JP2013525868A - System and method for determining sentiment expressed in a document - Google Patents

Abstract

A system for determining sentiment expressed in a document, a computer readable storage medium storing instructions therefor, and a computer-implemented method therefor are disclosed. A document is received from a plurality of documents. A sentence is identified in the document that includes at least one sentiment signature within a predetermined distance of at least one keyword from the keyword list, wherein the keyword list is extracted from multiple documents and phase transitions Filtered using an expression, at least one sentiment signature corresponds to a representation of at least one sentiment in the sentence. At least one category corresponding to at least one keyword of the sentence is determined that is included in the list of categories generated using the list of keywords. Based on the at least one sentiment signature, at least one sentiment corresponding to the at least one category is determined.
[Selection] Figure 1

Description

(Cross-reference of related applications)
This application claims priority to US patent application Ser. No. 12 / 977,513, filed Dec. 23, 2010, entitled “Systems and Methods for Determining Sentiments Expressed in Documents”. Insist. This application further claims priority to US Provisional Patent Application No. 61 / 284,820, filed December 24, 2009 under the name “Sentiment Platform”. This application further claims priority to US Provisional Patent Application No. 61 / 284,819, filed December 24, 2009 under the name “Mobile Sentiment Platform”. This application further claims priority to US Provisional Patent Application No. 61 / 393,813 filed on October 15, 2010 under the name “Sentiment Engine”. Each of the applications is incorporated herein by reference in its entirety.
The disclosed embodiments generally relate to determining sentiment expressed in a document.

  The Internet contains information on various subjects. This information is written by a specialist in a particular field or a temporary user (eg, blogger, reviewer, etc.). Search engines allow users to identify documents that contain information about subjects of interest to them. However, it is currently difficult to identify sentiments represented by these users on a particular subject.

US patent application Ser. No. 12 / 977,513 US provisional patent application 61 / 284,820 US Provisional Patent Application No. 61 / 284,819 US Provisional Patent Application No. 61 / 393,813

  A system for determining sentiment expressed in a document, a computer readable storage medium storing instructions therefor, and a computer-implemented method therefor are disclosed. A document is received from a plurality of documents. A sentence is identified in the document that includes at least one sentiment signature within a predetermined distance of at least one keyword from the keyword list, wherein the keyword list is extracted from multiple documents and phase transitions Filtered using an expression, at least one sentiment signature corresponds to a representation of at least one sentiment in the sentence. At least one category corresponding to at least one keyword of the sentence is determined that is included in the list of categories generated using the list of keywords. Based on the at least one sentiment signature, at least one sentiment corresponding to the at least one category is determined.

  Like reference numerals refer to corresponding parts throughout the drawings.

1 is a block diagram illustrating a network according to some embodiments. FIG. FIG. 2 is a block diagram illustrating a sentiment server according to some embodiments. 2 is a flowchart of a method for determining sentiment expressed in a document, according to some embodiments. 2 is a flowchart of a method for extracting a list of keywords, according to some embodiments. 2 is a flowchart of a method for generating a list of categories according to some embodiments. 6 is a flowchart of another method for generating a list of categories, according to some embodiments. 6 is a flowchart of another method for determining at least one category corresponding to at least one keyword of a sentence, according to some embodiments. 2 is a flowchart of a method for determining a category spectrum for a category, according to some embodiments. 2 is a flowchart of a method for selecting a plurality of documents from a collection of documents, according to some embodiments. 1 is a block diagram of a machine according to some embodiments. FIG.

  The description that follows includes exemplary systems, methods, techniques, instruction sequences, and processing machine program products that embody embodiments for purposes of illustration. In the following description, for the purposes of explanation, numerous specific details are set forth in order to provide a thorough understanding of various embodiments of the inventive subject matter. Nevertheless, it will be apparent to one skilled in the art that embodiments of the inventive subject matter may be practiced without these specific details. In general, well-known instruction instances, protocols, structures, and techniques have not been shown in detail.

