FI122369B - information - Google Patents

Description

Information Systems

Field of the Invention

The present invention relates to information technology and in particular to the sentiment analysis information system.

Background of the Invention

Most organizations need to monitor public opinion about the things they are involved with, whether it's products, reputation or image. For example, a company might be interested in whether people like the change in their product, or a campaign organization would like to know if 10 people have negative opinions about their candidate's ongoing divorce process.

Traditionally, this kind of monitoring is done either by asking a group of individuals directly about their opinions, or by looking through a collection of recordings on a particular subject. Each of these methods is costly, time consuming, and limited to a relatively small set of experiments because they involve human work. For example, in market research and risk management, automatic separation of positive or negative opinions, sentiments, would be preferable.

In addition, the explosive growth of online content such as weblogs and online forums has made it possible to access a vast source of public opinion information. However, the enormous amount of web content actually means that it is virtually impossible to credibly analyze web content using conventional methods.

The development of data processing tools has made it possible to produce devices in which the sentiments in different types of data units, such as recorders, HTML pages, etc., can be identified and stored autosmically. Automatic in this context means that a sequence of identification and storage functionalities can be implemented in an IT system. software, firmware, and / or hardware without the human mind being simultaneously involved.

The problem with existing information systems with respect to sentiment analysis o is that the results of the analyzes are not accurate and therefore

o

^ the results of queries based on the results are not informative enough to attract commercial interest, or do not justify the current costs of such queries. Locating the sentiment and analyzing the positivity or negation of the sentiment 2 is a relatively easy task for the human mind, but doing the same with an information system is nothing but trivial. Typically, the results obtained with an information system are very different from the results obtained through human analysis.

5 There are, of course, many challenges to the analysis of automated systems, but one of the major problems is that information systems are rarely sufficiently diverse to correctly interpret the positive and negative effects of different terms. Many conventional systems, such as text documents, rely on dictionary-based approaches 10 to list the relationships between terms and their corresponding polarity values. During the analysis, the polarity values of a recording or part thereof are retrieved from the dictionary and calculated using a predetermined equation. The resulting value is then used as the polarity value of the recording / recording portion.

15 However, such approaches neglect the fact that in many cases the same expression can convey many different meanings and sentiments, depending on where and how the expression is formed. When the human mind deals with an expression, such as a word, it functions holistically, taking into account the entire textual environment throughout the world. For example, with 20 text items, the term "small" conveys a different meaning depending on whether it's referring to a mobile phone ('Wow, my new Utopia phone is so small I can keep it in a jeans pocket!') Or room size ('The kitchen was so small even considering buying a house ').

BRIEF DESCRIPTION OF THE INVENTION It is therefore an object of the invention to provide an information system that provides more accurate automated sentiment analyzes. Accurate herein means that the results of the analyzes performed on the device better match the results obtained by combining manual analyzes made by one or more human minds. The object of the invention is achieved by a method and a system Characterized by what is stated in the independent claims.

Preferred embodiments of the invention are in the dependent claims

CM

the subject.

The invention is based on the idea of storing a plurality of comparison sentiment expressions, hints, each containing a sentiment expression and 35 sentiment context associated therewith. The device is arranged to detect a sentiment expression only when a sentiment term or set of terms occurs in 3 contexts corresponding to a given level of the hint context. The possibility of misinterpretations and associations in the device is significantly reduced and the accuracy of the device is improved.

An advantage of the method and apparatus according to the invention is that they enable rapid sentiment analysis and at the same time provide results that more accurately reflect the intent of the sources which gave the expression. The more accurate the results of the sentiment analysis are, the more valuable the results of the surveys for the recordings, and thus the overall performance of the sentiment analysis information system.

10 Enhanced accuracy also allows for a more efficient storage of reference sentiment expressions, as the interactive centimeter archive and tip-plus-context result approach human-like interpretations rather than overlapping them.

BRIEF DESCRIPTION OF THE DRAWINGS The invention will now be further described in connection with preferred embodiments, with reference to the accompanying drawings, in which:

Figure 1 provides a functional description of the information system;

Figure 2 shows the operating environment of the information system used in the embodiment of Figure 1; Figure 3 shows an exemplary device structure for implementing the device used in Figure 2;

Figure 4 shows the functional units of the acquisition section of the sentiment analysis information system of Figure 2;

Figure 5 illustrates a method implemented by the information system of Figure 4; 25 and Fig. 6 shows a self-controlling procedure that can be used as a c ...

^ vion 4 information system.

o cd

CM

x Detailed Description of the Invention

CL

It should be noted that the following embodiments are exemplary. Further, although the disclosure may at several points refer to "" some "," "one" g, or "" some "embodiment (s), the reference does not refer to the same embodiment (s), or the characteristic in question may not be applicable to a single embodiment. Individual features of the various direct embodiments may be combined to provide additional embodiments.

The invention relates to an information system which extracts certain information from input data units into a database. Several system architectures employing different data processing techniques may be used together or separately to implement embodiments of the invention. Data processing systems and techniques are constantly evolving and embodiments of the invention may require many changes which will, in principle, be apparent to those skilled in the art. Therefore, the words and expressions in this specification should be interpreted broadly, as they are intended only to illustrate, not to limit, the embodiments.

Figure 1 provides a functional description of an information system 100 according to the invention. Generally, the term information system relates to a combination of persons, records, and activities that automatically and manually process information and information within a particular organization. In this context, the information system 15 is more specifically related to the generic information technology component of the information system, and thus refers to any communication and / or computer related device or subsystem of interconnected devices that may be used to acquire, store, transmit, or receive information , micro-software, 20 and / or computing hardware.

