Introduction
This work is dedicated to the questions of creating the
engineering linguistic models of natural language for construction of
linguistic processors for different classes of information systems and to
description of the experience of the creation of linguistic ideas in the
systems, which relate to the artificial intelligence research field. In the
center of our attention are located the intellectual systems, developed on the
basis of the apparatus of the extended semantic networks (ESN) [Kuznetsov, 1986], [Kuznetsov et al., 2009a,b].
We call them ESN-systems. These systems were created by the association of
developers, including the authors of this article at the Institute of
Informatics problems of the Russian Academy of Sciences during the period of
two decades within the framework of research projects and applied systems,
oriented at the concrete subject areas and customers.
Intellectual ESN- systems contain the developed bases of
knowledge, in this case the knowledge is represented in the form of the records
in the language of the extended semantic networks, called ESN - structures.
Linguistic knowledge is, thus, a special case “of knowledge” and it is also
represented in the form of the records in the language of the extended semantic
networks. Basic structural element of the ESN is the named N-ary predicate,
called “fragment”. The whole set of language objects are given in the form of
predicate-argument structures, in this case the mechanisms for presentation of
embedded structures are supported, which gives very powerful presentation
mechanisms for describing the objects of different language levels.
In the process of analysis and synthesis of natural language
sentences the formal grammatical apparatus, similar to the dependency grammar,
is used. With this approach the words and the constructions, which perform the
role of predicates in the sentence, are the “support” elements, and the result
of the analysis of a sentence must become one predicate, which corresponds to
the predicate of the sentence (i.e. to basic verb in the tensed form or to
another basic predicate expression) in question. Thus, in the process of
analysis, in the first place, the processing is performed of the “action words”
and the “relation words”, i.e., of the verbs and other words, which have
syntactic-semantic valences. An example of a “relation word” the word “father”,
“friend”, and the like, i.e., in this case a “relation” is a word which assigns
strong clearly expressed syntactical-semantic expectations. Semantic analysis
in the engineering linguistic understanding is the process of translation of
natural language expressions into “internal” structures of the knowledge base
(KB) in our case these “internal” structures are the records in the ESN
language. Thus, a KB structure is the code of sense in the intellectual
information systems.
1. Basic
aspects of semantic modelling
The procedure of conceptual-linguistic simulation on the
basis of the ESN apparatus is based on the following principles: • a model must
be "open" , i.e., support the effective mechanism of expansion and
information update; • the model of the “sense” presentation should consider the
facts of extralinguistic reality, which in the form of rules and relations
compose a certain basic "world model"
and the concrete models of subject areas; • the model should be
practical, i.e., not overloaded by the detailed descriptions of connections and
relations between the concepts in order to ensure the possibility of its
realization, but at the same time, it should reflect the relevant information
for specific objectives.
┌──────────────────────────────┐
┌─────┤1. The analysis of the
texts │
│
└──────────────────────────────┘
│
│
┌──────────────────────────────┐
└────>┤2. Singling out the basic
con-│
┌─────┤ cepts and characteristics │
│
└──────────────────────────────┘
│
│
┌──────────────────────────────┐
└────>┤3. Constructing a subject
area│
│ vocabulary founded on the │
┌─────┤ basic “world model” │
│
└────────────────┬─────────────┘
│
┌───────────┴──────────────┐
│ │The basic “world model” │
│ │ and the language model │
│
└──────────────────────────┘
│
│
┌─────────────────────────────────┐
└────>┤4.
Establishment of the type-kind│
┌─────┤relations
between SA notions │
│
└─────────────────────────────────┘
│
┌─────────────────────────────────┐
└────>┤5.Formulation of the
situational │
│ rules in the form of IF…
THEN │
│ rules │
└─────────────────────────────────┘
Figure 1 – The flowchart of conceptual
linguistic modeling.
Realistic approach to the formulation of the problem
dictates the need of limitation to a domain-oriented subset of a natural
language. The essence of limitations consists in the following: - first,
analyzed text materials contain expert knowledge from particular subject areas
(we developed the systems for the subject areas for the diagnostics of the microcircuits
production failures, forecast in the social sphere, criminology, and others); -
in the second place, for the purposes of the maximally possible elimination of
ambiguity, dictionary is built according to the modular principle: there is a
certain most general common part (1-2 levels) completed by special dictionaries
for each particular subject area. The proposed model of lexical semantics is
based on the principle of the "nuclear" value realized in the context
of this subject area with the subsequent inductive supplementation of other
meanings (if they are actualized in the contexts in question). The taxonomy is
also used which is realized in the form of the hierarchical trees of the word
classes. The general "world model" of the system serves as the basis
for the subject area models. The classes of words, are subdivided into
concept/names, relations, actions, properties, characteristics of actions, time
and place locatives. The most general notion is “concept”, or universal class,
which is subdivided into object, the situation, process and others. The words
which relate to the classes of actions and relations, are represented as the
semantic-syntactic frames, which determine the predicate-argument structures
(government model). However, in the described approach (let us name it the ESN-approach) the range of argument values is
substantially extended. This extension consists in the fact that in the role of
arguments there can appear simple objects corresponding to the individual
words, structural objects which present word combinations, phrases and clauses,
and concept of "case" includes not only semantic, but also syntactic
aspects. Approach, based on ESN allows to reflect the arbitrary level of the
structures embedding it makes it possible to reflect the structural nature of
lexical semantics, which in this model has a hierarchical network structure.
