to handle them; including spell-check suggestions as translations for
2GM scoring—but each led to disjunctive ambiguity errors; words would be segmented with
extra letters when they shouldn’t have. Rules were devised as an attempt to exclude such
occurrences, but still without improving results. The one method which did prove to be useful
was the inclusion of common misspellings and their root words in our translation dictionary. As
with numbers, ordinals, acronyms, and contractions, common spelling corrections have thus been
included in translations for 2GM scoring.
8. RESOURCES USED
9. EVALUATION AND RESULTS
Performance is rated in terms of precision, recall, and f-score. For a single segmentation,
precision is defined as the number of correctly segmented words divided by the total number of
words in the proposed
Table 3: Performance evaluation in terms of accuracy

31
Table 4: Performance evaluation in terms of speed
Correct segmentations are based on a manually segmented list of 1129 hashtags. In the list, each
hashtag corresponds to a set of all valid segmentations (typically just one, but in some cases
alternate segmentations exist which are equally acceptable). During calculation of evaluation
metrics on a proposed segmentation, scores are calculated for each possible answer and the
highest results are returned.
Tables 3 and 4 list the results of methods that have been tested with the current dictionaries and
heuristics, as tested on our manually curated answer set. AH outscored the other three POS-based
methods. This makes sense, as ARK’s successful POS tagger should easily trump our HMMs in
terms of tagging accuracy, whereas the edge weights of the HMM should provide comparable or
better transition probabilities. As such, AH was the first of the four to be tested with the updated
heuristics and newly introduced translation scheme. AH was followed by HH for the latter’s far
superior speed.
For efficiency, possible segmentations had to be pruned twice before actually taking AH score
into consideration. First, heuristically. Second, by a combination of length and 2GM score. This
second pruning was based on an optimal convex combination of length and 2GM scores, the
values of which are depicted in Figure 1. The remaining segmentations were then disambiguated
by a convex combination of all three scores, length, 2-gram, and AH. Figure 2 displays the
relationship between possible combinations of the three and their effects on f-score. Although the
success of the AH method is largely due to the success of the pipeline system, Figure 2 shows
that it can perform as well as the 2GM method in disambiguation.
The same pruning system was used for the HH method, though with a slightly larger pruning set
size. Figure 3 mirrors the form of Figure 2, with HH score substitued for AH score. Though faster
than AH, HH has been less successful in disambiguation. In Figure 1, L refers to the weight
assigned to the length score, and T, the weight assigned to the 2GM score, can be simply
calculated as 1-L. Similarly, in Figures 2 and 3, L represents the weight of the length score, and T
is left to be calculated as 1-(L+A) or 1-(L+H), respectively, where A represents the weight of the
AH score and H the HH score. Performance evaluation in Figures 2 and 3 is based on the
segmentation of 232 hashtags chosen such that they could not be handled by simple heuristics.
They were chosen by running a heuristic segmenter on the manually curated collection and

32
gathering those on which the machine failed. Performance in Figure 1 is evaluated based on the
full answer set.
Figure 1: A range of combinations of length score and 2GM score with resulting performance measures
Figure 2: A range of combinations of length score, 2GM score, and AH score with resulting performance
measures

33
9.1 Comparison with Prior Approaches
Srinivasan et al. [22] do not provide direct results for hashtag segmentation, but state that recall in
search for tweets can improve by 5% or more when they segment the hashtags and use the tokens
obtained in search. They developed various models for segmentation of domain names, and
obtained average accuracies of 87.07, 87.98, 88.68 and 89.74% for the four bigram-based
methods (these numbers were obtained by adding up the values in the bottom part of Table 3 in
their paper and averaging). Our methods perform better compared to their average results, though
their numbers are for domain splitting and ours for hashtag splitting. The work that comes closest
to ours is by Bansal et al. [24], who segment Twitter hashtags with the goal of improving
extraction of links or relations among entities mentioned in tweets. They provide
Table 5: Performance evaluation in terms of accuracy for Brazilian Portuguese
a P@1 or accuracy value (which is the precision value we report) as 0.914. This is quite
comparable to the results we provide, but they do not provide any recall and F-score values.
Hence, it is difficult to fully compare their method to ours.
10. EXPERIMENTS WITH BRAZILIAN PORTUGUESE
We wanted to experiment with hashatags in at least one more language and we chose it to be
Brazilian Portuguese. Although often entirely English, tweets in Brazilian Portuguese range from
being entirely English to entirely Portuguese, sometimes combining both. Of course, working
with both languages brings in more difficulties, mainly more ambiguities.
Since these two languages are not very dissimilar, we hypothesized that many ambiguities could
be handled by methods previously described for English. We kept the original English dictionary
with approximately 150,000 words, but added to it a dictionary of approximately 100,000
Portuguese words obtained from free online resources. We added abbreviations, corporations,
first and last names to the Portuguese dictionary, just as we did for English. We also added the
50,000 most common 2-grams in Portuguese taken from Corpus do Portuguˆes, and downloaded
~100,000 hashtags in Brazilian Portuguese from Twitter using the Twitter API.
Although formal Brazilian Portuguese uses accents liberally, accents are often omitted in tweets
due to typing ease on mobile devices. To simplify matters, we removed all accents from our
datasets. For experiments reported here, we added Portuguese 2-gram data to our English 2-gram
data, although we acknowledge that this misses the possibility of mixed English and Portuguese
2-grams.
Correct segmentations are based on randomly chosen 1000 manually segmented hashtags. Results
from Brazilian Portuguese with MKM, MUM and 2GM methods are given in Table 5. Additional
experiments were not performed for Portuguese. We believe that the combination of two
languages and lack of as many resources as in English has for ready use, caused the results in
Brazilian Portuguese to be poorer.