  The embodiments described herein provide a technique for determining sentiment expressed in a document.

  FIG. 1 is a block diagram illustrating a network 120 according to some embodiments. The network 120 can generally include some type of wired or wireless communication channel that can couple processing nodes together. This includes, but is not limited to, a local area network, a wide area network, or a combination of networks. In some embodiments, the network 120 includes the Internet.

  In some embodiments, a server 100 is coupled to the network 120. Server 100 includes a document 102. Document 102 includes, but is not limited to, web documents (eg, hypertext markup language (HTML) documents and extensible markup language (XML) documents), text documents, spreadsheets, presentations, scanned documents. It can be any type of document, including (for example, scanned text), images, and the like.

  In some embodiments, an aggregator 104 is coupled to the network 120. Aggregator 104 includes a document 106. In some embodiments, aggregator 104 obtains at least a subset of documents 102 from server 100. For example, the aggregator 104 crawls the server 100 and retrieves at least a subset of the document 102 from the server 100.

  In some embodiments, sentiment server 108 is coupled to network 120. The sentiment server 108 is configured to determine sentiment expressed in the document, as described herein. Documents used by the sentiment server 108 include documents obtained from the server 100 (eg, by crawling the server 100), obtained from the aggregator 104 (eg, by purchasing documents from the aggregator 104). Documents, or combinations thereof, are included.

  FIG. 2 is a block diagram illustrating a sentiment server 108 according to some embodiments. The sentiment server 108 is configured to crawl the server 100 to obtain at least a subset of the document 102, and a crawling configured to determine the sentiment represented in the document. A module 204 (optional), a keyword module 206 configured to extract keywords from the document, a filtering module 208 configured to filter keywords and documents, documents, sentences, and / or keywords And a classification module 210 configured to classify. Note that the functionality of these modules may be combined. For example, the sentiment module 202 may include the functionality of a keyword module 206 and a filtering module 208. These modules are described in more detail below in connection with FIGS.

Determining Sentiments Expressed in a Document FIG. 3 is a flowchart of a method 300 for determining sentiments expressed in a document, according to some embodiments. Sentiment module 202 receives a document from a plurality of documents (302). For example, the plurality of documents includes at least a subset of the document 102, at least a subset of the document 106, or a combination thereof. The process of selecting multiple documents is described in further detail below in connection with FIG.

  Sentiment module 202 then identifies (304) sentences in the document that contain at least one sentiment signature within a predetermined distance of at least one keyword from the keyword list. At least one sentiment signature corresponds to a representation of at least one sentiment in the sentence. In some embodiments, the at least one sentiment signature includes at least one word (eg, an adjective) that indicates that an expression of the at least one sentiment is present in the sentence. In some embodiments, the sentiment signature has an associated polarity. For example, the polarity indicates whether the sentiment signature reflects a positive sentiment, a negative sentiment, or a neutral sentiment.

  Note that the sentiment module 202 can also identify probabilistic units that are larger or smaller than sentences. For example, the sentiment module 202 may identify a paragraph or phrase that includes at least one sentiment signature within a predetermined distance of at least one keyword.

  In some embodiments, the list of keywords is extracted from multiple documents and filtered using a phase transition equation. These embodiments are described in further detail below in connection with FIG.

  In some embodiments, the at least one sentiment signature is included in a list of sentiment signatures. The list of sentiment signatures may be manually generated.

  The classification module 210 then determines 306 at least one category corresponding to at least one keyword of the sentence. In some embodiments, at least one category is associated with a product, service, or combination thereof. The process for determining at least one category corresponding to at least one keyword of a sentence is described in more detail below in connection with FIGS. In some embodiments, at least one category is included in a list of categories that has been generated using a list of keywords. These embodiments are described in more detail below in connection with FIGS.

  The sentiment module 202 then determines 308 at least one sentiment corresponding to the at least one category based on the at least one sentiment signature. In some embodiments, the at least one sentiment is a representation of an opinion related to at least one category.

