=Paper=
{{Paper
|id=Vol-1395/paper_20
|storemode=property
|title=AMRITA — CEN@NEEL: Identification and Linking of Twitter Entities
|pdfUrl=https://ceur-ws.org/Vol-1395/paper_20.pdf
|volume=Vol-1395
|dblpUrl=https://dblp.org/rec/conf/msm/BNMRP15
}}
==AMRITA — CEN@NEEL: Identification and Linking of Twitter Entities==
https://ceur-ws.org/Vol-1395/paper_20.pdf
AMRITA - CEN@NEEL : Identification and Linking of
Twitter Entities
Barathi Ganesh H B, Abinaya N, Anand Kumar M, Vinayakumar R, Soman K P
Centre for Excellence in Computational Engineering and Networking
Amrita Vishwa Vidyapeetham, Coimbatore, India
barathiganesh.hb@gmail.com, abi9106@gmail.com, m anandkumar@cb.amrita.edu
¯
vinayakumarr77@gmail.com, kp soman@amrita.edu
¯
ABSTRACT bols (:-), #, @user), abbreviations, short words (lol, omg),
A short text gets updated every now and then. With the misspelled words, repeated punctuations and unstructured
global upswing of such micro posts, the need to retrieve words (goooood nightttt, helloooo). Hence these micro posts
information from them also seems to be incumbent. This were fed to the dedicated twitter tokenizer which accounts
work focuses on the knowledge extraction from the micro language identification, a lookup dictionary for list of names,
posts by having entity as evidence. Here the extracted en- spelling correction and special symbols [4][5] for e↵ective to-
tities are then linked to their relevant DBpedia source by kenization.
featurization, Part Of Speech (POS) tagging, Named En- 2.2 POS Tagger
tity Recognition (NER) and Word Sense Disambiguation
(WSD). This short paper encompasses its contribution to Due to the conversional nature of micro blogs with non-
#Micropost2015 - NEEL task by experimenting existing syntactic structure it becomes difficult in utilizing general
Machine Learning (ML) algorithms. algorithms with traditional POS tags in Penn Treebank and
Wall Street Journal Corpus [6]. O’Conner et al. used 25
POS tagset which includes dedicated tags (@user, hash tag,
Keywords G, URL, etc.) for twitter and reports 90% accuracy on
CRF, Micro posts, NER POS tagging [7]. The ability of resolving independent as-
sumptions and overcoming biasing problems make CRF as
1. INTRODUCTION promised supervised algorithm for sequence labeling appli-
cations [8]. TwitIE tagger: which utilizes CRF to build the
Micro posts are a pool of knowledge with scope in busi-
POS tagging model was thus used.
ness analytics, public consensus, opinion mining, sentimen-
tal analysis and author profiling and thus indispensable for 2.3 Named Entity Recognizer
Natural Language Processing (NLP) researchers. People use
CRF and SVM produced promising outcome for sequence
short forms and special symbols for easily conveying their
labeling task which prompted us to use the same for our ex-
message due to the limited size of micro posts which has
periment. Long range dependency of the CRF can also solve
eventually built complexity for traditional NLP tools [3].
Word Sense Disambiguation (WSD) problem over other graph-
Though there are number of tools, most of them rely on
ical models by avoiding label and casual biasing during learn-
least ML algorithms which are e↵ective for long texts than
ing phase. Both CRF and SVM allow us to utilize the com-
short texts. Thus by providing su↵cient features to these
plicated feature without modeling any dependency between
algorithms the objective can be achieved. We experimented
them. SVM is also well suited for sequence labeling task
the NEEL task with the available NLP tools to evaluate
since learning can be enhanced by incorporating cost models
their e↵ect on entity recognition by providing special fea-
[9]. These advantages provide flexibility in building expres-
tures available in tweets.
sive models with CRF suite and MALLET tools [10][11].
2. SELECTION OF ALGORITHMS
3. EXPERIMENTS AND OBSERVATION
2.1 Tokenization The experiment is conducted on i7 processor with 8GB
Tokenizing becomes highly challenging in micro posts due RAM and the flow of experiment is shown in Figure 1.
to the absence of lexical richness. It includes special sym- The training dataset consists of 3498 tweets with the unique
tweet id. These tweets have 4016 entities with 7 unique tags
namely Character, Event, Location, Organization, Person,
Product and Thing [1][2]. POS tag for the NER is obtained
from TwitIE tagger after tokenization which takes care of
Permission
Copyright to c make
2015 digital
held by or author(s)/owner(s);
hard copies of all or copying
part of this work for
permitted the nature of micro posts and provides an outcome desired
only for or
personal private and academic
classroom purposes.
use is granted without fee provided that copies are by the POS tagger model. The tags are mapped to BIO Tag-
Published as part of the #Microposts2015 Workshop proceedings,
not made or distributed for profit or commercial advantage and that copies
available online as CEUR Vol-1395 (http://ceur-ws.org/Vol-1395) ging of named entities. Considering the entity as a phrase,
bear this notice and the full citation on the first page. To copy otherwise, to token at the beginning of the phrase is tagged as ‘B-(original
republish, to post on May
#Microposts2015, servers or to2015,
18th, redistribute to lists,
Florence, requires prior specific
Italy.
permission and/or a fee. tag)’ and the token inside the phrase is tagged as ‘I-(original
Copyright 20XX ACM X-XXXXX-XX-X/XX/XX ...$15.00. tag)’. Feature vector constructed with POS tag and addi-
· #Microposts2015 · 5th Workshop on Making Sense of Microposts · @WWW2015
tional 34 features like root word, word shapes, prefix and
suffix of length 1 to 4, length of the token, start and end Table 1: Observations
Tools 10 Fold-Cross Development Time
of the sentence, binary features - whether the word contains
Validation Data (mins)
uppercase, lower case, special symbols, punctuations, first
Mallet 84.9 82.4 168.31
letter capitalization, combination of alphabet with digits,
SVM 79.8 76.3 20.15
punctuations and symbols, token of length 2 and 4 , etc.