34
11. POSSIBLE IMPROVEMENTS AND CONCLUSION
With the right tagset and enough supervised learning data, there is still hope for the success of the
HH method. Alternatively, CRFs, as shown in [17], have been used successfully as segmentation
tools, and moving from HMMs to CRFs, thus allowing arbitrary feature inclusion, could be a
more effective option. Larger n-gram data has not yet been tried to extend 2GM, and neither have
alternative lexical normalization strategies. With different definitions of distance, the hill-
climbing algorithm could also prove superior to the pipeline approach.
Much work is also left to be done in the task of unknown handling. Different training sets could
be explored, as well as sophisticated methods for learning threshold values. As far as spelling
correction goes, our solution is far from perfect. A distance measure that balances spelling error
correction against the creation of erroneous disjunctive ambiguities would likely improve
performance greatly. Physical keyboard proximity may even be useful to consider.
Hashtags typify a significant chunk of conversational language online. They have spread beyond
Twitter and into most popular social media sites. Some Twitter posts contain two, three or more
hashtags. Instagram posts can contain ten or more hashtags. Thus, ignoring hashtags in analyzing
texts in social media posts can make for impoverished understanding. One extension could be to
create a graph of relationships among
Figure 3: A range of combinations of length score, 2GM score, and HH score with resulting performance
measures
hashtags, allowing machines to first process a hashtag, then not only explore related topics within
that tag, but other topics within related tags.

