This contemporary work is done as a slice of the shared task on Entity Extraction from Social Media Text Indian Languages in Forum for Information Retrieval and Evaluation (FIRE2015). Nowadays people are extensively using social media platforms like Face book, Twitter, etc, to exchange their thoughts. The twitter messages are growing rapidly and their style and short nature present a new challenge in language technology field. This extensive amount of textual data is also increases the interest in Information Extraction (IE) on such textual data. Named entity extraction is one of the essential tasks in Information Extraction, aims to extract and classify entities from text. The performance of the present standard language processing tools is severely affected on Tweet messages. Hence, different improvised and nonimprovised algorithms are necessary for extracting these entities from the informal text. This paper deals with the extracting the Named Entities from twitter messages of four Indian Languages. The extraction of the Named entity relies mainly on the domain specific features and conventional features. A well known supervised algorithm, Support Vector Machine (SVM) is used to extracting the entities. • Theory of computation~Support vector machines • Computing methodologies~Natural language processing • Information systems~Information extraction • Human-centered computing~Social tagging systems
Named Entity Recognition (NER) is a task of extracting the
named entities like people, organization and location etc. from a
text. NER is an important task in Natural Language Processing
(NLP) system for question answering and information extraction
[
The rest of this paper is composed of four sections. In Section 2, we discuss about the dataset descriptions and explained the size of training data and testing data. The methodology and the features are explained in Section 3. We report our experiment results in Section 4 and conclude the paper with future directions in Section 5.
Given the four Indian languages tweets as an input, the task is to
extract the named entities and identify its type. The task
organizers released the Twitter dataset and its entities for four
languages namely Tamil, English, Hindi and Malayalam. There
are 22 types of entities present in the training dataset in which the
majority of entities from person, location and organization. The
size of the training and testing data is shown in the Table.1 in
which the most of the tweet contains URL and leading topics are
politics and cinema. The major issue in the dataset is that the size
of the training data is lesser than the size of the test data (except
for Malayalam) which leads to a drop in the accuracy level of the
system. We have converted the tweets into the conventional BIO
format and treated them as a sequence labeling problem. Table.1
also describes the token counts of training and testing data of all
the four languages. Count of primary entities in the training
dataset is shown in the Table.2.
divided into small chunks called as tokens. Training and testing
tweets are tokenized. Given named entities and tokenized training
tweets are merged to create BIO format in one token-per line
fashion. The BIO formatted sequence data is then tagged using
POS tagger of the respective languages. For Tamil Language in
house Twitter POS tagger and Amrita POS tagger [
This section describes our system submitted in Entity extraction from twitter messages task in FIRE 2015. In the proposed Twitter NER system, feature extraction plays an important role, as the accuracy of the system rely on the extracted features. Generally, for text mining and information extraction, preprocessing is necessary and it is mandatory for the twitter dataset. Preprocessing includes Normalization and Tokenization. Generally the sentences are divided into tokens based on the white spaces. These tokens are further normalized where superficial variations are removed. But, normalization of Twitter messages aims to standardize the non-standard words, spellings variations, expand informal abbreviations (e.g., tmrw for tomorrow), and standardize phonetic substitutions. Here we avoided stemming and Twitter data normalization. Case folding is appropriate for English language where case variations are exist but it is not required for Indian Languages where no such variation are exists. The methodology of the proposed system is depicted in the Figure.2. The training dataset contains two files which are raw tweets and type annotated extracted entities. The tweet file consists of “Tweet ID”, “UserID” and Tweets. The entity file consists of “Tweet ID”, “UserID”, Entity Type, Entity, Starting Index and length. We have combined these files and converted into the conventional BIO formatted text in which B-XXX tag indicates the beginning word of an entity type XXX and I-XXX is used for the following chunks of an entity. The tag other than named entity is tagged as O. In tokenization the tweets are further
Word Features: The current word is used as feature. Additionally, to provide context information, the previous and next words of a current word are also used as features (w-1, w0, w+1).
Named Entities: Named Entity tag of the previous word, next words are used as features (t-1, t+1).
Prefixes and Suffixes: Prefix and suffix information are helpful to identify the named entities, particularly for unknown words. The first and last n characters are used as features; characters are ranging from 3 to 5 (P3, P4, P5 S3, S4, S5).
End features: Here, we have used four different kind of features ( . , ! ?). If the token ends with dot or comma or exclamatory or question mark, then the corresponding feature is set to 1. Otherwise, it is set to 0. For example, for the token “help!” the feature is “0010”.
Non-alphabet features: If a token contains hash (#) or any digit or punctuation or acronym with periods, then the corresponding feature is set to 1. Otherwise, it is set to 0. For example, for the token “B.J.P” the feature is “0011”.
Other features: Length of the token and the position in current tweet is considered as features for all the four languages. Language Specific features are used only for Tamil and English. For Tamil, we check whether the token contains ‘ரூ ’ or ‘ரூ .’ followed by digit and any digit followed by -ல் ,-இல் ,கு , க்ு . For English, we consider the digit followed by “st” “th”. In addition to that for English language we have used brown cluster (size of 10) and orthographic features like Starts with capital, all capital and contains apostrophe as features.
We have participated in all the four Indian languages and submitted our outputs to organizers.
Run-1: All the four languages with aforesaid features and SVM training without Validation.
Run-2: Only Tamil system has been submitted. In feature extraction, instead of standard POS tagger, we have used our own Twitter POS tagger which is developed with small amount of manually annotated examples.
Run3: Only Tamil system has been submitted. Tamil Twitter POS tagger has been used. In addition to that, in order to tune the SVM parameters, we have taken first 1000 tokens (from BIO format) as development set and train the system accordingly.
P 55.23 POS Tags: POS tags contain the essential information for identifying and disambiguating named entities. The current word’s part-of-speech tag, the previous and next words (p-1, p0, p+1) part-of-speech tags are used as features.
We have used the Twitter POS tagger developed by Gimbel
et.al.[
In this paper, we have presented an approach to identify and extract the named entities on the Twitter messages in Indian languages. Due to the huge volumes of rapidly increasing tweets, text mining applications like entity extractions are immediately required to accomplish language processing tasks in informal short texts. This entity extraction is highly helpful in the field of information extraction and trend analysis. Here we have extracted the different types of named entities, such as person, location, organization, etc. We have converted the training data into sequence labeling format and we extracted rich language specific features for training the SVM classifier. We would also like to perform a through error analysis to understand the issues, so we can easily address them in future. We want to invest time in crafting the informal text features to improve Twitter NER for Indian languages. Other directions includes, using Twitter POS tagger for Indian languages, adding entity specific prefix suffix features and word embeddings with unlabeled Twitter data.
First and foremost, we would like to thank Forum of Information Retrieval and Evaluation (FIRE 2015) organizers for organizing a wonderful research evaluation workshop and giving opportunities for researchers to present their work on Text Analytics and Natural Language Processing (NLP). We also like to thank Computational Linguistics Research Group (CLRG), AU-KBC Research Centre, for organizing the Entity Extraction from Social Media Text Indian Languages (ESM-IL) Task.