FINDING OUT NOISY PATTERNS FOR RELATION EXTRACTION OF BANGLA SENTENCES

Relation extraction is one of the most important parts of natural language processing. It is the process of extracting relationships from a text. Extracted relationships actually occur between two or more entities of a certain type and these relations may have different patterns. The goal of the paper is to find out the noisy patterns for relation extraction of Bangla sentences. For the research work, seed tuples were needed containing two entities and the relation between them. We can get seed tuples from Freebase. Freebase is a large collaborative knowledge base and database of general, structured information for public use. But for Bangla language, there is no available Freebase. So we made Bangla Freebase which was the real challenge and it can be used for any other NLP based works. Then we tried to find out the noisy patterns for relation extraction by measuring conflict score.


INTRODUCTION
Natural language processing (NLP) is a branch of artificial intelligence which describes interaction of human and computer by manipulating human language. Its goal is to fill up the gap between human communication and computer understanding. Here, relation extraction (RE) is a fundamental topic of NLP. It is actually the task of finding semantic relationships between pairs of entities. Relation extraction is essential for many well-known tasks such as knowledge base completion, question answering, medical science and ontology construction [1]. There are so many unstructured electronic text data available on the web like newspaper, articles, journals, blogs, government and private documents etc. But unstructured text can be turned into structured by annotating semantic information.
Here, entities can be like person, organization, locations. We have to identify the entity types of a sentence. A relation is defined in the form of a tuple t = (e1, e2, ..., en) where the ei are entities in a predefined relation r within document D. Most relation extraction systems focus on binary relations. Examples of binary relations include born-in(Ruma, Dhaka), father-of(John David, Eric David) [8]. For relation extraction, we have different methods like supervised method, distant supervised method and unsupervised method.
In supervised method, sentences in a corpus are first hand-labeled for the presence of entities and relations between them. Lexical, syntactic and semantic features have to be extracted by the automatic content systems (ACE) to build supervised classifiers to label the relation between a given pair of entities in a test set sentence. Labeled training data is expensive to produce and thus limited in quantity. Another one is distant supervised method for relation extraction which aligns texts to the given KB and use the alignment to learn a relation extractor [3]. They use the large amounts of structured data sources (such as Freebase) as the distant supervision information. As these methods do not need a hand-labeled dataset and KBs grow fast recently, they are more efficient.
In our work, we tried to find out noisy patterns for relation extraction of Bangla sentences for distantly supervised method. For this method, we need seed tuples which we can get from knowledge base like (Freebase). But there is no available Freebase in Bangla. So we built Bangla Freebase which contains a large amount of relations.

PREVIOUS STUDY
Distant supervision can be introduced as an efficient method to scale relation extraction to very large corpora which contains a lot of relations. The authors proposed a sentence-level attention model to select the valid instances, which makes full use of the supervision information from knowledge bases [2]. And entity descriptions from Freebase and Wikipedia pages to supplement background knowledge have been extracted for their task. The background knowledge not only provides more information for predicting relations, but also brings better entity representations for the attention module. Three experiments have been conducted on a widely used dataset and the experimental results showed that their approach outperforms all the baseline systems significantly [2].
Modern models of relation extraction for tasks like ACE are based on supervised learning of relations from small hand-labeled corpora. The authors used a paradigm that does not require labeled corpora [3]. This paradigm can avoid the domain dependence of ACE style algorithms, and allow the use of corpora of any size. This experiment used Freebase which is a large semantic database of several thousand relations. The Freebase provides distant supervision. For each pair of entities that appears in some Freebase relation, all sentences containing those entities in a large unlabeled corpus have been selected and extracted textual features to train a relation classifier. Their algorithm combines the advantages of supervised IE and unsupervised IE [3].
There have been so many works of relation extraction entities in English. In this work, we have worked on relation extraction of Bangla sentences on which not so much research work has been done. So it will be very much beneficial for this language. This is the nobility of our work.