Linguistic knowledge is represented in the system dictionary and the
declarative modules of linguistic processor. In the ESN systems the function of
dynamically formed semantic dictionary which is expanded automatically by the
system in the course of concrete texts processing is also realized on the basis
of initial linguistic information. In Fig. 2 the “internal” description of the
verb in the semantic dictionary is represented. This dictionary is
automatically generated by the ESN-systems DIES2, LOGOS-D, IKS in the course of
natural language texts processing.
{(ВЫРАБАТЫВА895__)(DICSEM)
COORD(PROGNOZ1,RUS,ВЫРАБАТЫВА895__,S50_31_51_20,%)
SUB(UNIV,0+) SUB(UNIV,1+) SUB(UNIV,2+)
ВЫРАБАТЫВ(0-,1-,2-/3+)
INFI(3-) ПРИДЕТСЯ(3-) ПРИДЕТСЯ(3-/4+) FUT1(4-) SUB(СРЕД,5+)
Figure 2 -
An example of the presentation of the verb vyrobatyvat’
- “to manufacture” in the semantic dictionary.
2. The
linguistic processor for entities extraction
The
linguistic processor (LP) is realized by means of the language DEKL and is
controlled by the linguistic knowledge (LK) in the form of object dictionaries,
means of parametric tuning, and also the rules of extracting objects and
connections [Kuznetsov, Kozerenko 2003]. With the aid of LK the tuning of LP to
the appropriate categories of users and text corpora is accomplished. Concrete
realization appears as a result. Thus, the paper deals with the means of
constructing a class of processors with powerful mechanisms for their tuning
and updating. Further development of such processors (LP) is connected with the
development of implicit information, which we will consider in a narrow plan,
i.e. as the addition of the structures of knowledge by the new information,
which is absent or assigned implicitly. In this article the procedure of this
development is proposed, which consists in the use of LP for mapping NL texts
onto the structures of knowledge (ESN) and the use of the means of
logical-analytical processing (productions of the language DEKL) for the creation
of new information.
Advantages
and deficiencies of the proposed procedure will be examined on specific
objectives from the area of “criminology”, that is the role functions
establishment for the persons (participants) on the basis of the acts performed
by them or due to the participation in some specific events. We consider the
problem of assignment of properties to the persons (basing on their
participation in the acts of different kinds) - “the suffered”, “the suspect”
and others, if an explicit description of such properties is absent from the
text. For example, if it is said in the text “suffered Ivanov I.I.”, then
another task appears, i.e. extraction of some property in the process of
linguistic analysis and forming of the corresponding fragments in the knowledge
structure. In this article the discussion will deal with LP, customized for the
Russian language texts (NL), although the possibilities of LP are wider. There
is a sufficient test of tuning LP to the English language texts.
3. The
design method
The task
of the role functions establishment for
the information objects is a special case of the more general task, connected
with the estimation of objects according to their descriptions in the NL texts,
for example, with the estimation of the stability of enterprise (according to
the information from the Internet), by featuring political figures (positive or
negative depending on the statements in the press), by the estimation of the
role functions quality of product (basing on the statements of users) and so
forth. Quite frequently, it is not said directly whether something is bad, or good. As a rule, in NL texts the
events are described, the situations, in which one or other information object participated.
On the basis of them the estimation is done, which is often represented in the
form of a new (generated) property of object.
For the
solution of this problem different methods are used [Banko et al., 2007], [Clark et al., 2007], [Gildea, Palmer,
2002], [Pasca, Van Durme, 2007], [Punyakanok et al., 2008],
[Cunningham, 2005], [Han, 2004], [Han, Kamber, 2006] . The most common one is the method of
the new properties of objects development by using the syntactical-semantic
forms. For example:
<what-medicine>
caused allergy in <who-human
organism>…,
< what-medicine > has side effects … <who-person>
made scandal… and so forth.
