Automating the process of Named Entity Recognition has received a lot of attention over past few years in Social Media Text. Named Entities are real world objects such as Person, Organization, Product, Location. Identifying these entities in social media text is an important challenging task due the informal nature of text present on social media. One such challenge that is faced in recognizing named entities in Indian Social Media Text is Code Mixing. Code Mixing is usage of more than one language in a sentence. Being a multilingual country, people of India tend to know more than one language, which in turn results in the code mixing of text while expressing their opinions. This paper describes the proposed approach for shared task CMEE-IL (Code Mix Entity Extraction in Indian Language), FIRE 2016. Proposed algorithm uses a hybrid approach of a dictionary cum supervised classi cation approach for identifying entities in Code Mix Text of Indian Languages such as Hindi- English and Tamil-English.
Computing methodologies ! Natural language processing; Information extraction; Language resources; Machine learning;
India is a multilingual country with more than 1600 languages being spoken such as Hindi, Punjabi,Bengali, Telugu, Marathi, Tamil, Gujarati and many more. With the introduction of Indic keyboards and articles that use Indic languages, people started using Indic languages in their normal conversation and this has made people to converse easily on the internet. With such a scenario, it is common that people know at least one more language apart from their native language due to which there is a high possibility that people mix words from two or more di erent languages while writing or speaking something. This mixing of words in sentences is referred as code mixing. Code mixing is present where people speak in an informal way, like in social media.
Growing usage of social media platforms like Facebook, Twitter and WhatsApp has led to an increase of code mix data present because of interplay of Indian languages. Hence for making best use of this data it needs to be analyzed. Indic languages are used very much nowadays in general conversations but there are a few Natural Language Processing (NLP) resources that are available.When code mixed text is combined as well very less resources are present. Hence there is a greater need for developing NLP tool that can handle code mixed texts with Indic languages.
Entity recognition is a very important subtask of Information extraction and nd its applications in information retrieval, machine translation and other higher NLP applications such as coreference resolution. Named Entities are names of famous persons, organizations, locations, animals etc.Named entities have a many uses like in sentiment analysis,where recognizing named entities is important as they don't add much value to the statement. Similarly while tagging articles named entities are required for better search results. There are many such uses where named entities are used so named entity recognition is very important.Towards this, FIRE 2016 has organized a task for entity recognition in code mix text for Indian languages which identi es named entity in code mix text of English-Hindi and Tamil- English code mixed tweets. The Task was to identify the various entities such as person names, organization names, movie names, location names in a given tweet.
Rest of the paper is organized as follows. Section 2 explains the related work that has been done in the past few years. Section 3 presents the analysis of data set provided by CMEE-IL 2016 Task Organizers. Section 4 explains the Proposed Technique that have been performed for the task with block diagrams. Section 5 discusses algorithm to explain the procedure. Section 6 elaborates the evaluation and experimental results and error analysis. Section 7 concludes the paper and presents future work. 2.
In Named Entity recognition there has been signi cant
research done so far in English but same cannot be said
for Indian Languages due to rich morphology of indian
languages. Sujan et al [
In current scenario there has been a lot of work going on
recently found out trend of code-mixed texts in Indian
Languages. Parth et al. [
Data Set provided by task organizers contained two code mix data set, Tamil-English and Hindi- English. In Each dataset , the training data consisted of two les, A text le containing raw tweets along with their tweetID and UserID and another text le containing Annotations to the tweets present in the raw tweets le. The raw tweet les consists of 2700 tweets in the Hindi-English corpus and 3200 tweets in the Tamil-English corpus. All the tweets in the HindiEnglish corpus were already romanized whereas the tamilenglish corpus had a mixture of both tamil script and romanised script. There were 22 tags present in the corpus as mentioned in Table 1.
Named Entity (NE) Tag Person, Entertainment and Location occupies majority of the instances in Tamil -English corpus. Person tag comprises of Names of Famous Actors, Actresses, Politicians, New Reporters and Social Media Celebrities. Entertainment Comprises of Names of Famous TV shows and movies while Location consists of Names of Famous Cities, Indian Towns and Names of Countries.
Apart from these, there are some Numerical and Time based Tags that are present as well which comprise of the remaining part of Tamil-English data set. These tags include Count, Distance, date, money, month, time and year. Money represents numbers along with a monetary tag like '15 dollars'. Organization is another tag present which consists of names of organizations. The rest of the tags have very less annotations present in the annotated le.
Comparing Hindi-English with Tamil-English, The percentage of tags for the minority tags remains almost the same. But in the Hindi- English corpus the Tag Entertainment has the highest number of annotations present.It is followed by Person which is close to Entertainment.The rest of order remains the same but with varying percentages. 4.
A word level NE-recognition system is designed to recognise Named Entities in a tweet. The proposed methodology involves a pipelined approach for detecting each NE tag and has been divided into following four phases:
3. Gazetteer List Based Named Entity Recognition
4.1
The Data is pre-processed before detecting named entities. This is done to ensure that the data is uniform and the system can bene t from that. The preprocessing consists of creating a copy of the string in lowercase for uniformity. It also removes all links present in the tweet. This preprocessed tweet along with the original tweet is then passed to the next phase. 4.2
This phase of proposed algorithm identi es number based entity such as date, time, month, day, year, money, period, quantity, distance and count using a set of Regular Expressions Regular expressions are designed based on the common patterns observed in the annotations for these tags. Regular expressions work best in detection for these tags because there are limited variations possible for each of these Tags. For example the tag Day can be only one of the 7 possible days of a week in a language. So detecting them using regular expressions will be e cient. While checking for NE tags there is a possibility of having multiple tags attached to the same token. To remove ambiguity proposed technique checks for tags in a particular pre de ned order. 4.3
As shown in Table 1, apart from Entertainment, Location, Person and Organization, Rest of the Tags contain very less data that cannot be used to train a classi er. So Gazetteer Lists are used for identifying Named Entities with insu cient training data. Gazetteer lists are created from the annotations given to the training data. While checking in gazetteer list # and @ symbols were ignored. 4.4
The rest of the NE Tags are identi ed by creating feature vector for each token of a tweet. These feature vectors are then trained using a decision tree and extremely randomized tree classi er . The features considered for building feature vector are mentioned in Table 2.
