With the recent growth of Internet, mobile and social networks the spread of fake news and click-baits increases drastically. Today, the fact retrieval system is one of the most effective tools for identifying the information for decision-making. We propose the approach based on factual information systematization. Different interpretations of the same phenomenon, as well as the inconsistency, inaccuracy or mismatch in information coming from different sources, lead to the task of factual information extraction. In this work, we explore how can natural language processing methods help to check contradictions and mismatches in facts automatically. The reference model of the factbased analytical system is proposed. It consists of such basic components as Document Search component, Fact retrieval component, Fact Analysis component, Visualization component, and Control component.
The access to the Internet, as well as to social networks has been simplified for the last decade. It has led to information flow growth. Social networking sites give an opportunity for users to share content freely. As a consequence, fake news, hoaxes, and click-baits are spread, circulated, consumed and shared without critical thinking or fact checking. Regardless the form, the reverberations of inaccurate or misleading information could lead to major risks for the society.
Misinformation can be spread both intentionally and accidentally. Among the
interested stakeholders of fake information are politicians, marketing managers, sellers,
and users with unclear purposes. In media politicians use alternative facts and
posttruths in order to manipulate their audience`s opinions creating thus long-term
sustainable mindsets. Alternative facts are information with no basis in reality while
post-truth technics are defined as beyond the truth or irrelevant information [
As for social networking sites, there are several sources of fake information search. Users’ profiles contain a lot of misinformation. In many instances it is difficult to match pages of the same user at different social networks because of a clash in personal information. As well as that it is impossible to assess reliable posts presented in the newsfeed.
E-commerce is another sphere where fake information is common. Sellers at market places present the name and the description of products. The same product can be described in different ways by different sellers. Apart from that, sellers, in order to present their products at as many search requests as possible sellers use clashing information.
Thus, the problem of identifying and verifying contradictory or ambiguous information is crucial. The key idea of our work is to develop an approach for checking contradictions and mismatch in facts automatically. 2
In recent years, the reliability of information on the Internet has decreased significantly. It is particularly noticeable on social networks, where distorted, inaccurate or false information reaches and affects millions of users within minutes. Therefore the problem of fake information detection has become a popular research sphere. There are several challenges to automatic detection of fake news: determining if the facts in the news article are correct; analyzing the relations between the article headline and article body; estimating the inherent bias of a written text etc. Factual analysis of the text is designed to make possible the intellectual analysis of data extracted from the text flow. The solution of this task should lead to a synergistic effect, to the possibility of using existing information technologies.
The paper [
The work [
The need for automatic hoax detection systems is a vital task. In the paper [
The goal of the paper [
In the paper [
The research [
There are number approaches of information extraction from natural languages texts. We can highlight lack of automated semantic understanding and low consistency of extracted facts. Despite of existing data extraction solutions the task of extracting facts still is not solved. 3
Today, the fact retrieval system is one of the most effective tools for identifying the information for decision-making. When you refer to something as a fact you mean that you think it is true or correct. Factual information is information based on facts or relating to the facts. The reliability of automatically extracted facts is the main problem of processing factual information. It is especially important because of increasing density of text information flow in mass media and various social networks, forums and blogs. Different interpretations of the same phenomenon, as well as the inconsistency, inaccuracy or mismatch in information coming from different sources lead to the task of factual information extraction.
We can consider facts as structured objects. This record describes real-world entity with its attributes mentioned in text, usually, who did what to whom, where and when. So, the fact can be extracted from the textual information and can determine the attributes of the object or the relations between objects. The task of Information Extraction is to identify instances, relations, events and their relevant properties in natural language texts. We consider two types of facts. They can be described as triplets. The first kind of fact is a “Subject -> Relation -> Object”, where the subject is who acts, the relation defines action with the object. The second kind of fact is a triplet: “Object-Attribute-Value”, where the object is the entity about which the fact is fixed. The attribute is predetermined characteristic that identifies the object with the certain values.
The extraction of facts from weakly structured textual information includes the following steps:
1) Entity Extraction – extract words or phrases that are important for describing the meaning of the text (lists of terms of the subject domain, personalities, organizations, geographical names, etc.);
2) Feature Association Extraction is searching the links between the entities extracted;
3) Event and Fact Extraction is extraction of entities, recognition of facts and actions.
To implement the entity extraction, a standard linguistic processor is used. The issue is the extraction of information about the relations between entities. For this purpose we need to define a certain template that reflects the semantic links in the sentence. Based on the fact definition, it is possible to define the minimal semantic unit of factual search, which is a triad: agent-predicate-value. That is, the record of factual information must include a pointer to the fact search agent, the attribute or predicate of this object, and give a specific value of this attribute.
Such a definition makes it possible to extract concepts from weakly structured text sources of information and to represent relations between them in a structured way. The resulting structure is facts, both in the form of fairly simple concepts: keywords, personalities, organizations, geographical names, and in a more complex form, for example, the name of the person with her job and occupation.