The application of such forms to the NL texts consists in
the search for “estimating” or “characterizing” words (of type “scandal”) or
for word combinations of the type “caused allergy” (“it can cause allergy”), it
"has side effects” (“side-line actions”), “to make scandal” (“to brawl”)…
And then the environment is analyzed, i.e., the words, which stand to the left
and to the right, their semantic classes (objects are recognized by them) and
case forms. Estimations of information objects as a result are given. By the
first two forms the “quality of medicines” is estimated, while by the latter it
is recognized that a man performed “hooligan actions” or that he is
“suspected”. It is known that in NL many versions are possible for expressing
the same idea - with the aid of different syntactic constructions, verbal
groups, forms and so forth.
Therefore the number of estimating word combinations will be
sufficiently large. Moreover, the application of such forms requires different
forms of analysis - morphological (in order to reduce different word forms to
one form), syntactic (the trees are built of the selection of sentences in
order to isolate the connected components and to find place for the estimated
words) and semantic (in order to extract the objects, which are evaluated). The
use of syntactical-semantic forms is connected with certain difficulties caused
by special NL features: by the presence in texts of participial,
verbal-adverbial constructions, different explanations, facultative components
(time, place, purpose), anaphoric references and other language structures. As
a result, information objects are frequently disconnected from the estimated
words. Hence - the significant losses, which influence the quality of
estimation.
Example 1 (the text is taken from the summaries of incidents of the City Office
of Home Affairs, Moscow):
… Gorelov Peter Sergeyevich,
01.03.76 yr/bir, liv: c. Moscow, st. Young Leninists, h.71-6-12, does not work,
01.02.1998 yr. at 4.30 in
his house out of hooligan motives in the
state of alcoholic intoxication made scandal and broke window glass in the
apartment of Litvinova Galina Ivanovna, 20.07.1961 yr/bir,…
In this example the estimating (characteristic) words are
"made scandal” and "broke the window glass”, they are located at a
significant distance from the estimated person - “Gorelov Peter Sergeyevich”.
This limits the possibilities of applying the forms. It is required that the
initial extraction of components, which must not be considered in the forms:
the years of birth, addresses, specific properties (“he does not work”, “in the
state of alcoholic intoxication”), time, place and others, which requires sufficiently
deep text analysis with the extraction of objects, their properties and
attributes. In connection with the aforesaid, another more promising method is
represented - when evaluation is accomplished at the level of knowledge
structures. For their construction the objective-oriented LP is used producing
the structures of knowledge in which the objects are directly connected with
the events and the actions and excluding the above mentioned losses. For the
development of implicit information (role functions of objects) the rules of
the DEKL language are used which analyze the structures of knowledge (ESN) and
form new properties of objects. In this case the structure of knowledge does
not change, but it is only supplemented by new (useful) fragments.
4. Presentations of the meaningful portraits of documents
Within
the framework of the proposed procedure the development of the role functions
of objects (implicit information) is achieved at the level of the structures of
knowledge, called the meaningful portraits of documents (SS-documents). Let us
examine how such structures appear in the ESN formalism .
Example 2
(translation of the Russian text given below). Text N22 is taken from the summaries of incidents of the City Office of
Home Affairs, Moscow:
01.02.98 yr. 16-30 to the Home Office
applied citizen Mitrofanov Victor Mikhaylovich, 1955 yr. bir., liv.: Bohr
Highway 38-211, n/w. he stated that 01.02.98 yr. at 10-00 in house 3 at St. Fedosino
the unknowns being found in the drunk state made scandal, they expressed
themselves by unquotable swearing, they set dog. As a result of what Mitrofanov
applied to trauma care center, where the diagnosis was set: the bite of
foot.
The
objective-oriented LP performs the deep analysis of the text and automatically
builds its meaningful portrait (SS- document, transliterated):
DOC_(22,
“1-02-98”,
“SUMMARY; ” /0+) 0 (RUS)
OVD_(OVD/1+)
FIO(MITROFANOV],
VICTOR, MIKHAYLOVICH, 1955/2+) UNEMPLOYED (2-/3+) 3- (22, PROP)
ADR_(Borovskiy,
Sh., 38,211/4+)
PROZH.