  In clarifying the process described in connection with Figure 3, an example document containing the example sentence "The room was stinky and the carpets were dirty." Let's consider. The words “stinky” and “dirty” are sentiment signatures representing negative sentiment (eg negative polarity), the words “room” and “carpets” Is a keyword, and the predetermined distance is 3. The sentiment module 202 clarifies this exemplary sentence as if the sentiment signature “stinky” is two words away from the keyword “room” and the sentiment signature “dirty” is the keyword “carpets”. (304) because it is two words away from (carpet). The classification module 210 then determines a category corresponding to the sentence keyword (306). In this example, the classification module 210 determines that “hotel room” is the category for the sentence keyword. The sentiment module 202 then determines that the sentiment expressed for the hotel room (eg, category) is a negative sentiment (308).

In some embodiments for extracting a list of keywords, identifying a sentence in a document that includes at least one sentiment signature within a predetermined distance of at least one keyword from the list of keywords (304). Prior to this, the keyword module 206 extracts a list of keywords from a plurality of documents. FIG. 4 is a flowchart of a method 400 for extracting a list of keywords, according to some embodiments.

  The keyword module 206 extracts keywords from each of the plurality of documents (402).

であり、ここにｘ≧１である。幾つかの実施形態では、ｘは３である。キーワードモジュール２０６は、次に、キーワードの関連度が所定の閾値を超えていたら、当該キーワードをキーワードのリストに加える（４０８）。 For each keyword, the keyword module 206 performs the following operations. The keyword module 206 calculates the frequency f of keywords in a plurality of documents and the number N of documents containing the keywords (404). Next, the keyword module 206 calculates the degree of relevance of the keyword based on the frequency of the keywords in the plurality of documents and the number of documents including the keyword using the phase transition formula (406). . In some embodiments, the phase transition equation is

Where x ≧ 1. In some embodiments, x is 3. Next, the keyword module 206 adds the keyword to the keyword list if the degree of relevance of the keyword exceeds a predetermined threshold (408).

In some embodiments for generating a list of categories, the classification module 210 generates a list of categories prior to determining 306 at least one category corresponding to at least one keyword of the sentence.

  FIG. 5 is a flowchart of a method 500 for generating a list of categories according to some embodiments. The classification module 210 identifies a first set of documents that include at least one keyword from a list of keywords among the plurality of documents (502). Next, the classification module 210 identifies a set of keywords contained in at least a predetermined number of documents in the first set of documents (504). The classification module 210 then adds the set of keywords to a list of categories, each category including each keyword set (506). Thus, in these embodiments, the classification module 210 determines a category by identifying keywords that appear in at least a predetermined number of documents.

  FIG. 6 is a flowchart of a method 600 for generating a list of categories, according to some embodiments. The classification module 210 obtains a keyword pair that is a keyword pair that is related to each other in the keyword list and that is a unique keyword pair (602). Next, the classification module 210 identifies a set of keyword pairs, each set including at least one keyword that is common to all of the keyword pairs in the set ( 604). Until a predetermined termination condition is achieved, the classification module 210 iteratively combines sets of keyword pairs, and each combined set includes all keyword pairs in the combined set. At least one keyword that is common to each other is included, and such combinations are made (606). Thus, in these embodiments, the classification module 210 determines a set of keywords that are related to each other and repetitively combines pairs (or larger keyword groups) to form a category. For example, the classification engine 210 may generate the following keyword pairs from a list of keywords: {Paris, Romance}, {Paris, City of Love}, {Paris, France}, {Dog, Beagle}, {Cat, Siam} Identify The classification engine 210 then selects {Paris, Romance, City of Love} because the word "Paris" is common to the {Paris, Romance}, {Paris, City of Love}, {Paris, France} pairs. , France} is a relevant set of keywords (eg, category). Note that the classification engine 210 may determine that {Paris, Romance, City of Love} is a set of relevant keywords. The number of keywords associated with a particular category depends on several factors, including but not limited to the desired amount of specificity (eg, including 4 keywords). The category is more specific than the category including three keywords), the number of documents associated with the specific category, the number of sentences associated with the specific category, and the like. In some embodiments, the predetermined termination condition depends on the degree of specificity desired for the category. The more keywords used to describe a category, the more specific the category is (eg {Paris, Romance, City of Love, France} is {Paris, Romance, City of Love} More specific).

