Semantic Oriented Linguistic Processors
· The Linguistic Processor provides the deep analysis of texts in natural language with knowledge extraction, which are interested by users in his subject fields. Process of text analysis by the Linguistic Processor is presented on following pictures:
In the paper the new class of Semantic Oriented Linguistic Processors (LP) for knowledge extraction from texts (about criminal events, correspondence of mass media and so on) are considered. Processor LP transforms the texts to the structures of Knowledge Base (KB). Processors LP select from texts the semantic significant information which is interested by users, i.e. the information objects (named essences), their quality and quantity characteristics and interconnections. For example, they may be persons, their addresses, telephones, organizations and so on.
2. Components of Semantic Oriented Linguistic Processor
A variety of semantic search forms is supported. The linguistic processor LP comprises the following components.
1) The component of lexical and morphological analysis. It extracts words and sentences from the text, performs lemmatization of words (normal form establishment) and constructs the semantic network presenting the space structure of text (SpST), which reflects the sequence of words, their basic features, beginnings of sentences and the presence of space character lines. The component uses a two-level general ontology and a special collection of subject dictionaries (the dictionary of countries, regions of Russia, names, forms of weapons, and other items specific for the supported domains). The component performs semantic grouping of the words and assigns them additional semantic attributes.
2) The component of syntactic-semantic analysis. It converts one semantic network (SN) into another one which represents the semantic structure of text (SemST) the, i.e., the relevant semantic entities and their connections. The SemST is called the meaningful portrait of document. It comprises knowledge structures of the knowledge base which serve the basis for implementing different forms of semantic search : the search by features and connections, the search for the entities connected at different levels, the search for similar persons and incidents, the search by distinctive characteristics (with the use of ontologies).
The component is controlled by the linguistic knowledge (LK), which determines the process of text analysis. LK includes special form of contextual rules which ensure the high degree of selectivity with the extraction of entities and connections. The functions of this component are as follows:
- Extraction of information objects from the flow of NL documents: persons, organizations, actions, their place and time, and many other relevant types of objects.
- The establishment of the connections of between objects. For example, persons are connected with organizations (PLACE_OF_WORK), by addresses (LIVES, REGISTERED). Or figurants of criminal cases are connected with such objects as the type of weapon, drugs (TO HAVE).
- The analysis of finite and nonfinite verbal forms with the identification of the participation of objects in the appropriate actions. For example, one figurant gave drugs to another figurant, and this is the fact linking them.
- The establishment of the connections of actions with the objects by place or time (where and when some action or event occurred).
- The analysis of the reason-consequence and temporary connections between actions and events.
3) Expert system component (ES). On the basis of semantic networks the new knowledge pieces are
constructed in the form of additional fragments (ESN). For example, the ES extracts the area of a person
activity (in accordance with the assigned classifier) from the text of resume for each autobiography. The
experience of the person’s work is evaluated. The correlation of a criminal incident to the specific type is
accomplished with the analysis of the criminal actions of ES: the following facts are revealed - the nature of crime, the method of its accomplishment, the instrument, and so forth (in accordance with the classifiers of the criminal police).
4) Reverse linguistic processor, which converts the meaningful portrait of document (semantic network) into NL-text (for users) or the XML- file (for KB). In this case the necessary replacements of symbols, service words (names of objects) are achieved, the markers of beginning and end of the objects, actions, sentences. Conversion is achieved without the loss of information.
5) The base of linguistic and expert knowledge (KB). It contains the rules of the text analysis and expert solutions in the internal presentation. They determine the work of the linguistic processor. Processor LP has several such bases, which are activated depending on subject areas and user tasks.
3. The Entities and Links for Extraction
The set of the entities to be extracted depends on the tasks of a user. At the same time the quality of a linguistic processor is to a considerable degree determined by the possibilities for this extraction. The processor LP supports more than 40 types of semantic entities which can be extracted automatically.