(it is 2nd, 4)
ADR_(UL,
FEDOSINO, HOUSE, 3/5+)
FIO
(" " , " " , " " , NESKOLKO/6+)
UNKNOWN
(6)
DRUNK
(6-/7+) 7 (2, PROP_)
SCANDAL
(6, PYANYY/8+)
IS EIGHTH
(22, ACT_)
TO REPORT
(IT IS 2ND, 8-/9+) 9 (22, ACT_)
DATA_(1998,02,
~01, " 10-00" /10+)
When (9,
10)
TO TURN
(1, GR- N, 2-/11+) 11- (22, ACT_)
DATA_(1998,02,
~01, " 16-30" /12+)
When
(11-, 12-)
EXPRESS
(6, UNQUOTABLE, [BRAN]/13+) 13- (22, ACT_)
TO SET
(6, [SOBAKA]/14+) 14 (0, ACT_)
TO TURN
(IT IS 2ND, IN, [TRAVMPUNKT]/14+) 14 (0, ACT_)
TO PLACE
(DIAGNOSIS, BITE, [NOGA]/16+) 16 (0, ACT_)
PREDL_(22,11-,
4, 3-, 9, 13-, 14-/17+) 17- (2,15,341)
PREDL_(22,15-, 16-/18+) 18- (6,342,448)
A meaningful portrait consists of the
elementary fragments, arguments of which are words in the normal form
(necessarily for the search and processing). Each elementary fragment has its
unique code, which is written in the form of the number with the sign + and is
separated by a slash line. For example, in the fragment OVD_(OVD/1+) the sign
1+ is its code (but 1 is the reference
to it). Fragments DOK_(22, “1-02-98.TXT”, “SUMMARY; ” /0+) 0 (RUS) indicate
that the meaningful portrait is built on the basis of the Russian-language text
of document with number 22 of the file of 1-02-98.TXT”, which was processed as
the summary of the incidents (linguistic knowledge depend on this). The
following fragments present police department OVD_(… /1+), person’s surname,
name and patronymic FIO (… /2+), person’s specific property UNEMPLOYED (2-/3+),
address ADR_ (… /4+) and so forth; the signs 2+, it is 2nd, 3+, 3-,… are the
codes of the fragments, with the aid of which their connections and relations
are assigned. For example, the fragment PROZH (live) (it is 2nd, 4) represents
the relation that the person (represented as FIO with code 2+) lives at the
address (fragment [ADR]_ with code 4+). Actions are represented in the form of
fragments of the type SCANDAL (6, PYANYY/8+) it is 8 (22, ACT_), where it is
represented that “person (FIO with code 6+), being drunk, made scandal”. With
the aid of it is the fragment 8_(22, ACT_) indicates that the first fragment is
SCANDAL (…./8+) presents the action and relates to the document with the number
22. A
similar role is played by the fragments of the type 3-(22, PROP_), by which the
properties are noted. The codes of fragments also serve for the idea of time,
scene of action and cases, when one action is included in the composition of
another. For example, the fragment TO REPORT (it is 2nd, 8-/9+) represents that
the person (code 2+) “reported” (code 9+) about the action (code 8+), i.e.,
about “made scandal”. The following fragments DATA]_(… /10+) when (9, 10)
represent the time (DATA_), which relates (when) to the action “to report”.
Special role is played by the fragments PREDL_(...), which correspond to the
sentences. They are filled up with the words, which did not enter the
information objects (in this example they are absent), or with the codes of
objects themselves.
To these
fragments the indicators of their position in the text are added. For example,
the fragment PREDL_(22,11-, 3-, 9, 13-, 14-/17+) 17- (2,15,341) represents the
fact that the objects with codes 11- (corresponding to the action “to turn”),
3- (corresponding to the property “unemployed”) and others are located in the
sentence, which begins from the 2nd line of the text of the document and they
occupy the place from the 15-th to the 310-th byte. These means of positioning
are necessary for the work of the reverse linguistic processor (LP).
Analyzing this example, it is possible to make the following
conclusions: 1) In SS- document the estimating (characterizing) words occur
either in one fragment with the object - SCANDAL (,…), or the next one, i.e.,
the codes of the actions, in which the object participates, are nearby in
PREDL_(… 9, 13-, 14…). In this case the possibility of composite actions is
considered. 2) On the actions, represented as SCANDAL (,…), it is possible to draw
the conclusion that the discussion deals with “that suspected”, and TO REPORT
(,) - that the person is “suffered” or “the applicant”. Such conclusions are
easily arrived at with the aid of the rules IF… THEN (productions) of the
language DEKL, which are the basis for the extraction of role functions. 3) The
particular difficulties of dividing the text into the sentences occur (in the
old version).