Several techniques can be used to determine the category corresponding to the keyword that determines the category corresponding to the keyword.

  In some embodiments, the classification module 210 determines (306) at least one category corresponding to at least one keyword of the sentence using a support vector machine.

  In some embodiments, the classification module 210 determines (306) at least one category corresponding to at least one keyword of the sentence using a neural network.

  FIG. 7 is a flowchart of a method 700 for determining at least one category corresponding to at least one keyword of a sentence, according to some embodiments. The classification module 210 obtains a category spectrum that includes a plurality of category spectra, each of which includes the frequency of occurrence of a keyword in a keyword list corresponding to each category (702).

であり、ここにｉは単語ＩＤに対応しており、頻出度_1iは第１のカテゴリについての単語ＩＤ「ｉ」に対応するキーワードの出現頻度であり、頻出度_2iは第２のカテゴリについての単語ＩＤ「ｉ」に対応するキーワードの出現頻度である）。 Next, the classification module 210 determines a category spectrum for the sentence based on the at least one keyword (704). In some embodiments, the classification module 210 normalizes the sentence category spectrum. The classification module 210 then calculates (706) the dot product of the category spectrum of the sentence and each category spectrum among the plurality of category spectra. The classification module 210 then determines 708 at least one category as a category corresponding to at least one dot product that exceeds a predetermined threshold. The category spectrum can be represented by a pair of {word ID, frequency}, the value of the word ID corresponds to a unique keyword, and the frequency of occurrence corresponds to the appearance frequency of the keyword corresponding to the word ID. Please pay attention to. For example, it is assumed that the keyword “Paris” has a word ID of 8 and the appearance frequency is 1002. Then, the category spectrum includes {8,1002} pairs. Note further that the category spectrum can also be represented visually. For example, on the 2D plot, the x-axis can be the word ID and the y-axis can be the frequency. Furthermore, it should be noted that the dot product of the two category spectra is the sum of the product of the appearance frequencies of the keywords for each word ID (for example,

Here, i corresponds to the word ID, the frequency _1i is the appearance frequency of the keyword corresponding to the word ID “i” for the first category , and the frequency _2i is the frequency for the second category. The frequency of occurrence of the keyword corresponding to the word ID “i”).

  In some embodiments, prior to obtaining a plurality of category spectra, the classification module 210 determines a category spectrum for each category. FIG. 8 is a flowchart of a method 800 for determining a category spectrum for a category, according to some embodiments. The classification module 210 obtains a collection of documents corresponding to the category (802). Next, the classification module 210 extracts keywords from each document in the document collection (804). The classification module 210 then filters the keywords using the phase transition equation to derive filtered keywords (806). Next, the classification module 210 determines the frequency of appearance of the filtered keywords in the document collection (808). In some embodiments, the frequency of occurrence of each keyword is divided by the total number of counts that the keyword was found in the entire aggregate across all categories. In some embodiments, a threshold is then applied to the resulting category spectrum. In these embodiments, if the amplitude of the total keyword count in a category divided by the total keyword count across all categories is greater than a predetermined threshold, the amplitude is reset to a new value by an equation or algorithm. Is done. In some embodiments, a threshold is then applied to the resulting spectrum. In these embodiments, if the amplitude obtained by dividing the keyword count in a certain category by the total page count of pages in which the keyword appears in all categories is larger than a predetermined threshold, the amplitude Is reset to a new value by an expression or algorithm. The new value is set to zero to eliminate common words that are very commonly used across all categories.