Information system 100 takes a plurality of data units D1j, performs a processing function F (D1j), and outputs a plurality of records Sj (DI1). The information unit D1, refers to a block of data encoded in electronic form to enable computer hardware and computer programs in the information system to convert, store, protect, process, and transfer it and retrieve it securely in the processes of the information system 100. The data unit may be, for example, a text, video or audio file. Each data unit D1, comprises one or more data elements, ik, which can be automatically detected by the information system and processed as separate parts of the data unit. Automatically in this context ^ 30 means that the detection and processing functions can be performed in the information system.

X

System software, firmware, and / or computing hardware without the human mind simultaneously performing support operations.

S Some data elements in the data unit D1, can be associated with oo g identifiable subjects. The data element iekvoi refer to the topic yourself or 35 to provide a conception of the term party in relation to the topic. The party represents the understanding of sentiment, personal belief or assessment, which is not based on 5 display or reliability of the information and which is expressed in the unit used in the detection means. For example, in a text file, sentiments are expressed by verbal expressions of the language used in the text file.

The information system processing function F (D1j) is arranged to detect the data elements ik from the 5 data units entered, and to distinguish from the detected data elements the subjects and the sentiments associated with these subjects. The cent is processed to a polarity value that indicates the level of acceptance in the perceived sentiment of the subject. The processing function information F (D1j) is also arranged to create a record Sj (Dli) comprising the detected topic identification information and the sentiment polarity value associated with the detected topic. Information system 100 is arranged to provide a credit record Sj (D1j) to a database from which automated sentiment analysis can retrieve it.

The invention will now be described by means of an embodiment of an information system according to the invention. An embodiment of the invention uses the sen-15 sentiment analysis information system (SAIS), which is arranged to process data units encoded as a text file. A text file contains characters that can be interpreted as words, punctuation, and other special characters. Words are made up of character quences and can have semantic meaning in natural language. One or more 20 words may refer to the same semantic meaning, and one word may have one or more distinct interpretations. For information system processes, characters can be digitized and digital information can be converted back to characters.

Figure 2 illustrates the operating environment of the information system used as the embodiment of Figure 1. As described above, the information system receives from the data units from at least one data unit source. In Figure 2, SAIS is connected to a plurality of remote communication parties CP1, CP2, CP3, CP4 and a plurality of local-to-L communication parties CP5, CP6.

o

^ In order to communicate with remote communication parties, SAIS is connected to ^ 30 networked communication systems that enable the exchange of information between SAIS and

£ between the telecommunication parties CP1, CP2, CP3, CP4. To exchange information SAIS

and the telecommunication parties CP1, CP2, CP3, CP4 comprise terminals which S, together with the related software, applications and content, operate on the telecommunication network NET and are supported by the telecommunication network NET.

CM

The telecommunications network NET typically comprises at least one or more backbone networks providing network services and one or more access networks through which several different users can access these network services. The core network herein refers to a transmission medium consisting of networks, system hardware and infrastructures arranged to provide connections between access networks and between service providers and access networks. The access network represents the necessary transport bearer options for providing telecommunication services to the terminals of the telecommunication network. Here, the terminal represents a uniquely identifiable system unit that is the source and sink of data streams for one or more specific protocol layers. The basic elements of a telecommunications system are well known to those skilled in the art and will not be described in further detail herein.

In order to communicate with the local communication parties, SAIS is connected to the local interface LIF, or contains a local interface LIF, through which it can exchange information with CP5 and CP6. The LIF may comprise a user interface through which users have access to a limited set of SAIS Function-15 and. LIF may also include some non-network based communication capabilities that allow information to be exchanged directly between CP5, CP6 and SAIS communication devices. Examples of such mechanisms include Bluetooth, Infrared Data Association (IrDA), Radio Frequency Identification (RFID), Digital Enhanced Cordless Telecommunications (DECT), Dedicated Short Range Communications (DSRC), HIPERLAN, H1-PERMAN, Institute of Electrical and Electronics Engineers (IEEE) 802.11, Ultra-wideband (UWB), Institute of Electrical and Electronics Engineers (IEEE) 802.16 (WiMAX), xMax, ZigBee, Institute of Electrical and Electronics Engineers (IEEE) 802.16 (WPAN), Wireless USB.

25 As also described above, SAIS provides output processed records. For output functions, SAIS is connected to a database system DB, where it can store data. For simplicity, Fig. 2 shows a database system * L as one locally connected database. However, it is clear that

o

The DB database may also be integrated with SA1S., or it may be arranged over a communications network NET available to SAIS. Without departing from the scope of protection, a database system containing any combination of these various types of database elements may be used.

S Some communication parties may be sources of information and provide information units to SAIS for processing and storage. Some communication parties

CM

35 may be inquiry clients seeking information for sentiment analysis from the SAIS database system. Some communications parties may act as both sources of information and as inquiry customers. The object of the invention is to improve the acquisition side of the information system so that the results of the surveys more accurately reflect the human person's perception of the centimeters expressed in the information units without necessarily using the human mind in the process. It should be noted that although Figure 2 illustrates SAIS as a system that performs both procurement and litho operations, it is possible, without departing from the scope, to divide these parts into separate systems that share the same database system. The acquisition portion may be dedicated to extracting and storing sentiments from the data units, and the applicable portion may be dedicated to performing queries in a distributed database system. The acquisition part of an information system used as an embodiment will be discussed in more detail below.