The reduction “of n/r” (with the point at the end) was not
understood as the end of a sentence. 4) The linguistic processor (LP) correctly
identified the pronoun “he”, and also
it knew how to reveal the participation of the subject (“unknowns”) by the actions “to
be evinced by unquotable swearing” and “to
set dog”, which also characterize subject. At the same time the LP could
not connect the action “diagnosis was set”
with the person - “Mitrofanov…” (the
code is 2-nd). In this case an example proved to be successful. Also the
processor LP (with its linguistic knowledge - LK) was developed for the tasks
of the criminal police, connected with different forms of the objective searches:
the search for similar participants (addresses, and so forth), search according
to the connections, precise search for objects, for the search by signs and
other identifiers. In this case the analysis of some complex NL forms was not
required, i.e. the cases of the enumeration of the objects participating in the
uniform actions (they are described by one verb), the enumeration of the
actions of one object and others in contrast to the aforesaid, with the
extraction of role functions for each object the indication of its
participation in each action is required. Hence it follows that with the use of
the proposed procedure the more qualitative extraction of role functions is
directly connected with the works on improvement of LP in the aspect of the
development of objects and their actions. In many instances the numerous errors
caused the inaccuracies in SS-document, e.g.: the absence of punctuation marks
or their presence (where it was not required), the inappropriate reductions,
gaps in the words and many others. The fact is that the documents, entering the
summaries of incidents, are composed on the spot by people (militiamen) of
different degree of literacy. Hence – the additional noise and loss. Thus,
meaningful portraits are the collections of fragments of ESN which represent
the sufficiently high level of formalization of NL texts and are convenient for
the working - with the aid of the instrument means - DEKL. Besides LP which
analyzes texts and builds SS-documents, there is a reverse linguistic processor
(LP) which on the basis of the fragments of the SS- document generates the NL
texts presented to the user.
5. The means for the establishment of the role functions
Within the framework of the proposed
procedure (instead of the application of syntactical-semantic forms to the
documents) the rules are used for logical conclusion and transformation of the
knowledge structures - the SS- documents, in which there are no morphological
features (of type who, whom,…), and the subjects and the objects are
distinguished by their arrangement in the fragments of ESN, which present
actions. The names of fragments present the nature of actions.
Syntactical-semantic forms are transformed into the fragments of ESN which
determine conversions and logical conclusion achieved by productions of
language DEKL. Such fragments play the role of the logical-semantic shell,
which determines conversions and logical conclusion on the basis of
SS-documents. After filling of the shell by ontological-fragmental knowledge (OFK)
which consist of the mentioned fragments (ESN), the program is formed, which
accomplishes the development of role functions and completion of the
SS-document by the appropriate fragments. With this approach it is possible to
avoid many difficulties, connected with the design features of NL and the
specific character of the use of syntactical-semantic forms. There are many
versions of construction of the shells and representation of the corresponding
knowledge which are distinguished according to the degree of their generality.
Let us examine the version which is at present realized and verified.
Case 1. The role functions are
determined by the names of actions. In this case for the extraction of objects
(participants) which should be assigned properties (role functions), the
fragments of the following form are used :
INTERPRET (MAN_2, FIO, "
suffered") FORMA_CC (MAN_2, CLASS_D4, " ") CLASS_D4 (TO TURN, TO
STATE, TO REPORT, TO PASS AWAY,…)
The first fragment INTERPRET (…)
means that from the SS- document it is necessary to extract the fragments of
the form FIO (…), that correspond to participants, and to analyze the
possibility of assigning them the property "suffered". Such
participants are conditionally designated as MAN_2. The second fragment
FORMA_CC (…) specifies the conditions for assigning this property to MAN_2,
determined by the constant CLASS_D4. In the third fragment CLASS_D3 (…) the
words are given which present actions. It is represented that the words belong
to the class CLASS_D3. If the participant occurs in one of the enumerated
actions, then to this participant the property "suffered" is
assigned. This participation is revealed via the analysis of the SS-document.
If there is a fragment TO TURN (…, it is n-th,…) in it, the argument of which
is the code FIO (… /N+), then the fragment N-("suffered") is added
that represents the role function of the corresponding participant. Conformably
for the SS- document represented in example 2 the analysis will occur as
follows. Consecutively the extraction of fragments FIO (…) corresponding to the
participants is performed. First FIO (MITROFANOV,… /2+) will be extracted . Its
code is 2- is the argument of the fragment TO APPLY (1, GR- N, 2-/11+), that
presents the action. In connection with this to SS- document the fragment 11-
("suffered") will be added, which via the reverse LP will be
transformed into the statement that “Mitrofanov Victor Mikhaylovich is a
suffered person”. These actions are realized within the framework of the
logical-linguistic shell.