  The classification module 210 then normalizes the frequency of appearance of the filtered keywords to derive a category spectrum for the category (810). In some embodiments, the category spectrum is normalized such that the area under the spectrum of each category spectrum is the same. By doing so, the comparison bias between categories is reduced.

であり、ここに、ｘはｎ−グラムのサイズであり、ａ、ｂ、及びｃは、所定の方程式のピークがサイズ２のｎ−グラムとサイズ３のｎ−グラムの間となる所定の値である。幾つかの実施形態では、ｂの値は１と２の間であり、ｃの値は１と２の間である。フィルタリングモジュール２０８は、文書のｎ−グラムスペクトルが基準ｎ−グラムスペクトルに所定の閾値内で整合している場合は、当該文書を前記複数の文書に加える（９０８）。フィルタリングモジュール２０８は、文書のｎ−グラムスペクトルが基準ｎ−グラムスペクトルに所定の閾値内で整合していない場合は、当該文書を捨てる（９１０）。 The document 102 and / or document 106 from which the document is selected may include documents that have little or no value. For example, document 102 and / or document 106 may include a machine-generated document to enhance search engine ranking. Machine-generated documents typically do not contain a representation of sentiment. It is often desirable to filter out these types of documents. Thus, in some embodiments, prior to receiving documents from a plurality of documents, the filtering module 208 selects a plurality of documents from a collection of documents. FIG. 9 is a flowchart of a method 900 for selecting a plurality of documents from a collection of documents, according to some embodiments. For each document in the document collection, the filtering module 208 performs the following operations. Filtering module 208 extracts n-grams from the document (902). Next, the filtering module 208 determines an n-gram spectrum for the document based on the extracted n-grams, which indicates the frequency of occurrence of the n-grams as a function of the size of the n-grams (904). Filtering module 208 then determines whether the n-gram spectrum of the document is matched within a predetermined threshold to a reference n-gram spectrum defined by a predetermined equation (906). In some embodiments, the predetermined equation is:

Where x is the size of n-grams, a, b, and c are predetermined values where the peak of a given equation is between an n-gram of size 2 and an n-gram of size 3 It is. In some embodiments, the value of b is between 1 and 2 and the value of c is between 1 and 2. Filtering module 208 adds the document to the plurality of documents if the n-gram spectrum of the document matches the reference n-gram spectrum within a predetermined threshold (908). If the n-gram spectrum of the document does not match the reference n-gram spectrum within a predetermined threshold, the filtering module 208 discards the document (910).

Exemplary Computer System FIG. 10 is a machine , by way of example in the form of a computer system 1000, within which a machine may perform any one or more of the methodologies discussed herein. Figure 2 depicts a block diagram of a machine on which a set of instructions is executed. In alternative embodiments, the machine may operate as an independent device or may be connected (networked) to other machines. In some networked deployments, the machine may be operating in a server-client network environment with server or client machine qualification, or in a peer-to-peer (or distributed) network environment. It may be working as.

  A machine can execute a set of instructions (sequentially or otherwise) that specify actions to be taken by the machine. Further, although only one machine is shown, the term “machine”, individually or conjointly, refers to an instruction (single or combined) to perform any one or more of the methodologies discussed herein. It should be construed to include some collection of machines that perform the set (s).

