The progression of social media contents, similar like Twitter and Facebook messages and blog post, has created, many new opportunities for language technology. The user generated contents such as tweets and blogs in most of the languages are written using Roman script due to distinct social culture and technology. Some of them using own language script and mixed script. The primary challenges in process the short message is identifying languages. Therefore, the language identification is not restricted to a language but also to multiple languages. The task is to label the words with the following categories L1, L2, Named Entities, Mixed, Punctuation and Others This paper presents the AmritaCen_NLP team participation in FIRE2015-Shared Task on Mixed Script Information Retrieval Subtask 1: Query Word Labeling on language identification of each word in text, Named Entities, Mixed, Punctuation and Others which uses sequence level query labelling with Support Vector Machine. • Theory of computation~Support vector machines • Computing methodologies~Natural language Processing • Information systems~Information extraction • Humancentered computing~Social tagging systems
This paper describes our system for FIRE 2015 Shared
Task on Query Word Labeling on Mixed Script
Information Retrieval. The faster growth of internet in
current period the Webpages are not limited to English,
social media content in other languages increasing rapidly
[
The problem of language identification is researched for
half century (Gold,1967) and code switching for several
decades. But there has been less work on automatic
language identification for mixed script analysis in social
media websites and forums. Research showed that the
predominant language used in Twitter and Face book in
their earlier days was English [
In training data set, input query is constructed and annotated with their label. The query is written in roman script. Input query and annotated set are given as a part of the Subtask. The training data contains annotation and input file each have 2908 sentences (Tokens 54,088). The test data contains 792 sentences (Tokens 11,999). Tokens are person name, location, organization and abbreviation comes under NER label.
Tokens which constructed two parts, each coming from a different language are labelled as MIX, Emoticons, hash and punctuation are labelled as MIX. Foreign languages are labelled as O. There is no extra data set is used in this task. Input query many contain mixture of 1 or 2 languages, named entities, mixed, punctuation and others. Table 1 contains the counts for mixed, punctuation and others with overall token count. The languages token count is mentioned in Table 2. Named Entities have nine different tag set and total count of the NER tokens are mentioned in the Table 3.
We participated in the Query Word Labeling task which is described very briefly as follows: Suppose that q: w₁ w₂ w₃ is a query which is written in Roman script. The words, w₁ w₂ etc., could be standard English words or transliterated from another language L = {Bengali (Bn), Gujarati (Gu), Hindi (Hi), Kannada (Ka), Malayalam (Ml), Marathi (Mr), Tamil (Ta), Telugu (Te)}. The task is labeling the words as En or L. In Query word labeling we used Support Vector Machine classifier to predict language of a particular word which belong to either Indian language or English. As the training corpus is very huge, the words from the corpus are taken as features. As a method of preprocessing, the input raw data taken as token per sequence is annotated with corresponding tag set. This annotated data set is assigned as input for the machine from which the features are extracted. Various features are taken for better labelling of language. The three prefixes and suffixes of the current word, length of the present token, position of current word are taken as features. Punctuation, comma, colon/Semi Colon, dot and word starting with ‘@’ and ‘#’ are taken as binary features. This set of feature has been mainly used to identify Indian languages. They constitute checks on token endings in terms of presence of certain characters. Along with the features machine also learns from the training data set which is already labelled. When it comes to test data, same preprocessing step is carried out. Annotated test data is given as input for Support Vector Machine Classifier and classified output is taken. Sample output is given in Table 4. The query and corresponding tag set is given as a training data of the shared task and these are annotated as preprocessing procedure. Flow of the proposed system is illustrated in Fig 1. From the annotated data the features are extracted. Along with the extracted feature sequence of lines are given as an input for the Support vector machine classifier in which in creates a module file. The test data and module file is given to the classifier and output is extracted. Further the output is processed in which the utterance id is properly paired with test data.
Fig 1: Proposed system flow diagram.
In this paper, we described our system for Subtask 1 in FIRE 2015 - Query Word Labelling on Mixed Script Information Retrieval. The query word labelling is very useful in search engines. We used SVM classifier to identify languages, Punctuations, NEs, Mixed and Other. SVM uses set of features guaranteeing reasonable accuracy for mixed languages query and other tags. In proposed language identification system, the word sequences are divided into tokens which are trained using SVM classifier and the system is evaluated against the given test data. System is elevated separately for each tag in language pair, Mixed and Named Entities using Recall, Precision and F1Score.Concentration is required more on Mixed Script and NEs. As a future work words from the language dictionary and word as distributed vector can also be included as feature which will improve the accuracy of the system. Overall scores for tags set is mentioned in Table 5. Average F -Measure Weighted F-Measure 8.3333 36.3964 76.6231