Algebra of finite predicates is used as a mathematical tool for describing discrete,
determinate and finite objects or processes from real world [
An analysis showed that the most natural and convenient tool for modeling natural
language relations is the algebra of finite predicates that operates with letter variables
[
Mathematical relation is the basic concept of logical mathematics. A logical
network is a processor and it performs various actions on relationships [
Any relation can be interpreted meaningfully as knowledge about the fact, expressed by some utterance. The fact is an exhaustive description of the actual state of all places interested to us. Knowledge of the fact only limits the many possible states of places. A statement about a fact can be true or false. It is true if the characterizing relation contains the actual set of place states and otherwise is false. 4
The model of fact extraction from natural language text can be presented in the
following way. The fact is considered as a triplet: “Subject -> Predicate -> Object”. The
predicate defines a relation, and the subject and object defines two entities. In the
developed model we introduce a set of grammatical characteristics of the sentence
words. To represent the triplet of fact we use approach proposed in [
We suggest the model of facts extracting based on the method of comparator
identification [
We discover that the descriptions of the same commodities in the trading platforms can be presented in a different way. We notice that such description can be presented as a triplet “object-attribute-value”. So, we can consider the second kind of facts. The description of commodity is represented as number of words; usually it is not a sentence, and a table with some characteristics of commodity.
Let E – be the set of structural elements of a web page, W – the set of words. Then RSEARCH E W – is the binary relation "is used for search". Let Eq E – the set of elements of the web page that are selected for estimation and Wq W – a set of words that match the topic of the search.
Binary relation RSEARCH eqi ,wqj | eqi Eq ,wqj W defines a "word-element" pairs. For that q pair, the words belong to the set of words which correspond to the topic and the elements which belong to a set of selected elements.
Let wpj Wp – a set of words extracted from the web page. Then the predicate which evaluates the binary of "element-word" pair:
1, if (eqi , wpj ) RSEARCH , Pw eqi , wpj
0, if (eqi , qpj ) RSEARCH .
The predicate that defines the presence of control words in a particular element: Pe eqi Pw eqi , wp1 Pw eqi , wp2 ... Pw eqi , wpn .
The web page estimation combines the estimates for each item and determined by the predicate:
Pq Peq1 Peq2 ... Pe
qs .
The page estimation is based on a data source model. The presence of different combinations of words in different combinations of elements of the web page is estimated. Let RSOURCE E W – be a binary relation "is used for sources selection", this is given as follows:
RSOURCE ei ,wj | ei Es ,wj wqi ,wqi W q , where wqi – set of words according to the element ei .
The predicate that estimates a pair of "element-words" is defined as:
1, if (ei , p p ) RSOURCE , P ei , w j
0, if (ei , p p ) RSOURCE .
The predicate that estimates an item using different word combinations: P( ei ) ( P( ei ,wp1 ) P( ei ,wp2 ) ... P( ei ,wpj )) ( P( ei ,wpj1 ) ...) ... Web page estimation for various combinations of elements is given by predicate: PS ( P( e1 ) P( e2 ) ... P( es )) ( P( e j ) ...) ...
The binary relation "elements and corresponding words that were extracted from the source page for representation the template model, RPAGE E W , RPAGE ( e1 , w1 ),...,( es , wj ). The function of transforming the word combinations into a value template from the set of "standards" C c1 ,...,cm is given as: c1, if (wi1 wi2 ...) (wj1 wj2 ...) ...
(ei , wj ) RPAGE : F(ei ) ...
cm, if (wim wim ...) (wjm wjm ...) ...
ei ,c j | ei E p ,c j C. set C c1,..., c m
A set of elements of a web page that contains a certain standard of a is given as E p e j E | c F( e j ),c C. Let RPATTERN E C – the binary relation "elements contain benchmarks", at the same The template predicate looks like:
1, if (e1e2e3 (E(e1, eF ) E (e2 , eI ) E (e3 , eO ))) 1, Ppattern
0, in other case. where
E( e1 ,eF ) 1, e EF , , 0, e EF ;
1, e E , E( e2 ,eI ) I 0, e E ;
I , and E( e3 ,eO ) 1, e EO ,
0, e EO .
Let K k j – a set of indicators that are relevant to this signs, then the predicate M ( k ,k j ) determines, whether the summary template has indicator k j from this set. Let I i – the set of indicators according to the given one, then predicate M ( i,i ) determines whether the generic template contains the data for this indicator i .
We can propose the reference model to factual information retrieval and analysis (fig.1). The main concepts are facts that are some knowledge about real-world objects, web-pages which contain text, indicators for representing attributes, and values of those attributes. The appropriate models must formalize the factual data processing.
In order to realize proposed reference model, the software should be developed. We suggest the basic components which are presented in figure 2. The developed software consists of such basic components as Document Search component, Fact retrieval component, Fact Analysis component, Visualization component, and Control component. Our future work is to make implementing and experimenting with the proposed model.
Summarizing, we can say that factual analysis is a rather complex system that has great potential and functionality. The tasks under which the data are built are designed to facilitate the work of analysts, to carry out filtration as well as structuring of huge volumes of information, which in our time are one of the main tasks of a person.
As result, we can underline that the task of identifying instances, relations, events and their relevant properties in natural language texts is still live issue. In general, we consider two kinds of facts. Despite existing data extraction solutions the task of extracting facts still is not solved. We propose to use predicate algebra and method of comparator identification to create a model of searching and extracting factual data. The future work will be devoted to research the similarity of facts and mismatch identification. We hope to develop a mathematical tool based on the relation of tolerance to make a conclusion about similarity or mismatch in the set of extracted facts. 6