35
REFERENCES
[1] N. Xue, “Chinese word segmentation as character tagging,” Computational Linguistics and Chinese
Language Processing, vol. 8, no. 1, pp. 29–48, 2003.
[2] J. J. Webster and C. Kit, “Tokenization as the initial phase in Natural Language Processing,” in
COLING–Volume 4, 1992, pp. 1106–1110.
[3] A. Farzindar and D. Inkpen, “Natural language processing for social media,” Synthesis Lectures on
Human Language Technologies, vol. 8, no. 2, pp. 1–166, 2015.
[4] M. Kaufmann and J. Kalita, “Syntactic normalization of twitter messages,” in International
Conference on Natural Language Processing, ICON, Kharagpur, India, 2010.
[5] B. Han and T. Baldwin, “Lexical normalisation of short text messages: Makn sens a# twitter,” in
ACL-HLT, Volume 1, 2011, pp. 368–378.
[6] B. Han, P. Cook, and T. Baldwin, “Automatically constructing a normalisation dictionary for
microblogs,” in EMNLP and CoNLL, 2012, pp. 421–432.
[7] K. Gimpel, N. Schneider, B. O’Connor, D. Das, D. Mills, J. Eisenstein, M. Heilman, D. Yogatama, J.
Flanigan, and N. A. Smith, “Part-of-speech tagging for twitter: Annotation, features, and
experiments,”in ACL-HLT, Volume 2, 2011, pp. 42–47.
[8] O. Owoputi, B. O’Connor, C. Dyer, K. Gimpel, N. Schneider, and N. A. Smith, “Improved part-
ofspeech tagging for online conversational text with word clusters,” in NAACL-HLT, 2013, pp. 380–
390.
[9] B. Sharifi, M.-A. Hutton, and J. Kalita, “Summarizing microblogs automatically,” in NAACL, 2010,
pp. 685–688.
[10] D. Inouye and J. K. Kalita, “Comparing twitter summarization algorithms for multiple post
summaries,” in Third IEEE International Conference on Social Computing, SocialCom. IEEE, 2011,
pp.298–306.
[11] J. Nichols, J. Mahmud, and C. Drews, “Summarizing sporting events using twitter,” in ACM
International Conference on Intelligent User Interfaces, 2012, pp. 189–198.
[12] M. Cataldi, L. Di Caro, and C. Schifanella, “Emerging topic detection on twitter based on temporal
and social terms evaluation,” in Proceedings of the Tenth International Workshop on Multimedia
Data Mining, 2010, p. 4.
[13] W. X. Zhao, J. Jiang, J. He, Y. Song, P. Achananuparp, E.-P. Lim, and X. Li, “Topical keyphrase
extraction from twitter,” in NAACL-HLT–Volume 1, 2011, pp. 379–388.
[14] A. Go, R. Bhayani, and L. Huang, “Twitter sentiment classification using distant supervision,”
CS224N Project Report, Stanford, vol. 1, p. 12, 2009.
[15] L. Barbosa and J. Feng, “Robust sentiment detection on twitter from biased and noisy data,” in
COLING: Posters. Association for Computational Linguistics, 2010, pp. 36–44.
[16] J.-Y. Nie, M.-L. Hannan, and W. Jin, “Unknown word detection and segmentation of Chinese using
statistical and heuristic knowledge,” Communications of COLIPS, vol. 5, no. 1, pp. 47–57, 1995.
[17] F. Peng, F. Feng, and A. McCallum, “Chinese segmentation and new word detection using
conditional random fields,” in COLING, 2004, p. 562.
[18] P. Koehn and K. Knight, “Empirical methods for compound splitting,” in European ACL–Volume 1,
2003, pp. 187–193.
[19] G. Berardi, A. Esuli, D. Marcheggiani, and F. Sebastiani, “ISTI@ TREC Microblog Track 2011:
Exploring the Use of Hashtag Segmentation and Text Quality Ranking,” in TREC, 2011.
[20] I. Ounis, C. Macdonald, J. Lin, and I. Soboroff, “Overview of the TREC-2011 Microblog Track,” in
Proceedings of the 20th Text REtrieval Conference (TREC 2011), 2011.
[21] G. D. Forney Jr, “The Viterbi algorithm,” Proceedings of the IEEE, vol. 61, no. 3, pp. 268–278, 1973.
[22] S. Srinivasan, S. Bhattacharya, and R. Chakraborty, “Segmenting web-domains and hashtags using
length specific models,” in Proceedings of the 21st ACM International Conference on Information
and Knowledge Management, 2012, pp. 1113–1122.
[23] I. Tsochantaridis, T. Hofmann, T. Joachims, and Y. Altun, “Support vector machine learning for
interdependent and structured output spaces,” in Proceedings of the Twenty-first International
Conference on Machine Learning. ACM, 2004, p. 104.
[24] P. Bansal, R. Bansal, and V. Varma, “Towards deep semantic analysis of hashtags,” in Advances in
Information Retrieval. Springer, 2015, pp. 453–464.

36
[25] Y. Zhang, C. Li, R. Barzilay, and K. Darwish, “Randomized Greedy Inference for Joint
Segmentation, POS Tagging and Dependency Parsing,” in NAACL-HLT, 2015, pp. 42–52.
[26] N. Habash and O. Rambow, “Arabic tokenization, part-of-speech tagging and morphological
disambiguation in one fell swoop,” in Proceedings of the 43rd Annual Meeting on Association for
Computational Linguistics, 2005, pp. 573–580.
[27] `J. Hatori, T. Matsuzaki, Y. Miyao, and J. Tsujii, “Incremental joint approach to word segmentation,
pos tagging, and dependency parsing in chinese,” in Proceedings of the 50th Annual Meeting of the
Association for Computational Linguistics: Long Papers-Volume 1, 2012, pp. 1045–1053.
[28] B. Bohnet, J. Nivre, I. Boguslavsky, R. Farkas, F. Ginter, and J. Hajiˇc, “Joint morphological and
syntactic analysis for richly inflected languages,” Transactions of the Association for Computational
Linguistics, vol. 1, pp. 415–428, 2013.

SEGMENTING TWITTER HASHTAGS

More Related Content

What's hot (17)

Similar to SEGMENTING TWITTER HASHTAGS (20)

Recently uploaded (20)

SEGMENTING TWITTER HASHTAGS