Due to today's information overload, the user is particularly nding it di cult to access the right information through the World Wide Web. The situation becomes worse when this information is in multiple languages. In this paper we present a model for information extraction. Our model mainly works on the concept of speech tagging and named entity recognization. We represent each word with the POS tag and the entity identi ed for that term. We assume that the event exists in the rst line of the document. If we do not nd it in the rst line, then we take the help of emotion analysis. If it has negative polarity, then it is associated with an unexpected event which has negative meaning. We use NLTK for emotion analysis.
The overload of today's information, throws enormous di culty to access the
right information especially through the World Wide Web. The situation
becomes worse when the information is in multiple languages. When a user
accesses a document written in multiple languages and the user faces di culty in
nding facts, it is important to remove all information from the data except only
the facts that match the user's interest. To extract various types of information
from document pertaining to speci c languages and domains, Information
Extraction (IE) systems are primarily used [
Most IE systems process texts in sequential phases (or \steps" ) like lexical
and morphological processing, identifying and typing appropriate names,
analyzing large syntactic components, nal extraction of anaphora and co-referent, and
relationships with domain-release events and lessons [
In Event detection, Topic Detection and tracking are one of the major
component of a broad initiative. Written and spoken news stories are primarily belong
to the category of Topic Detection and tracking interest based broadcasting news
[
Recent research in this area starts with the notion that statistical machine
learning is the best way to solve information extraction problems. To nd
structured information with uncontrolled or semi-structured text are the primary
objective of information extraction. [
Our model primarily works on the concepts of Part of Speech tagging and Named-entity recognition. We represented each word with POS tag and identied entity for that word. We assumed that the event exists in the rst line of the document. If we do not nd it in the rst line, then we took help of sentiment analysis. If it has negative polarity, it was found to be associated with an unexpected event that has a negative sense. We used NLTK for sentiment analysis.
Most of the information is easily extracted by Named Entity Recognition such as Time-argument(Date), and Place-argument(Location). The Speed-argument and Casualty extraction are rst tagged by \CD" and further distinguished with the help of NER tag. 4
We developed a modular method for information extraction from the Indian language. The dataset provided by organisers is collected from various social sources such as blogs, microblogs, social media and newswires that are either in Roman script or codemixed. Here, Indian language is mixed with English. We have worked on English and Hindi languages. The key components of proposed work (Figure 1) are described as follows.
Dataset the task organiser provide contains three languages: English, Hindi and Tamil. Training dataset is in the xml format and testing data in raw text. The statistics about given dataset are shown Table(1). We have used Stanford POS Tagger for the part of speech tagging to English dataset. We have used Hindi treebank dataset to make POS Tagger applying conditional random eld algorithm. Our entire system relies on POS tagger. The part of speech tagging is the rst step to perform extraction task.
Similar to POS Tagging, Stanford NER Tagger was used for named entity recognization to English dataset. For Hindi language, similar to POS Tagger we use Hindi treebank dataset to create NER Tagger. By default, the Stanford NER Tagger neither use part of speech nor gazette to extract locations. The pair of POS Tag and NER Tag along with word is helpful to extract information about Person, Date,..etc. 4.3
When we want to extract full information related to the event such as casualties, depth,..etc. then it is necessary to identify the whole phrases. Thus, the complete phrase is extracted by Stanford Parser. Sentiment analysis the extracted phrase that has tagged as a verb followed by a cardinal number but not date NER. This may have information about casualties. This extracted pattern will fail when some positive events follow the casualties' pattern. To avoid such situation we check the sentiment polarity, the threshold value for sentiment checking is 0.5.
The proposed method is a framework for information extraction from unstructured user generated contents on social media. Our information extraction systems analyse human language text as linguistics structure in order to extract information about di erent types of events, time, place, casualties and speed. We Selected the sentences from the dataset and perform the POS Tagging, NER Tagging and Chunking then extracted the phrases from the above-POS, NER and Chunk.
{ Time extraction: We assume that if we got date tag in NER tag then that is the correct date otherwise we match the day name or day abbreviation in calender library. Apart from that we also match the word with Today, Tomorrow and yesterday strings. { Event extraction: We took the instance from the dataset and applied the lemmatizer on the selected instance and extracted the most frequent 10 lemmas that should not be the stopword and punctuation mark and for a selection of an event among those lemma that should be the noun and belongs from rst sentence of instance. { Place extraction: Those extracted phrases, who must contain the noun, proper noun and preposition POS tag from the dataset and the selected phrases that shows location in NER tag and those words should start with a capital letter then it represents the place. { Casualties extraction: The phrase which we selected for the extraction must hold the cardinal number, verb, noun and preposition POS tag then only we will select that phrase. After the data analysis, we found the cardinal number and verb, the window size should be from 1-5 but there is some selected phrase which represents time and the next we check the NER tag should not be Date tag. { Speed extraction: for the selection of phrase, the phrase which contains the cardinal number, noun and preposition then we selected those phrase. Those phrases should not match with extracted Time and Casualties and the word which hold the cardinality number's word we check that to measurement unit for the remaining phrase. The measurement unit is created by a manual dictionary which keeps the almost measurement unit. 5
The evaluation matrix for the Event Extraction problem is shown in table 2. The task organiser includes the Precision(P) and Recall(R), F-measure: T P P recision = T P +F P
T P
Recall = T P +F N
F measure = 2 PPrreecciissiioonn((PP))+RReeccaallll((RR)) Where TP is truly positive, FP is false positive, FN is a false negative. We have extracted ve information out of seven information i.e Speed, Casualties, Time, Event, Place.
We clearly saw that our Hindi model is not performing well as compare to the English model because there is no free existence of Hindi POS Tagger and NER Tagger available yet. 6
We have discussed our Rule-Based methodology used to solve the task of information extraction from newswires and social media text in Indian languages. Our methodology tested on Hindi and English language and have derived some insights from the achieved results. The achieved F-measure are 39.98% and 45.07% for Hindi and English respectively. We believe that the incorporation of probabilistic approach with Rule-Based will improve the results.