Case 2. Role functions are
determined by the actions and elucidating words. For this the same fragments are
used, as in the first case, but during the enumeration of the names of actions
the additional fragments which present actions with the possible elucidating
words, are introduced:
INTERPRET (MAN_1, FIO, “suspect”)
FORMA_CC (MAN_1, CLASS_D3, "
")
FRAUD (USER, POKUPATEL/15+)
TO SET (DOG/16+)
TO BE EXPRESSED (UNQUOTABLE,
SWEARING, OBSCENE,… /17+)
CLASS_D3 (IS DELAYED, TO BE
SOUGHT,…, 15,16-, 17-)
The given fragments determine
actions of the extraction of persons (MAN_1), by which the property of
“suspect” is assigned. For this at the level of the knowledge structures their
participation is analyzed in the actions “is
delayed”, “to be sought”, and also
in the composite actions: “to set dog”,
“to be expressed unquotable…”, “to be expressed by swearing…” and
others. In example 2 the code of fragment FIO (" " , " " ,
" " , FEW]6+), that represents the unknown persons is the argument of
the fragment TO SET (6, DOG/14+), representing action “to set” with the elucidating word “dog” – “sobaka”. Therefore the fragment 6 is added ("suspect"), that represents that “the unknown persons are suspected”, and
through the reverse LP the explanation to this conclusion is offered, see
below. A similar conclusion will be made on the basis of the fragment TO BE
EXPRESSED (6, UNQUOTABLE, SWEAR/13+), but with other explanations.
Case 3. The actions determine the
role functions of several persons. For this (additionally to the fragments
INTERPRET) the fragments are added: CLASS_D1 (TO STRIKE, TO BEAT UP,…) FORMA_CC
(MAN_1, CLASS_D1, MAN_2), where FORMA_CC (…) indicates the need of the search
of two persons - "suspect"
and "suffered" (MAN_1 and
MAN_2), that participate in one action, which are mentioned in the fragment
CLASS_D1 (…). For example, “certain
person struck another…”. In the appropriate fragment TO STRIKE (…) the code
FIO (…) that corresponds to the first person will stand in front of the second.
The given fragments of ESN compose the knowledge OFK which are constantly
supplemented - due to the filling of classes by the new words-actions and with
the elucidating words. The process of filling is sufficiently simple. If role
function is not revealed, then it is necessary to look in the SS- document in
which the action of one or another participant (by the text its role is easily
determined) occurs. Further, the corresponding constants are located, by which
are supplemented the classes of knowledge OFK. Subsequently it is intended to
automate the process of completing the knowledge OFK as follows. In the text
the words, which determine role functions, are noted. Further, in the formed
SS-document the corresponding constants which supplement knowledge OFK are
located.
6. Factors of processor quality
The quality of a linguistic processor is determined by a
number of factors. First, the possibility for isolation of objects and
connections. These are the types of objects being isolated, their quantity. The
Semantix processor identifies up to 40 types of objects, including very complex
ones, which correspond to actions and events. With an increase in the quantity
appear the additional difficulties, connected with collisions of the extraction
rules of: some rules can seize the words, which relate to other objects and
those extracted by other rules. It becomes important to consider the order of
the application of rules, including of the rules of identification. In the
second place, an important factor is the selectivity of rules and procedures of
the identification: the factor of the noise and losses. By noise we mean the
presence of excessive words in the objects. Losses are the situations when an
object is not revealed or revealed partially: in the text there are the words,
which did not enter into the object. In the Semantix processor 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 objects 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 objects), and also tuning to the new objects. Due to the
complexity of analysis this tuning should be achieved through the linguistic
knowledge (LK). The latter should have all means for increasing the selectivity
of rules and necessary conveniences in the plan of their creation and
correction. Ideally, with the aid of LK
the tuning to the special features of language as well as to the standard
language forms should be ensured. The Semantix linguistic processor ensures the
analysis of the Russian and English lanuage 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 objects being extracted. The application of rules of extraction
is connected with the search for the necessary words, where sortings are
required. The greater the number of objects and rules, the greater the time of
analysis. In the Semantix processor 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 permissible contexts (to the
left and to the right of 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 objects extracted. More than 40
different types of objects are supported by the Semantix processor. The
subject areas represented in the text documents are as follows.
·
Documents about terrorism in the
Russian language. The analysis of the documents, in which the discussion deals
with the terrorist acts and the groups. This feature supports the extraction of
40 types of objects, their connections and the degree of participation in the
criminal actions.
·
Documents about terrorists in the
English language. The objects and links include persons (their family name,
name, patronymic – FNP), posts, organizations, terrorist groups, instruments of
crime, time and place of events and so forth, and also connection with and
participation in the actions.
·
Summaries of incidents. Is
ensured the extraction of figurants, their connections, organizations, dates,
documents, numbers of bank accounts, details of weapons, etc. with the
indication of their participation in particular criminal actions.
·
Accusatory conclusions,
information about the criminal cases. Objects are identified along the entire
field of text. Their connections and criminal actions are revealed.
·
Government communications, media
issues. Persons, dates, organizations, positions and other significant
information and also connections and participation in the actions are selected.
·
Autobiographies in the Russian
language. From the Russian language resumes all attributes of people, periods
of time and place of their work, studies, language proficiency and so forth are
extracted.