CREATING BANGLA FREEBASE
Freebase is a large collaborative knowledge base and database of general, structured information for public use. Its structured data had been harvested from many sources, including individual, user-submitted wiki contributions. Its aim is to create a global resource so that people (and machines) can access common information more effectively [9]. It is available in English. Actually in Freebase, the triple format is like (e1, r, e2) where e1 and e2 are the two entities and r defines the relation. So relation can be found in a known KG and can generate large amount of data [4]. Since we mentioned before that we created our own Bangla Freebase which contains a large number of relation with the help of Wikidata query service and SPARQL query language. It is a large collection of knowledge base database. Today the number of Bangla articles in the internet is growing day by day. So it has become a necessary to have a structured data store in Bangla. It consists of different types of concepts (topics) and relationships between those topics. These include different types of areas like popular culture (e.g. films, music, books, sports) location information (restaurants, locations, businesses), scholarly information (linguistics, biology, astronomy), birth place of (poets, politicians, actor, actress) and general knowledge (Wikipedia). Here we collected more than 100 relations according to our need .By using SPARQL query, anyone can find out their required realtion. So this knowledge base is very much helpful. It will be much more helpful for relation extraction or any kind of NLP (Natural Language Processing) works on Bangla language.

Wikidata Query Service
Wikidata is a website that belongs to the Wikimedia family of websites. Data from Wikidata is available in RDF dumps. Actually RDF stands for Resource Description Framework which is a general method for describing data by defining relationships between data objects and it allows data integration from multiple sources. RDF has triple format which is a set of three entities that codifies a statement about semantic data in the form of subject-predicate-object expressions [7]. Wikidata is a place to store structured data in many languages. The basic entity in Wikidata is an item. An item can be a thing, a place, a person, an idea or anything else. Wikidata has identifier numbers for entities and properties.

Entity Identifier
As Wikidata treats all languages in the same way, items don"t have names, but generic identifiers. Each identifier is the letter Q that is followed by a number. For example, the item about the capital of Japan is called neither "Tokyo" nor "anything" but Q1490. But to give it a humanreadable name, each item has a list of labels in each language associated with it. So we"ll see that the English (en) label at Q1490 is "Tokyo", also has corresponding word for the Japanese (ja) label, the Bangla (bn) label and so on.

Property Identifier
Every item has a list of statements associated with it. Each statement has a "property" and a "value". There is a long list of possible properties. Like items, properties have generic identifiers, but they begin with the letter P and not Q. For example, the property to indicate the country is P17, and it has the label "country" in English. The value of P17 (country) for Q1490 (Tokyo) is Q17 (Japan, etc.).

SPARQL Query Process
It is necessary to extract information from complaints, either scraped from the Web or received directly from the client for many companies nowadays. The aim is to find inside them some actionable knowledge. There is a query language, SPARQL to extract information from natural language documents, pre-annotated with NLP information. SPARQL stands for SPARQL Protocol and RDF Query Language. It is an RDF query language and able to retrieve and manipulate data stored in Resource Description Framework (RDF) format. SPARQL allows query to consist of triple patterns. It was made a standard by the RDF Data Access Working Group (DAWG) of the World Wide Web Consortium. SPARQL allows users to write queries against what can loosely be called "key-value" data [6]. We have to follow these steps: • We can retrieve data according to our need by making query and for this, we have an online query service engine known as Wikidata query service. The URL of online query service is https://query.wikidata.org/. • SPARQL is a standard query language technology which is endorsed by the World Wide Web consortium for querying any linked data information source. For making a query, we have to understand the SPARQL query. Here, one query has been added thorugh which we will retrieve all the poets who are the citizen of Bangladesh.

METHODOLOGY
Relation extraction is very much significant in NLP based work to extract the relation between two entities [11]. There are lot of methods for RE. In our work, we use distant supervision for relation extraction. In distant supervision, an already existing database, such as Freebase (knowledge-database) is prepared. Then we gather examples for the relations we want to extract. Thus our training data will be prepared. For example, Freebase contains the fact that Paris is the capital of France. We then label each pair of "France" and "Paris" that appear in the same sentence as a positive example for "capital_of_the_country" relation. A large amount of (possibly noisy) training data can be generated. In the research work we needed seed tuples which are collected from Bangla Freebase made by us. Distant supervised method has been used which is very much efficient. In each seed tuple, there are two entities and their relation. There may be different types of entities like person, organization, location, films etc. We then extracted features from the sentences containing those entities in a large corpus. So we can say, our goal is to extract relation between two entities from sentences in a triple format and map the triple elements existing in a knowledge base [5]. After that we made decision that these extracted features are valid or not for each relation by measuring conflict score.