Fig. 3 shows an exemplary device structure for implementing a device, the embodiment of which in Fig. 2 is SAIS. The device comprises a control unit 31, a processor element comprising an arithmetic logic module, a plurality of special registers and control circuits. The device also comprises a memory unit 32, a medium on which computer-readable data or programs or user data can be stored. The memory means typically comprise both memory modules, which allow both reading and writing (RAM), and memory modules, the contents of which are read-only (ROM). The device 20 also comprises an interface unit 33 having an input unit 34 for supplying data to internal processes within the device and an output unit 35 for printing data from internal processes within the device.

Examples of said pickup unit 34 in SAIS include staple units that serve as a gateway for information transported to its external access points. To locate information locally, the acquisition unit 34 may also comprise a keyboard or touch screen, a microphone or the like. To provide information locally, the delivery unit 35 may also comprise a display for μ. screen, microphone or similar.

The control unit 31, the memory unit 32 and the interface unit 33 are connected ^ 30 electrically to each other for received and / or stored data.

X

£ for systematic execution of operations according to predetermined, substantially programmed processes of the device. In the solutions of the invention, these functions include functions for implementing SAIS logical units, 00 g functions and interfaces as shown in Figures 4-6.

Units and blocks of a device may be implemented in one or more integrated circuits, such as application specific integrated circuits (ASICs). Other 8 embodiments of hardware are also useful, such as a circuit constructed from discrete logic components. A hybrid of these can also be applied. All modifications and structures for implementing the function of the embodiment may be performed as routines, which may be implemented as added or updated software routines, application circuits (ASICs) and / or programmable circuits. Software routines can be downloaded to the device.

It should be noted that Figure 3 only shows the elements required to represent the present embodiment. It will be obvious to a person skilled in the art that communication devices comprise a number of additional elements and functionalities 10 not explicitly described herein. In addition, the blocks represent logical and functional units that may be implemented within or in conjunction with one or more physical units, whether depicted in Figure 3 as one or more blocks.

The computer programs may be stored on a computer or processor 15 readable computer program distribution medium. The computer program medium may include, but is not limited to, an electrical, magnetic, optical, infrared or semiconductor system, device, or transmission medium. The computer program media contains at least one of the following media: computer readable media, program storage media, storage media 20, computer readable memory, random access memory, erase! programmable read only memory, computer readable software distribution package, computer readable signal, computer readable communication signal, computer readable print product, and computer readable compressed software package.

Figure 4 shows the SAIS of the sentiment analysis information system of Figure 2

functional units of the acquisition part. The data units are fed through various local or network interfaces to the data unit receiver DR. The information unit D1j is a combination of the elements iej, wherein the element iej can be considered at any level as an automatically separable part of the information unit D1j. DR retrieves the input data unit D1, and possibly separates it from the data unit levels

X

The metadata may be associated with the information unit D1, for example for use in subsequent steps of the procedure S and / or to be provided as additional information with the resulting sentiment records Sj (D1j).

C \] 35 In the present embodiment, the information unit is written in Text in Natural Language, for example English, and contains only textual information, not paternal media such as pictures or audio clips. A text record contains a set of expressions T. Expressions typically include words in their basic form t_i = (base), but they can also contain other natural language information, such as word class or t_i = (base, pos). For example, the words 5 'small', 'smaller', 'smallest' may be different expressions or they may all be described as t = ('small') or t = ('small', adj).

The DR enters the record detector separator DT of the information unit Dii, which analyzes the character sequences in the recording and determines a plurality of expressions therefrom. Expressions can be obtained automatically using natural language process-10 sounding methods, such as word class notation, object / subject ratio determination of verb and noun phrases. The distinction between phrases and text documents has been extensively discussed (http://www-nlp.stanford.edu/links/statnlp.html) in the field of tal analysis and is not described in detail here. Any mechanism suitable for distinguishing expressions may be used without departing from the scope of protection.

Let's say the recording contains the phrase The new Utopia phone is really small. However, it didn't have good pre-installed ring tones. I think I'll go buy it Tomorrow. In DT, the following set of expressions is distinguished from the phrase: 20 'the' 'new' 'Utopia' 'phone' 'be' 25 'really' 'small' t i 0

CM

'However'

Is- 9 'it'

CD

™ 30 'do' £ 'not', - 'have' 1 'good' o 'pre-installed' ow 35 'ring' 'tone' 10 c) τ 'think' Τ 5 'will' 'go' 'buy " 'it' 'Tomorrow' 10

To streamline the context analysis functions, DT may arrange the expressions of the original text record into the expression record D corresponding to the set D = {term}, where tenrij = (ti, i) is the i term of the record and t is the term term.

15 ('the', 0) ('new', 1) ('Utopia', 2) ('phone', 3) ('be', 4) 20 ('really', 5) ('small', 6 ) ΓΛ7) ('however', 8) ('it', 9) 25 ('do', 10) ('not', 11) o ('have', 12) | 4. ('Good', 13) o ^ ('pre-installed', 14) w 30 ('ring', 15) £ ('tone', 16)

Si <”17> S (T, 18) 00 § ('think', 19) w 35 (T, 20) ('will', 21) 11 ('go', 22) ('buy', 23) ( 'it', 24) ('Tomorrow', 25) 5 ('.', 26) DT enters the expression-separated record into the sentiment separator SE. Usually, SAIS stores a set of sentiment expressions and a polarity value associated with the sentiment expression. The SE detects at least one centimeter expression from the set of terms in the recording. The SE determines the polarity value corresponding to the sentiment term it detects. The SE also combines sentiment expression with a target audience comprising one or more target terms of the recording. The SE calculates the polarity value from the target term or the polarity value of the sentiment term combined with the target terms. The SE then provides the sentiment records, where the cent record contains at least the expression of the target term, the calculated Pola-15 runtime value of the target term, and, alternatively, some additional information that may be useful for the search and use phase operations.