  The computer system 1000 of this example includes a processor 1002 (eg, a central processing unit (CPU), a graphics processing unit (GPU), or both) and a memory 1004 that are in communication with each other via a bus 1008. Memory 1004 includes volatile memory devices (eg, DRAM, SRAM, DDR RAM, or other volatile solid state memory devices), non-volatile memory devices (eg, magnetic disk memory devices, optical disk memory devices, flash memory devices, Tape drives, or other non-volatile solid state memory devices), or combinations thereof. Memory 1004 may optionally include one or more storage devices located remotely from computer system 1000. The computer system 1000 can further include a video display unit 1006 (eg, a plasma display, a liquid crystal display (LCD), or a cathode ray tube (CRT). The computer system 1000 can further include an input device 1010 (eg, a keyboard, a mouse, etc.). , Trackball, touch screen display, etc.), output device 1012 (eg, speaker), and network interface device 1016. The above-described components of computer system 1000 include a single housing or case (dotted line in FIG. 10). Alternatively, a subset of the components may be placed outside the housing, eg, the video display unit 1006 and the input device 101. An output device 1012 is present in the outside of the housing, it may be connected to bus 1008 via an external port or connector can outside access housing.

  Memory 1004 embodies any one or more of the methodologies or functions described herein, or data structures and instructions 1022 (eg, utilized by any one or more of the methodologies or functions). A machine-readable medium 1020 on which one or more sets of software) are stored. One or more sets of data structures store data. Note that machine-readable media refers to storage media that can be read by a machine (eg, computer-readable storage media). The data structures and instructions 1022 may also be wholly or at least partially resident in the memory 1004 and / or the processor 1002 executing by the computer system 1000, and the memory 1004 and the processor 1002 are also machine-readable tangible. It is a component of the medium.

  Data structures and instructions 1022 may also be sent and received via network 120 and via network interface 1016 using a number of well-known transfer protocols (eg, Hypertext Transfer Protocol (HTTP)). Network 120 may generally include some type of wired or wireless communication channel that can couple together processing nodes (eg, computer system 1000). This includes, but is not limited to, a local area network, a wide area network, or a combination of networks. In some embodiments, the network 120 includes the Internet.

  Particular embodiments are described herein as including logic or a number of components, modules, or mechanisms. A module may comprise either a software module (eg, code and / or instructions embodied on a machine-readable medium or in a transmitted signal) or a hardware module. A hardware module is a tangible unit capable of performing a specific operation, and can be configured or arranged in a specific manner. In an exemplary embodiment, one or more computer systems (eg, computer system 1000) or one or more hardware modules (eg, processor 1002 or group of processors) of a computer system are software (eg, The application portion) is configured as a hardware module that operates to perform the specific operations described herein.

  In various embodiments, the hardware module can be implemented mechanically or electronically. For example, a hardware module is dedicated that is configured permanently (eg, as a special purpose processor such as a field programmable gate array (FPGA) or application specific integrated circuit (ASIC)) to perform a specific operation. The circuit configuration or logic may be provided. A hardware module is further included within programmable logic or circuitry (eg, a general purpose processor 1002 or other programmable processor) that is provisionally configured by software to perform a specific operation. Thing) may be provided. The decision to implement a hardware module mechanically in a dedicated and permanent configuration or in a temporary configuration (eg, configured by software) is a cost and time consideration. It will be understood that it will be driven by matters.

  Thus, the term “hardware” is a tangible entity that is physically constructed to operate in a particular manner and / or perform the specific operations described herein. It should be understood that it covers entities that are permanently configured (eg, wired) or provisionally configured (eg, programmed). Consider an embodiment in which hardware modules are provisionally configured (eg programmed), where each piece of hardware must be configured or instantiated at any time. It doesn't have to be. For example, when the hardware module includes a general-purpose processor 1002 configured using software, the general-purpose processor 1002 may be configured as different hardware modules at different times. Thus, the software can configure the processor 1002 to, for example, form a specific hardware module at some times and a different hardware module at different times.

  Modules can provide information to other modules and receive information from other modules. For example, the modules described may be considered to be communicatively coupled. If multiple such hardware modules are present at the same time, communication could be achieved through signal transmission (eg, via an appropriate circuit or bus) that connects the modules. In embodiments where multiple modules are configured or instantiated at different times, communication between such modules may store information in or retrieve information from memory structures accessible to the multiple modules. It can be realized through doing. For example, one module performs an operation, and the output of the operation is stored in a memory device to which the module is communicatively coupled. Then, another module accesses the memory device at a later time, retrieves the stored output, and processes it. Modules may also initiate communication with input or output devices and may operate on resources (eg, collections of information).