Some examples of basic entities types and connections extracted by LP are given below:
• persons (by family name, given name and patronymic - FNP) with their role features (criminal, victim);
• the verbal description of the persons, their distinctive signs;
• address, posting information attributes;
• date(s) mentioned;
• weapon with its special features;
• telephone numbers, faxes, e-mails with their subsequent standardization;
• the means of transport with the indication of the vehicle type, its state number, color and other attributes;
• passport data and other documents with their attributes;
• explosives and narcotic substances;
• organizations, positions;
• quantitative characteristics (how many persons or other objects participated in an event);
• the numbers of accounts, sums of money with the indication of the currency type;
• terrorist groups and organizations;
• participants of terrorist groups with the indication of their roles (leader, head of, etc.);
• the armed forces, assigned for antiterrorist combat (Military_.Force);
• event (criminal, terrorist, biographical, and so on) with the indication of the information objects participation in them;
• time and the place of events;
• the connection between different types of information objects (with whom a person works in an organization, or lives at the same address, in what events participated together with other objects, etc.).
For extracting objects all versions of an object name including the contracted form possible in the text were considered. Standard objects (names, dates, addresses, types of weapons and others) are reduced to one (standard) form. The identification of objects is performed taking into account brief designations (for example, separate surnames, patronymics, initials), anaphoric references (indicative and personal pronouns, for example, "this person", "it...") definitions and explanations (for example, "the mayor of Moscow Luzhkov" is identified with the subsequent words "mayor", "Luzhkov"). For the extraction of events and connections the analysis of verbal forms, participial and adverbial constructions is carried out. An important task is the identification of entities in the entire text, the use for these purposes of indicative pronouns, brief names, anaphoric references.
4. Factors of Processor Quality
The quality of a linguistic processor is determined by a number of factors.
The first one is the facility of entities and connections establishment. The processor LP outperforms the existing systems by the number of the supported semantic entities types. It identifies more than 40 types of entities including very complex ones, which correspond to actions and events, for comparison the competitors’ best result is about 15 types.
The second important factor is the selectivity of rules and procedures of identification: the factor of the noise and losses. By noise we mean the presence of excessive words in the entities. Losses are the situations when an entity is not revealed or revealed partially: in the text there are the words, which did not enter into the entity. In the processor LP the rules are arranged in such a way that they ensure the high degree of selectivity and the minimization of noise and losses with the large number of the entities being selected.
The third factor is the possibility and the labor expense for tuning to a corpus of texts (for increasing the selectivity of rules for extraction of entities), and also tuning to the new subject domains and types of entities. Due to the complexity of analysis this tuning is achieved through the linguistic knowledge (LK).
The linguistic processor LP ensures the analysis of the Russian and English language forms with the aid of the uniform language model.
The fourth factor is the speed of linguistic processor operation, i.e., the time of text analysis. The speed is determined by the design features of a processor (by means of search time decrease), and also by the number of entities being extracted. The application of rules of extraction is connected with the search for the necessary words, where sorting is required. The greater is the number of entities and rules the greater is the time of analysis. In the processor LP there are different means of sorting time decrease.
Besides the program, there are also means of control by linguistic knowledge. It is indicated for each rule, what words should be searched for the initiation of the process of its application. The constraints in the form of the expected contexts (to the left and to the right of the revealed words) are assigned. These features ensure sufficiently high speed (fractions of a second for 1 KB of text) with a sufficiently large number of entities extracted.
The system features the entire complex of means which ensure rapid tuning to the applications (including the introduction of new entities and connections) taking into account the demands of customers. Note that in the mentioned processors the entities are brought to the standard form with the indication of the types of components. A sufficiently in-depth analysis of sentences is conducted with the development of verbal forms, and also with the identification of entities of the entire text. The analysis of complex language structures is ensured: forms with verbal nouns, participial and adverbial constructions, coordinated terms, etc. is supported by the expert component. The processor LP can be used as a stand-alone (independent) module. At present the first release of the English language version of the information semantic-oriented linguistic processor LP has been developed.