According to the invention, sentiment expressions are further combined with context, and this context is used to determine whether the polarity value of the sentiment term can be used to calculate the polarity-20 of the target term or terms. In the following, one embodiment of the embodiment of the invention will be discussed in more detail. In an embodiment, the hints database enables context-based analysis.

Hint Database CDB is a functional unit that is available to the SAIS Sentiment Separator SE and contains several hints. 25 A hint basically comprises a reference, a reference context, and a reference polarity value. The shape of the hint elements follows the shape of the data elements o of the recording to be analyzed, so that the similarity between the hint elements and the set of one or more elements of the recording can be determined simultaneously for analysis by a human mind. In principle, a hint can comprise ^ 30 elements of any type of media, including audio and video. In the present embodiment related to textual recordings, the terms are based on textual sentences, and the clues consist primarily of expressions of adjectives, verbs, nouns, emoticons or exclamation points or exclamation points containing a positive or negative charge.

C \ J

Thus, to detect terms that are potential candidates for sediment, the SE reads the terms from the recording and selects a set of 12 from one or more terms. SE verifies that this term set matches the tips in the Hints Database. If a match is found between the term set and the clue, the terms are considered a sentiment candidate and the SE will take them for further analysis. If not, the terms are not considered to express sentiment and will be ignored by SE 5 at this stage. A hit check comprises selecting a candidate set consisting of one or more terms to be checked from an expression extracted record, selecting a matching hit criterion to use with a hint similarity function and a hint similarity function, and calculating the selected hint similarity function from the selected candidate set and. If the result matches the positive match criterion selected for the similarity function, a hint match is found.

It is generally assumed that the record D represents a data unit, a ceC hint, P (D) a set of all data element sets, and m: P (D) x C → R a similarity function between the data element set and the hint. The hint can be considered to match the record set {ie,} if m ({iei}, c)> um, where 5m is a predetermined threshold for the hint hit.

In the present embodiment, the hint similarity function can be implemented as a predetermined function that defines a set of terms from a set of terms and a set of expressions in a hint reference, and compares these determined sets. As a simple example, the similarity function may comprise a character-by-character comparison between the term expression and a hint reference expression, and give a Boolean value of 1 if all expression characters match or a Boolean value of 0 if at least one of the compared expression characters is different.

The hint similarity function can be selected separately for each of the 25 expressions or sets of expressions to be analyzed, or the same similarity function can be used in all analyzes. Without departing from the scope, other types of hint similarity functions and different sized recordings and hint detection sets can be used. For example, the hint similarity function need not (^) necessarily be a Boolean function; of different levels between the expression sets

(M

30 Similarities (ie hits) can be expressed using functions giving real numbers £. A hint match can be detected if the level of similarity has exceeded a predetermined threshold. Methods and functions for determining similarity between reference terms and target terms are described in more detail, for example, in PCT / FI2007 / 050511.

(M

35 When comparing the terms in a recording with those in the reference, typically, but not necessarily, it is required that the terms appear in the recording 13 in the same order as given in the reference. It may also be required that there be no other expressions between successive expressions. In many languages, this may still be too strict a requirement, and in order not to miss references due to irrelevant filler words, it is acceptable that there is a gap between the expressions. For this reason, it is possible to add a spacing value to the hints that defines the allowed size for the distances between hit expressions.

Thus, a hint can more formally be defined as a set of data elements that include a reference containing a looseness value sei and an artistry sequence (t, 0), a reference polarity value pe R, and a reference context o,. Tip size 10 c | is the length of the sequence (t, c). Hint c can be considered to match a set of terms in a recording if the reference expressions are detected in the term set expressions, preferably in the same order, and a limited distance apart. In this embodiment, the acceptable distance corresponds to the allowable number of irrelevant expressions between the expression of the candidate set 15 in the reference expression and the next expression of the candidate set in the next reference.

In other words, denote the hint c = (p, s, (tjc), Oj). Let DK = {(tj, i) | ieK} is a set of terms where Kc {0,1, ..., | D | -1}. Let kjK be the largest element, c, with D | <if | c | = | DK | and Export [0, | c | -1], 20 if - tki> ki + l - ki <s

Suppose, for example, that a hint comparison is performed on term sets that contain only one term, and the similarity function for the hint hit includes the Boolean function described above. Let's also assume that the hint database contains the following tips: 25 c_0 = (1,0, ("good"), {}) c_1 = (0,1, ("good", "morning"), {}) § c_2 = ( 1.0, ("small"), {'' phone '})

When the SE begins to search through the terms in the previous example record, the first term in cd that corresponds to the hint in the hint database is the term x 30 (“small”, 6). This term is now considered to be the candidate bearing the expression sentiment.

Conventional dictionary-based sentiment separation techniques suffer from the use of general-purpose dictionaries. Clearly, one and the same dictionary cannot be used for different types of recordings with different ways of expressing sentiment, and inaccurate or even misinterpretations can be made in the automatic production of sentiment records. To alleviate this problem 14, each hint further includes a definition of the context in which the hint applies. Similarly, for each term of a recording, the context in which it occurs can be determined. A hint can only be considered relevant to a group of one or more terms in a recording if the context of the term in the recording is sufficiently similar to the context of the hint. If the contexts are to an acceptable degree similar, a match is found and the hint is said to be related to the term. The related hint is suitable for producing sentiment records.

Let C be a set of contexts of all possible contexts, the context of the element set {e,}, the context of the CueC hint u, and the similarity function between the contexts r: CxC → R. The hint is related to the set of elements {e,} if r (C {£ i}, Cu)> 5r, where δΓ is a specified threshold for contextual similarity.