  The various operations of the example methods described herein may be tentatively performed (eg, software, code, and / or stored on a machine-readable medium to at least partially perform related operations). Can be performed by one or more processors 1002 configured (permanently) or configured permanently. Such a processor 1002 may be a processor-implemented (computer-implemented) module that operates to perform one or more operations or functions, whether provisionally or permanently configured. May be included. The modules referred to herein comprise processor-implemented (or computer-implemented) modules in some example embodiments.

  Also, the methods described herein may be at least partially processor-implemented (or computer-implemented) and / or processor-executable (or computer-executable). For example, at least some of the method operations may be performed by one or more processors 1002 or processor-implemented (or computer-implemented) modules. Similarly, at least some of the method operations are stored in a computer-readable storage medium and governed by instructions that cause one or more processors 1002 or processor-implemented (computer-implemented) modules to execute. May be. Performing certain types of operations is distributed across one or more processors 1002 and can be deployed across multiple machines, not just within a single machine. You can also. In some example embodiments, the processor 1002 is located at a single location (eg, in a home environment, a corporate environment, or as a server farm), and in other embodiments, the processor 1002 is , Distributed across multiple locations.

  Although embodiment (s) have been described in connection with various implementations and uses, these embodiments are for illustrative purposes and the scope of the embodiment (s) is described above. It is not limited to implementation and utilization. In general, the embodiments described herein could be implemented with equipment consistent with any hardware system (s) defined herein. Many variations, modifications, additions and improvements may be implemented.

  Multiple instances may be provided for a component, operation, and / or structure described herein as a single instance. Finally, the boundaries between the various components, operations, and data storage are rather discretionary, and specific operations are shown in the context of specific exemplary configurations. Other functionality assignments are envisioned and will fall within the scope of the embodiment (s). In general, structures and functionality presented as separate components in the illustrative configurations may be implemented as a combined structure or component. Similarly, structures and functionality presented as unitary components can be implemented as separate components. These and other variations, modifications, additions and improvements fall within the scope of the embodiment (s).

Exemplary Embodiments The following discussion includes non-limiting exemplary embodiments for determining sentiment expressed in a document.

  A. An initial list of seed words (seed word list) is formed to assist in identifying relevant aggregate text. A list of seed keywords for a particular subject is obtained from source and subject matter experts, including but not limited to subject glossaries.

  B. Source text is obtained from documents in the aggregate. For example, the aggregate may include documents from a website (eg, a blog, a review website, etc.). The words in each document are compared to the seed word list. If the number and / or ratio of words in the document that match a member of the seed word list meets or exceeds a predetermined threshold, the document is added to the accepted document list. Otherwise, the document is discarded.

  C. The initial seed word list extracts words from the seed word list and keywords and N-grams from documents in the accepted document list (eg, using standard keyword extraction and N-gram extraction tools). Expanded by combining with the list of words obtained by. Keywords and N-grams are further refined using the process described below to calculate the node structure of the selected document before combining the seed word list results.

  D. The document node structure is created to improve the initial set of keywords and N-grams. This process involves extracting information from words (eg, text format) to create a multifractal data structure. The multifractal data structure is one of data structures that exhibit an in-scale scale. The pattern on one scale is slightly different from the pattern found on another scale. Once the data is extracted and its multifractal structure is determined, the data structure is stored in memory so that the computer can access it and perform operations using the data structure. This data structure represents the natural structure of information and is useful for efficiently identifying related information. The data structure is assembled as follows.