·
Autobiographies in the English. From the
English language resumes are all attributes of people, periods of time and
place of their work, studies, language proficiency and so forth are extracted.
·
Documents of media issues in English. From the English
language texts the persons mentioned in media issues, positions, organizations,
dates, terrorist and anti-terrorist groups, weapons, events, their time and
place, different connections and other features are extracted.
In the processors of the Semantix, Lingua-Master, “Criminal”
systems up to 40 types of objects are extracted with high accuracy and minimum
noise. For example, the system "Criminal" was verified on about 500
thousand incidents from the summaries of Moscow Criminal Police Department, and
on the basic objects showed the unique results: the coefficient of noise, i.e.
excessive words in the objects) is not more than 1-2% and losses are not more
than 1%. The Semantix Processor was fixed on a smaller quantity of documents
dealing with the terrorist activity, and therefore there can be more noise and
losses in it. But this can be quickly fixed. The fact is that to consider
everything which can be encountered in the NL texts is impossible. Therefore,
in the first place, the representative collections of test documents are
extremely important, and in the second place, the means of fixing or tuning of
linguistic processors are as follows: the employment of hybrid approaches
comprising hand-made rules and statistical means for rapid correction and fine
adjustment of linguistic knowledge.
In our systems there is an entire complex of such means
which ensure rapid tuning to the applications (including the introduction of
new objects and connections) taking into account the demands of customers. A sufficiently in-depth analysis of sentences
is conducted with the development of verbal forms, and also with the
identification of objects of the entire text. The analysis of the complex
language structures is ensured: forms with
verbal nouns, participial and adverbial constructions, coordinated
terms, etc. is supported by the expert component. The Semantix processor can be
used as a stand-alone (independent) module. At present the first release of the
English language version of the object - oriented linguistic processor Semantix
has been developed.
Conclusion
The objective-oriented linguistic processor for entities
extraction can be used in different areas of application. It has a number of
essential advantages, the main is that it outperforms the recently appeared
systems by the number of the supported object types, there is an entire complex
of logical and statistical means which ensure rapid tuning to the applications
(including the introduction of new objects and connections) taking into account
the demands of customers. Our furter efforts are connected with the developing
a simultaneous bilingual search and extraction features.
The procedure of the role
functions extraction centered at the analysis of knowledge structures is
sufficiently promising from the point of view of the knowledge bases technology
development. The current task is to improve its performance.
The methods were
tested on the basis of the summaries of incidents which contain about three
thousand documents (each document consists of 10 - 80 lines). In the case of
the summaries processing the documents with the mentioned enumerations (there
were about 10% of them) were withdrawn and in the remained texts the gaps in
the words were removed. At the current moment the program which realizes the
proposed procedure gave about 80% of correct recognition of role functions, and
about 65% of complete explanations with the indication of all acts. But these
numbers rapidly change for the better due to the means (the LK and OFK
knowledge) of tuning the LP to special features of the subject area texts. For
this not much time is required. Let us note that tuning itself to the
extraction of the role functions of persons from the mentioned summaries (with
reaching the indicated percentages), required about two weeks of the work of
one person. The development and fixing of the shell itself took about four
days. The subsequent development is connected with the improvement and the
tuning of LP to the work with complex NL forms. At present the extraction of
actions is interfered with causal word combinations of the type “out of the hooligan motives”, “owing to the hostile relations” and so
forth, which at present are introduced into the system. Difficulties appear
with the transfer of the subject of action to other actions to which the
subject is not assigned explicitly, but its presence is implied.
The second direction of research and development is
connected with the extension of the shell features to the solution of other
problems connected with the estimation of objects depending on the nature of
statements about them in the texts of description. Within the framework of the
studies conducted it is also intended to tune the shell to the work with the
English language texts. Since the meaningful portraits of the English language
and Russian language texts have the identical structure (SS-documents), this
tuning cannot be labor-consuming.
Bibliography List
[Kuznetsov, 1986] Kuznetsov, I.P. Semanticheskie Predstavleniia. Moscow: Nauka,
1986, 290 p.
[Kuznetsov et al.,
2009] Kuznetsov I.P., Efimov D.A., Kozerenko E.B. Tools for Tuning the Semantix Processor to Application Areas // Proceedings of ICAI'09, Vol. I. WORLDCOMP'09, July 13-16, 2009, Las Vegas, Nevada, USA. - CRSEA Press, USA, 2009. P.
467-472.