In the present embodiment, the hint contains a reference context Oj for context analysis which contains none, one or more context expressions, which hit terms comparable to those in the surrounding context region 15. Therefore, when a hint match between the expression and the expression or expressions in the hint database is found, SAIS determines the context area around the candidate's term or terms in the recording, and then checks for a match between the terms in the context and the reference context expression or expressions in the hint. The text expressions of the reference reference-20 tips are stored and maintained in the reference database, and in the present embodiment, from the viewpoint of the sentiment separation procedure, comprise predetermined sets of expressions. The context of the terms can be defined in many ways without departing from the scope. The following is a mechanism that utilizes a sequence of expressions in a detached recording.

In the detection-separated recording of the present embodiment, term D includes an indication of the order of its expression in the recording. This expression

C \ J

^ is used to define a region-specific term context. Thus, the terms comprising the terms of the recording can be considered as a sequence, and the term / c context comprises expressions up to k steps from the term. In other words, the ter-

| The min / c context comprises the terms contained in the detached recording D

(2k + 1) inside the window. For example, the term 0 context contains the expression of the term C 1 J itself, and the context 1 term is a set containing the term itself and the terms o immediately to its left and right.

^ 35 Generally, if D = {term} is a recording, the term term, k-context is (-'term = {tx | (tx, X) eD A hk <X <i + k}.

15

The k-context of a term set is a union of contexts.

In the present embodiment, the 3-context of the term ("small", 6) found to carry the sentiment expression is: C ("small", 6) = {"phone", "be", "really", "small", " "," If "} 5 In this embodiment, context similarity is computed using the basic cosine similarity f. If we denote the set of all expressions by T and P (T) the set of all expression sets, and give C-ieP (T) and C2sP (T) be expression contexts, expression context similarity f: P (T) x P (T) -> [0,1] can be defined as: f 1, if Ci = 0VC2 = 0 f (C'i, 02) = <ICiDCal otherume l V & isÄcTt 'ot > lermse 10

The equation indicates that the identity of an empty set with respect to all other contexts can be 1. This allows you to define clues that are sufficiently similar to any context. Such clues are called global clues. An example of a global tip is the expression "15" "good", which can be taken to mean a positive stance in any given context.

The cosine similarity in context 3 with C3 (-Smaii ", 6) with the hit tip c_2 = (1,0, (" small "), {" "phone"}) is:

Ismail-, 6V {V ™> "" = ^ 20 Assume in the system used in the embodiment, the threshold for context similarity is set to ó = 0 such that any value greater than zero verifies the similarity. This means that SE detects the term ("small", 6) to be related to hint c_2. This means that tips c_2 can be used to create sentiment records for terms that are within the scope of sa-S 25 racket ("small", 6).

Generally, sentiment refers to a targeted opinion whose scope extends beyond a set of elements in a recording. In this embodiment, the ^ sentiment may have a positive, neutral, or negative polarity which

X

£ represents a real number, with negative numbers representing negative cents and positive numbers positive cents. In this embodiment, the sentiment record comprises a set of data elements (k, a, t, p) where 00 δ keD indicates the effect range of the sentiment, the aeA sentiment factor, the teD sentiment object and the piR sentiment polarity. The sentiment indicates a negative sentiment if (iff), p <0, and the sentiment indicates a positive sentiment 16 if (iff) p> 0. Otherwise, the sentiment may be considered a neutral sentiment.

Usually, the recording is mostly filled with neutral sentiments, that is, statements that have no negative or positive payload. For Sen-5 Sentiment Records, the system aims to detect and record negative and positive sentiments from the large mass of neutral sentiment in the recording.

Thus, once the relationship between a recording term and a hint is determined, the sentiment associated with the term can be applied to one or more sets of 10 target terms that extend beyond the scope of the sentiment. To make this possible, the SE is combined with the sphere separator SPE. Once the SE has detected the sentiment based on the information in the hint database CDB, it enters the term into the scope delimiter, which defines the scope of the sentiment. Scope determination is highly dependent on the me-15 slide type and / or recording type used, and different scope determination methods can be applied without departing from the scope. In the embodiment system, the effect delimiter is arranged to see that the sentiment extends beyond the sentence of the text recording. When the SPE receives a term ("small", 6) from the SE, it defines terms that belong to the same sentence and returns a set of {("the, 0), (" new ", 1), (" Utopia " , 2), ("phone", 3), ("be", 4), ("really", 5), ("small", 6), (".", 7)}

It is clear that not all terms are relevant to sentiment records, and creating sentiment records for any affected term would be a waste of resources. Preferably, the SAIS comprises a target delimiter TE, which is provided with an object delimiter function that receives the set of SEs described above and returns a subset comprising only the target delimiter terms considered suitable for sentiment record queries. The target difference function may vary within the protection circuit. In the embodiment used, the TE is arranged to define and distinguish all substantive C / I vit / pronouns that precede the hit term in the domain of influence. Automatic definition of word categories is well known to those skilled in the art and will not be discussed further herein.

S In this embodiment, the word classes are defined by the TE, but can also be defined by DT. In the present embodiment, the TE returns the target terms ("Utopia", 2) and ("phone", 3) to 35 SE.

17

The sentiments expressed in the text may be generated by one person per recording. However, it is possible that the text contains multiple quotes and references that point to multiple opinion holders. For correct referencing at the query stage, SAIS may also comprise a functional element factor separator AE. The AE is provided with a factor extraction function that receives hit terms for the entire effect area and determines the factor indication to be included in the sentiment record. The factor separation function may vary within the protection circuit. In the embodied system, the AE using a simplified solution that searches ED from the received metadata to determine whether the Al-10-month invoking the communication partner (verkkojutteluidentiteetti '' loveMyPhone "), and use this identity to any of sentiment expressed by the record.