English dictionary feature is used to identify the presence and absence of an english word i.e if it is a english word it is 1 and if it is a non english word then its 0. Python dictionary called pyenchant 1 was used for this identifying this feature. Also for pre x su x feature a dictionary was built using most common pre xes and su xes and presence of these pre xes and su xes (length = 1 to 3) in tokens were identi ed using the same. Gazetteer list feature checks for the presence of the token in gazetteers list of the remaining tags and uses this as a feature. Previous token tag was also taken into account to check structure of the tweet. Using all 1http://packages.python.org/pyenchant/
Algorithm 1 explains the proposed technique for Named Entity Recognition of code mixed text. The System rst pre-processes the input by removing the website and twitter links (implemented by callable: Link remover ) and then converts the tweet into lowercase (implemented by callable: Case conversion). We check for all numerical features like date, time, money, quantity, period, distance, day and count (implemented by callable: check Numerical).Before adding it to the nal predictions we check for overlapping Tags and remove them (using add without repetition). This is the second phase of the system.
In the third phase, we tokenize the tweets (using the function Tokenize) and check if any of the token is present in any of the gazetteer list (using check gazetteer List) and add them to the nal list of tags of that tweet.This ends the third phase of the system. In the nal phase we create feature vectors for each tweet and then predict using the classi er (clf) already trained using a training data. The classi er(clf) used are decision trees and extremely randomized trees.
Algorithm 1 Algorithm for Identifying Named Entity Recognition for Code Mixed Text in Indian Language 1: Input: Code-Mixed tweets list , S 2: Output: Predicted Named Entity Labels, P 3: Initialization: P=[], toks=[], NE Data=[], NE Tags=[] 4: for i=0 to S.length do 5: Link remover(S[i]) 6: Case conversion(S[i]) 7: end for 8: for i=0 to S.length do 9: d=check Numerical(S[i]) 10: add without repetition(d) 11: tok =Tokenize(S[i]) 12: for j=0 to tok.length do 13: g = check gazetteer List(tok[ j ] ) 14: add without repetition(P , g ) 15: f = Create Feature vector(tok[ j ]) 16: c=clf.predict(f) 17: add without repetition(P,c) 18: end for 19: end for
CMEE-IL, FIRE 2016, had proposed the task of Named Entity Recognition in Hindi-English and Tamil-English Code mixed text. Three runs were submitted for the task evaluation for each language pair. Four versions were created for di erent runs submitted. In all the versions, numerical feature were detected using numerical function as explained in section 4.2. Rest of the tags were classi ed using di erent versions created as speci ed in Table 3.
The rest of the tags were classi ed using gazetteer list phase as mentioned in section 4.3.This was done due to low amount of training data available for few tags such as plants, disease, locomotive etc as mentioned in Table 1. All the Variations for the proposed algorithm were evaluated using F-Score for each language pair and nally Run1, 2, 3 for Hindi-English used versions 1, 2 and 4 respectively whereas Run 1, 2, 3 for Tamil-English used versions 1, 3 and 4 respectively. 6.1
As shown in Table 4, run 2 performed the best for HindiEnglish Proposed System. Similarly, run 2 performed best for Tamil-English as well as indicated in Table 5. Based on the F-Score, it can be concluded that algorithm with more gazetteer List and Extremely Randomized forest (Version 2) performed well in case of Hindi- English. But in case of Tamil-English, algorithm with less Gazetteer List and Decision Tree (Version 3) proved to be e ective.
Precision value could be less because of string matching with elements of the gazetteer lists. Also, it can observed that recall value is low for all the runs, this could be due to less number of Named Entities for any sentence which in
Few phases in proposed approach might have attributed to misclassi cation for few tags. One such phase can be the gazetteer list phase of the proposed method which is a dictionary based approach and has disadvantages associated with it. When there is less data present in the dictionary the precision will be low for the system. So there is a need for more elements in the list for a better recognition system. Also, if there is an ambiguity in tags then there a chance of misclassi cation. For example if we say there is a token 'Honey' it can represent a Person like 'Honey Singh' or as a tag material. This ambiguity can only be solved using a classi er.
In this paper, a hybrid approach of a dictionary cum supervised classi cation approach for identifying entities in Code Mix Text of Indian Languages such as Hindi- English and Tamil-English is submitted for the task CMEE-IL,FIRE 2016. The proposed system used a pipelined approach to identify the named entities. There are four variants of the system based on the number of tags, the classi er can detect and the classi er used. Further improvisation can be done by incorporating features related to the structure of the sentence. POS Tagging and Chunking of the tweets has not been included, this an be another improvisation in the proposed algorithm. Although we need better POS Taggers for Code mixed Languages which can tag both romanised and non-romanised tweets for the same.