Name Entity Recognition
For our work, we had to identify the entity for each sentence. For entity identification, we used word level features (e.g., token, prefix, suffix), list lookup features (e.g., gazetteers). Gazetteers include names of countries, major cities, common people name, organization name etc.

Preparation of Corpus
A corpus is a collection of real world text in linguistics research. The intuition of our distant supervision approach is to use Freebase to give us a training set of relations and entity pairs that participate in those relations. In Freebase there are hundreds of relations. Freebase works as seed tuple. In our seed tuples, the Bangla synonym for the relations had been used. The seed tuples have different relations like "birth-place", "working-place", "actor", "film-director", "film- For each pair of entities in the seed tuple that appears in some Freebase relation, we used Wikipedia because it is relatively up to-date. We found out all sentences containing those entities in Wikipedia or large unlabeled corpus and collected them. Then we worked on them and extracted textual features. A part of our corpus looks like below:

Preparing Data and Pattern Identification
Before identifying patterns, we performed some tasks. We had to prepare the data. The following steps are performed: • Tokenizing: We tokenized the data by inserting spaces between words and punctuations.
• Cleaning: Then we cleaned the data by removing empty lines, extra spaces and some lines that were too short or too long.
• Entity identification: In a given sentence, we identified the entity types and extracted the relation patterns between them.
• Chunking: In preprocessing, consecutive words with the same named entity tag are "chunked", like Ruma/PERSON Rahman/PERSON. So in a sentence if Ruma and Rahman appear together, these will be chunked together like [Ruma Rahman]/PERSON.
• Lexical features: Then to find out the pattern, we worked on lexical features of each sentence. Like:i) the sequence of words between the two entities is very much important. ii) A flag indicating which entity came first in the sentence iii) A window of n words to the left of Entity 1 iv) A window of n words to the right of Entity 2 These lexical features help us to identify the patterns of each relations. For example,

Conflict Score Formula
The main goal of our work is to identify the valid patterns for relations and find out noisy patterns existing there. The task can be done by identifying the conflict score for each pattern. Bangla Freebase is considered as seed tuples. Each seed tuple contains two entities and their relation. Our corpus has sentences containing these entities to get distant supervision. Our research work takes 5 relations. They are place_of_birth, place_of_work, living_place, film_director, film_producer and for the relations, we used the Bangla synonym of the words A threshold value is fixed which is 0.3. If the conflict score is less than or equal to threshold value then it is a valid pattern. Otherwise the pattern is invalid or noisy for a relation. For person and organization entities we take three relations. We know it is not necessarily for a person that his working place and birth place will be same. It helps us to find out the conflict patterns.

RESULT AND EVALUATION ON CONFLICT SCORE
We showed our results on five relations for Bangla sentences. The threshold value is 0.3. If the conflict is less than or equal to 0.3 it is considered as the valid pattern for that relation. Otherwise it is the noisy pattern. In future work we will remove these noisy patterns to train our relation classifier. The conflict scores have been mentioned in the following: Table 4. Valid pattern identification for (place_of_birth) relation

FUTURE WORK
Here in this research, we have worked on relation extraction patterns of Bangla sentences. RE is used in information extraction. As Bengali articles are increasing in the Web, this work holds very much significance for Bangla language based research work. Bangla language is very much enriched. In future, we will build a classifier where noisy patterns for any relation will be removed.

CONCLUSIONS
Relation extraction is a significant topic in NLP. In this work, we made a Freebase for Bangla which is a large collection of structured data by using Wikidata query service. It will be much more helpful for further research work like in natural language processing of Bangla, where the researchers need to get seed tuples. Actually researchers in areas such as name entity extraction data mining, Semantic Web, information retrieval, question answering, genetic disease analysis, ontology creation and analysis can use this to support their work. With the help of Freebase, we get our seed tuples. We worked on Bangla sentences for relation extraction using distant supervision. Then we found out the noisy patterns using conflict score.