Finally, SAIS further comprises a polarity separator PE, which calculates a polarity value for sentiment records based on a related hint. The PE 15 is provided with a polarity difference function that receives the target terms and determines the associated polarity value to be included in the sentiment record. The polarity difference function may vary within the protective circuit. In its simplest form, the related polarity can be directly linked to the target term. Typically, however, it is necessary to check whether the scope 20 contains negation and whether it has an effect on the calculations for the target term. In the embodiment system, this is done by checking the PE for the 3 context context part preceding the hit sentiment term to contain the negation expression "not". If the expression is found, the reference polarity of the associated hint can be changed by multiplying by -1, otherwise the reference polarity value of the related hint 25 can be used directly. It is also possible to further improve the accuracy of the PE by interpreting the polarity decreasing as it moves away from the hit sentiment term. This deterioration can be accounted for, e.g.

CM

r-L is the polarity value of the target term corresponding to the reference polarity value of the associated hint, divided by the distance between the artistic expression and the expression of the target term in the range of 30. In the example, the reference polarity value for the associated hint c_2 is 1, and ί is the distance from the target term (“Utopia”, 2) to the hit term (“small”, 6), so the polarity value associated with the target (“Utopia”, 2) is 14. "Phone", 3) has 3 in the hit term ("small", 6), so the polarity value to be associated with the target term 00 g ("small", 6) is 1/3.

(M

35 The sentiment records generated by the first hit term are: 18 s_0 = ({('the ,, 0) l (' new ,, 1), ('Utopia') 2)> ('phone'> 3), ('be'I4 ), (, really'l5), ('smair> 6), ('., l 7)}, LoveMyPhone, {("Utopia", 2)}, 1/4) s_1 = ({('the ,, 0), (, new ', 1), (' Utopia ', 2'), ( 'phone', 3 '), (' be ', 4'), ( 'realiy', 5), ( 'smair, 6); ('.'> 5 7)}, LoveMyPhone, {("phone", 3)}, 1/3)

Sentence-carrying text expression can become negative or neutralized if the expression is part of a larger expression. This relationship of hints can be expressed in the form of sub-sequences. Let u and v be the exact sub-order for hints and <cue for all 10 hints. Hint u is a subtype of a hint v and hint v is a hint of u if (iff) u <cue v. An over hint is considered to be more detailed than its sub-tips. For more accurate results and more efficient handling, the system can be configured to use the most detailed hint that hits the element.

15 For example, the next relevant term for record D would be the term ("good", 13). Hint cue c_0 = ((1,0, ("good"), {}) now has a hint of c_1 = (0,1, ("good", "morning"), {}), and PE tries to get more accurate results first check the c_1 hit The following two terms ("pre-installed", 14) and ("ring", 15) do not contain expressions that would match c_1's other expression "" morning. " On the other hand, the hint c_0 clearly falls within the term "good" (13).

The SE used as an embodiment separates the following statement: {('however', 8), ('it', 9), ('do', 10), ('not', 11), ('have', 12), (' good ', 13), (' pre-installed ', 14), (' ring ', 15), (' tone ', 16), ('. ', 17)}, and TE separates the target term (' it ' , 9), which is the only noun / pronoun in La.

The related hint c_0 has a polarity of 1 and a distance of "it" from o "" good "of 4. PE detects the expression (" not ", 11), so the

CM

The polarity factor is -1/4.

° The resulting sediment thus becomes: ™ 30 s_2 = ({('how-ever', 8), ('it', 9), ('do', 10), ('not', 11), ('have ', 12), (' good ', 13), (' pre-install f ed ', 14), (' ring ', 15), (' tone ', 16), ('. ', 17)}, LoveMyPhone, {("it", 9)}, -1/4}

An alternative way to consider the expression "not" (11) would be

CM

^ also takes into account the location of the negative expression. The polarity value would then be o = -1 / (11-9) =. 1/2.

^ 35 No other relevant terms are found in the example recording of the present embodiment. Using a context-sensitive approach, the interpretation of the combination of the sentiment 19 and the target term within the scope of the sentiment approaches the flexibility of the human mind and makes the polarity values associated with the target terms more accurate for further analysis. The method is relatively simple and can therefore be implemented with the aid of general-purpose information systems.

Figure 5 illustrates an embodiment of the method used by SAIS of Figure 4 to create sentiment records. The details of the steps can also be checked from the description in Figure 4. The sequence begins when SAIS is up and running ready to take tips and information-10 ropes. A set of one or more clues is stored (step 500) in a reference database and SAIS goes into readiness (step 502) to receive new information units. The hint contains the reference, reference context, and reference polarity value. When a new information unit is received (step 504), its expressions are separated (step 506) into a record D comprising a set of terms D = {termj}. The relevance of the terms D-15 of the recording D to the recorded clues is checked (step 508). If (step 510) a hint match is not found, the information unit does not contain the sentiment and the procedure goes back to step 502 to wait for the next information item.