After constructing the feature vector for individual tokens in CRFSuite 88.9 85.2 4.12
the training set and by keeping bi-directional window of size
5, the nearby token’s feature statistics are also observed to
help the WSD. The final windowed training sets are passed that has direct relation with NER accuracy. The utilized
to the CRF and SVM algorithms to produce the NER model. TwitIE tagger shows promising performance in both the to-
The development data has 500 tweets along with their id and kenization and POS tagging phases. The special 34 features
790 entities [1][2]. The development data is also tokenized, extracted from the tweets improves efficacy by nearing 13%
tagged and feature extracted as the training data for testing greater than the model with absence of special features. At
and tuning the model. The developed model performance linking part, this work is limited using dot product similarity
which could be improved by including semantic similarity.
5. REFERENCES
[1] Rizzo, Giuseppe and Cano Basave, Amparo Elizabeth
and Pereira, Bianca and Varga, Andrea, Making Sense
of Microposts (#Microposts2015) Named Entity
rEcognition and Linking (NEEL) Challenge., In 5th
Workshop on Making Sense of Microposts
(#Microposts2015), pp. 44–53, 2015.
[2] Matthew Rowe and Milan Stankovic and Aba-Sah
Dadzie, Proceedings, 5th Workshop on Making Sense of
Microposts (#Microposts2015): Big things come in small
packages, Florence, Italy, 18th of May 2015, 2015.
[3] Dlugolinsky S, Marek Ciglan and M Laclavik,
Figure 1: Overall Model Structure Evaluation of named entity recognition tools on
microposts, INES, 2013 , pp. 197-202. IEEE, 2013.
is evaluated by 10- fold cross validation of training set and [4] Bontcheva K, Derczynski L, Funk A, Greenwood M A,
validated against the development data. The accuracy is Maynard D, and Aswani N, TwitIE: An Open-Source
computed as ratio of total number of correctly identified en- Information Extraction Pipeline for Microblog Text, In
tities to the total number of entities and tabulated in Table RANLP, pp. 83-90, 2013, September.
1. [5] Brendan O’Connor, Michel Krieger and David Ahn,
P
correctly identif ied entities TweetMotif: Exploratory Search and Topic
Accuracy = ⇥ 100 (1)
total entities Summarization for Twitter, ICWSM, 2010.
MALLET incorporates O-LBFGS which is well suited for [6] Tim Finin, Will Murnane, Anand Karandikar, Nicholas
log-linear models but shows reduced performance when com- Keller and Justin Martineau, Annotating named entities
pared to CRFsuite which engulfs LBFGS for optimization in Twitter data with crowdsourcing, 2010.
[12][13]. SVM’s low performance can be improved by in- [7] Kevin Gimpel, et al, Part-of-Speech Tagging for
creasing the number of features which will not introduce Twitter: Annotation, Features, and Experiments,
any over fitting and sparse matrix problem [9]. HLT’11, 2011.
The final entity linking part is done by utilizing lookup dic- [8] John La↵erty,Andrew McCallum and Fernando Pereira,
tionary (DBpedia 2014) and sentence similarity. The en- Conditional Random Fields: Probabilistic Models for
tity’s tokens are given to the look up dictionary which results Segmenting and Labeling Sequence Data, 2001.
in few related links. The final link assigned to the entity is [9] Chun-Nam John Yu, Thorsten Joachims, Ron Elber
based on maximum similarity score between related links and Jaroslaw Pillardy, Support vector training of protein
and proper nouns in the test tweet. Similarity score is com- alignment models, in Research in Computational
puted by performing dot product between unigram vectors Molecular Biology, 2007.
of proper nouns in the test tweet and the unigram vectors of [10] Naoaki Okazaki, CRFsuite: a fast implementation of
related links from lookup dictionary. Entity without related Conditional Random Fields (CRFs), 2007.
links is assigned as NIL. [11] McCallum and Andrew Kachites, MALLET: A
Machine Learning for Language Toolkit,
4. DISCUSSION http://mallet.cs.umass.edu, 2002.
This experimentation is about sequence labeling for entity [12] Galen Andrew and Jianfeng Gao, Scalable Training of
identification from micro posts and extended with DBpedia L1-Regularized Log-Linear Models, ICML, 2007.
resource linking. By observing Table 1, it is clear that CRF [13] Jorge Nocedal, Updating Quasi-Newton Matrices with
shows great performance and paves way for building a smart Limited Storage, Mathematics of Computation, Volume
NER model for streaming data application. Even though 35, Number151, pp:773-782, 1980.
CRF seems to be reliable, it is dependent on the feature
65
· #Microposts2015 · 5th Workshop on Making Sense of Microposts · @WWW2015