[Kuznetsov et al., 2009] Kuznetsov I.P., Kozerenko E.B., Kuznetsov K.I.,
Timonina N.O. Intelligent System for Entities Extraction (ISEE) from
Natural Language Texts // Proceedings of the International Workshop on Conceptual Structures for Extracting Natural Language Semantics -
Sense'09, Uta Priss, Galia Angelova
(Eds.), at the 17 International Conference on Conceptual Structures (ICCS'09),
University Higher School of Economics, Moscow, Russia, 2009. P. 17-25.
[Kuznetsov, Kozerenko, 2003] Kuznetsov, I.P.,
Kozerenko T.B. The system for extracting semantic information from natural
language texts // Proceeding of International Conference on Machine Learning.
MLMTA-03, Las Vegas US, 23-26 June 2003 г., p. 75-80.
[Banko et al., 2007] Banko M., M. Cafarella, S. Soderland, M. Broadhead, and O.
Etzioni. Open Information Extraction from the Web // Proceedings of the 20th
International Joint Conference on Artificial Intelligence (IJCAI-07), 2007. P. 2670–2676.
[Clark et al., 2007] Clark P., P. Harrison, and J. Thompson. A Knowledge-Driven
Approach to Text Meaning Processing // Proceedings of the HLT-NAACL 2003
Workshop on Text Meaning, 2007. P. 1–6.
[Gildea, Palmer, 2002] Gildea D. and M. Palmer. The
necessity of syntactic parsing for predicate argument recognition. In
Proceedings of the 40th Annual Conference of the Association for Computational
Linguistics (ACL-02), Philadelphia, PA, 2002. P. 239–246.
[Pasca, Van Durme, 2007] Pa¸sca M. and B. Van Durme.
What You Seek is What You Get:
Extraction of Class Attributes from Query Logs // Proceedings of the 20th
International Joint Conference on Artificial Intelligence (IJCAI-07), 2007. P.
2832–2837.
[Punyakanok et al., 2008] Punyakanok V., D. Roth, and W. tau Yih. The Importance of
Syntactic Parsing and Inference in Semantic Role Labeling // Computational
Linguistics 34(2), 2008. P. 257–287.
[Cunningham, 2005] Cunningham H. Automatic
Information Extraction // Encyclopedia of Language and Linguistics, 2cnd ed.
Elsevier, 2005.
[Han, 2004] Han J., Pei Y. Yin, and Mao R. Mining
Frequent Patterns without Candidate Generation: A Frequent-Pattern Tree
Approach,” // Data Mining and Knowledge
Discovery, 8(1), 2004. P. 53–87.
[Han, Kamber, 2006]Han J. and Kamber, M. Data Mining: Concepts and Techniques //
Morgan
Интеллектуальные системы извлечения сущностей,
основанные на расширенных семантических
сетях
Кузнецов И.П., Шарнин М.М., Козеренко Е.Б., Мацкевич А.Г., Николаев В.Г., Сомин Н.В.
Учреждение Российской академии наук
Институт проблем информатики РАН, Москва, Российская Федерация
igor-kuz@mtu-net.ru
В данной работе приведены модель,
механизмы и реализации интеллектуальных систем извлечения
знаний - сущностей и их связей - из текстов на естественном языке. В основе проектных
решений лежит аппарат расширенных семантических сетей. Инструментом создания
систем служит оригинальных язык логического программирования – ДЕКЛ.
Введение
Дается описание
семантико-ориентированного лингвистического процессора, извлекающего из текстов
естественного языка информационные объекты, их свойства и связи и формирующего
на этой основе структуры знаний. Одно из направлений развития таких процессоров
связано с выявлением имплицитной информации, которая рассматривается в узком
плане - как выявление новых свойств объектов, заданных в неявном виде.
Предлагается методика такого выявления, основанная на анализе структур знаний.
В качестве примера рассматривается выявление ролевых функций фигурантов на базе
их описаний в сводках происшествий.
Основная часть
Научное направление, связанное с
обработкой произвольных текстов ЕЯ (в заданной предметной области) развивается с учетом задач определенной
категории пользователей: большинство пользователей – специалистов в своей
области - интересуются лишь конкретными вещами для своих задач. Например,
следователям важны фигуранты, их места жительства, телефоны, приметы, действия
лиц (с указанием места, времени), связи и др. Такая
информация является основой для оперативно-розыскных действий, поиска по
связям, различных видов анализа. Специалистов по кадрам интересуют организации,
где человек работал, кем и когда это было. Подобную информацию будем называть сущностями, информационными объектами, которые различаются по типам. Например,
лица и фигуранты – это объекты одного типа, адреса – другого и т.п.
Заключение
Анализ методики проводился с использованием сводки происшествий, содержащей
около трех тысяч документов (каждый документ содержит от 10 до 80 строк). Предлагаемая
методика, основанная на анализе структур знаний, представляется перспективной для
развития технологии баз знаний.