If (hint 510) a hint match is found, SAIS performs a relevance check where it determines (step 512) the context area around the hit term and 20 checks (step 514) the relevance of the term expressions and hint context expressions in the context area. If a relevance match is not found (step 516), the hint cannot be associated with any term in the affected area, and the procedure returns to step 508 to search for more hint hits. If (step 516) a re-levance match is found, the sentiment term is recorded (step 518) and the associated range of influence is determined (step 520). The target area is searched (step 522) for target terms. If (step 524) no target term is found, there is no single term to which the sentiment can be combined and the procedure returns. 508 to search for more clue hits. If (step 524) the target term is found, ^ SAIS determines (step 526) the sentiment factor, and calculates (step 528) the associated ™ 30 hint from the reference polarity value for the target term polarity value. Information about the range of action, the factor, the target term, and the calculated polarity value are stored (step 530) as data elements of the sentiment record and are given (step 532) to the sentiment database of S SAIS. The procedure then proceeds to step 522 to search for another target term for the sentiment record.

CM

35 As mentioned above, the reference database for the Sentiment Separation Procedure represents a static unit that provides a set of clues and clue context expressions that can be used in comparisons. The Hints Database, which enables context-conscious separation of expressions, provides a data structure in which both references and their associated contexts can also be easily retained with or without minimal parallel support of the human mind. In the previous embodiment, it was assumed that the expert had entered the cues into the database and stored them for further use in sentiment separation operations.

However, it is clear that for any full-scale system, a hints database based solely on man-made hints will not be technically operational. A more automated approach is made possible by separating clues from meaningful sentiment information. Assume that SAIS includes a CM additional functional element for tip management. It is further assumed that the expert enters a set of sentences in the CM that are marked either positive or negative. The sentences could be: Ί think Idoc is a very good MP3 player ', 15 The theater has a good sound system'

The new Swecar performed well in tests' and each of them would be marked positive. To detect references, CM includes a hint detection function and separates adjectives "" new "and" good "from the entered sentences and generates positive (polarity ^) global hint candidate dates with zero spacing. c_0 = (1,0, ("good"), {}) c_1 = (1,0, ("new"), {}) CM then determines the frequency of these Hint candidates in the tagged sentences. The adjective '' good 'has a frequency of 2/3 (sentences) and the adjective' 'new' has a frequency of 1/3. The CM can then either accept or reject the candidates based on their frequency.

The problem with an automated procedure is that the tips generated may not contain the context information needed to obtain more accurate results.

C \ J

However, the system according to the invention provides self-steering

£ mechanism with the existing hint database and sentiment separator SE

is used to generate new clues with context information. Figure 6 shows an S self-controlling procedure that can be used in the SAIS of Figure 3. In a self-ignoring procedure, the CM receives the recording and the existing

CM

35 and extract one or more sentiment records from the recording, as described above in Figures 4 and 5. In the example of Figure 6, the input 60 is 'Idoc 21 is not a good MP3 player because it is so small', and the hint database CDB comprises one hint c = (1,0, ("good"), {}). When using this hint as described above, SE will provide a sentiment record: s_0 = ({("ldoc", 0), ("nono", 1), ("be", 2), ("a", 3), (" good ", 4), (" mp3 ", 5), (" player ", 6), (" be 5 cause ", 7), (" it ", 8), (" be ", 9), (" so ", 10), (" small ", 11), (". ", 12)}, A, {(" ldoc ", 0), (" nono ", 1)}, 1/3) with target terms ("Ldoc", 0), ("nono", 1) and polarity 1/3. In addition to providing the sentiment record to the sentiment database, the sentiment s_0 is also passed to CM, which distinguishes between "good" and "small" 63, and create 10 of them as positive and negative candidate hints 64 ("good", 1), ("good", - 1), ("small", 1), ("small", - 1).

CM can be arranged to accept only the cue expressions that appear frequently in the texts and thus before accepting the expressions "good" and "small", CM checks to see if they occur frequently enough in the delivered recording. CM 15 complements candidate expression records 65 to include expression of the candidate's frequency in positive and negative sentiments ("good", 0/1, -1), ("" good ", 1/1, 1), (" "small", 0/1 , -1), ('' small ', 1/1, 1) Since the teaching material comprises a positive sentiment of one recording, positive candidates get a frequency of 1/1, since both adjectives appear once in a positive-20 sediment. zero, and negative candidates are thus eliminated, so the remaining hint candidates 66 are ('' good ', 1/1, 1) and (' small ', 1/1, 1).

CM will also check if the hint candidates are new. In the example of Figure 6, the expression for "candidate" ("good", 1/1, 1) already exists in the existing 25 clues, and thus the clue with the reference "good" does not need to be added to the clue database anymore. the expression '' small '' is therefore to be considered as a candidate expression for the purpose of 67 clues.

CM

^ Target terms for context analysis are defined from the area of influence of the sentiment record in the same way as in the separation of sentiment, e.g.

CQ

^ 30 why by recording all the nouns ("ldoc", "nono", "MP3", "" Player ") in the region of influence. These nouns are added to the hint as a set of reference context -r expressions. In fully automated cue production, the expression of the hint candidate 67 and the reference context expressions are compiled directly into a hint 68, which can be stored in the 00 o hint database CDB. The polarity value for the reference can be obtained directly from 3535 or by calculating the polarity value of the hint candidate. As an additional control, procedure 22 may include adding a step of checking the tip 68 by a specialist before storing it in the tip database CDB.

It will be obvious to a person skilled in the art that as technology advances, the basic idea of the invention can be implemented in many different ways. The invention and its embodiments are thus not limited to the examples described above, but may vary within the scope of the claims.

δ

CM

1 ^ o

CD

CM

X

cc

CL

CM

LO

00

o

o

CM

Claims (15)

An information system comprising: an expression separator arranged to receive a data unit and to distinguish therefrom a plurality of expressions; A tip database arranged to store one or more tips, the tip comprising a hint and a hint polarity value, the hint comprising one or more expressions; a sentiment separator arranged to separate from said expression a candidate expression set comprising about -1 one or more expressions; compare the candidate expression amount with the tips database tips; in response to detecting a match between the candidate expression amount and the tip, seek the hint's polarity value; use the hint polarity value when calculating a polarity value for a blend amount of the data unit one or more blend expressions; provide sentiment items, the item comprising the amount of melon and the polarity value calculated for the amount of melon, characterized in that the tip also includes a tip-specific implication context comprising one or more expressions; and the sentiment separator is arranged to: determine an area comprising the candidate expression amount and definite neighbor expressions for the expressions in the candidate expression amount; compare the expressions in the particular area with the hint's contextual context; 25, on the basis of the comparison, decide whether the implied polarity value should be used in calculating the polarity value for one or more of the data units in the target expression. O Information system according to claim 1, characterized in that the expression separator is arranged to determine the expression's position in the g data unit and to convey the expression to the sentiment separator as a term comprising the expression and an indication of the position of the expression; The sentiment separator is arranged to use the indication of the position of the expression in its separating, calculating and / or determining functions. 28 Information system according to claim 1 or 2, characterized in that the expression separator is arranged to determine the word class of the expression and to convey the expression to the sentiment separator as a term comprising the expression and an indication of the word class of the expression; The sentiment separator is arranged to use the indication of the word class of the phrase in its separating, calculating and / or determining functions. 4. Information system according to any one of claims 1-3, wherein η η βίο is characterized by the implication and the candidate expression amount being a successive amount of expression, and the implication further comprises a validity value which determines the allowed number of unexceptional expressions between the candidate's expression that matches an expression. in the hint and the candidate's next expression that hits the next expression in the hint. Information system according to any of claims 1-4, characterized in that the sentiment separator is arranged to compare expressions in the defined range with the hint context of the tip by using a similarity function, whereby the similarity function's determined result is linked to a positive hit or to a positive hit. negative hit. Information system according to claim 5, characterized in that the similarity function is cosine similarity obtained from f 1, if Ci = 0V (72 = 0 / (C1, C2) 177 - oi / ie 'where C1 is the implication context and C2 is the context formed by the expressions in the defined range. 7. Information system according to claim 5 or 6, characterized in that in response to the result of the similarity function relating to a positive hit, the sentiment separator is arranged to determine a scope for the amount of candida, the scope comprising a certain amount of expression which - x 30 adds to the candidate expression amount. 8. Information system according to claim 7, characterized by cvi that the scope comprises expressions belonging to the same set of candidate expression terms. 9. Information system according to claim 8, characterized in that the sentiment separator is arranged to separate at least one meal expression from the scope; and using the distance of the position of the template expression from the candidate expression amount when calculating the polarity value of the target expression. An information system according to claim 9, characterized in that the sentiment separator is arranged to check whether the scope contains a negative expression; and in response to detecting a negative expression to the extent of adjusting the polarity value of the template expression according to the detected negative expression. Information system according to any of claims 7-10, characterized in that the sentiment separator is arranged to separate from the scope a tip candidate amount of one or more expressions. Information system according to claim 11, characterized in that the sentiment separator is arranged to detect from the scope a plurality of one or more tip context expressions and to link the tip context expressions to the amount of tip candidate. Information system according to claim 12, characterized in that the sentiment separator is arranged to calculate a polarity value for the tip candidate amount and to save the tip candidate amount, the tip context expressions and the polarity value as a tip in the tip database, where the tip candidate amount is used as the basis for the subtitle expression, the tip text expression, the implied polarity of the calculated polarity values. A method comprising: operating an information system; a data unit is received and a plurality of expressions are separated therefrom; One or more tips are stored, the tip comprising a hint and a hint polarity value, the hint comprising one or more expressions; from multiple expressions, a candidate expression set comprising one or more expressions is separated; ^ the candidate expression amount is compared to the tips database tips; in CD 30 in response to detecting a match between the candidate expression amount x and the tip, the hint's polarity value is searched for; cc the designation polarity value is used in calculating a polarity value for a blend amount of the data unit's one or more blend expressions; 0 g of sentiment items are given, the item comprising the grind quantity and the polarity value calculated for the grind quantity, characterized in that a tip-specific hint context comprising one or more expressions is included in the tip; and an area is determined comprising a candidate expression set and determined neighbor expressions for the expressions in the candidate expression set; The expressions in the particular range are compared to the hint's context; On the basis of the comparison, it is decided whether the hint polarity value should be used in calculating the polarity value for one or more of the data unit's template expressions. 15. A computer program product that is readable on a computer and which encodes a computer program containing instructions for performing a computer process for controlling functions in an information system, characterized in that the process comprises: an information system is operated; A data unit is received and a plurality of expressions are separated therefrom; one or more tips are saved, the tip comprising a hint and a hint polarity value, the hint comprising one or more expressions; from multiple expressions a candidate expression set comprising one or more expressions is separated; The amount of candidate expression is compared with the tips database tips; in response to detecting a match between the candidate expression amount and the tip, the hint's polarity value is searched for; the hint polarity value is used in calculating a polarity value for a template amount of one or more template expressions of the data unit; 25 sentiment entries are given, the item comprising the amount of mel and the pole value calculated for the amount of mel, characterized in that -r- a tip-specific implication context comprising one or more expressions ^ is included in the tip; and a region is determined which comprises a candidate expression set and c in determined neighbor expressions for the expressions in the candidate expression set; The C 1 x expressions in the specified range are compared with the hint's contextual context; On the basis of the comparison it is decided whether the hint polarity value M "g is used in calculating the polarity value for one or more of the data unit's o expressions.