1. A word list (seed word list) of k seed category names is created.
2. Each of the k category name words is put into a crawler, and N (k) web pages are taken out from each of the k category name words.
3. Each of the N (k) pages is taken, a list of unique words is created, and the number of times these words appear on each page is counted. Stop words (such as “the”, “his”, “and”, etc.) are removed at this stage. In addition, the number of words is reduced to nouns and adjacent adjectives.
4). M (for example, 10) most frequently occurring words are taken from each of N (k) pages, and the most frequently occurring words on the first page are indexed (for example, 1, 2, 3, 4, 5, 6, 7, 8 , 9, and 10), assigning an index (eg, 11, 13,..., 21) to the most frequent word on the second page, and so on, N (1) items from the first category The index is allocated until the page runs out. A series of words can be identified by a single index (called a subject group). Repeat for N (2) pages of the second category name and continue until all k categories are exhausted.
5. Compute the sum of each of the N (k) pages for all k and name the value N words.
6). Create a graph with a horizontal axis x and a vertical axis y.
7). Create a matrix of N words × N words padded with zeros, and call this W.
8). Take word data from the first web page and calculate the word link coefficient using the following method.
a. For each of the most frequent words listed, a unique pair (x, y) called a word link is created, where x is not the same word as y.
b. Combining word links with page link coefficients, so that high word link and page link coefficients have a higher selection score than low word link and page link coefficients. The page link coefficient is rooted in the link structure and content analysis of the page link entering the web page. This selection score is the value used to pick up the tag from the data. If the page link coefficient is not available, it is set to a default value such as 1.0.
9. For arithmetic processing purposes:
a. The word coefficient of the word link coefficient is expressed as (x, y) = (1, 3), (1, 4),. . . , (1, M + 1). x and y are word indexes from the list defined above.
b. (X, y) = (2, 3),. . . , (2, M + 1), and finally the process is performed for the (M) th and (M + 1) th words, that is, (x, y) = (M, M + 1). For the second category of pages, the first coefficient entry is (x, y) = (M + 2, M + 3). Each of the coefficient entries represents the connectivity of two words that are found or associated with each other in two words from the document.
c. Repeat step 9b for all pages of the subject, except in the following situations:
i. Whenever the current word can be found in the previous x-word index, the coefficients are all found in the previous index and in such a way that large coefficient combinations tend to be larger than small coefficient combinations Combined. For example, the word “dog” corresponds to the word index 4 and is accompanied by a word coefficient (4, 8) of W1, and the word “dog” appears on another page as well. Thus, with index 25 and W2 word coefficient (25,8), all the connectivity coefficients of index 25 are combined into the x value of index 4 for each corresponding index of y. In this case, W (4,8) will be combined with W (25,8) and the result will be assigned to a new W (4,8). After the transfer and combination of coefficients is complete, W (25,8) is set to a small value such as 0.0. This rule applies to all other instances of the “dog” instance that occurred after the first instance.
ii. Duplicate words found on the y word index may be incorporated into the original y word index in the same or similar manner as was done for the x word index. If they are not combined, the effect of each word will be seen as a function of the subject because each subject is within a certain range of the y-index.
10. By calculating a word ranking score from each word link coefficient, the words are ranked in such a way that the word with the highest and highest word link coefficient is ranked higher than the word with the lowest and lowest word link coefficient.
a. Let R (i) be the ranking of a given word node.
b. Let W (i, j) be the connectivity coefficient between words.
c. Calculate as follows: for each i, R (i) = R (W (i, j)) and the ranking is the word link weight for each j (already includes page ranking weight) Function).
d. For tags and groups, select words with greater weight.
11. If you want to analyze the data visually in a graphic format, do the following:
a. Dots proportional to the score based on the word and page link coefficients are denoted by (x, y) = (1, 3), (1, 4),. . . , (1, M + 1) coordinates. The same is true for (x, y) = (2,3),. . . , (2, M + 1) and finally for the Mth and (M + 1) th words, (x, y) = (M, M + 1). For the second category of pages, the first dot will be (x, y) = (M + 2, M + 3). Each dot represents the connectivity of two words that are found or associated with each other in two words from the document. The x-axis and y-axis are word indexes from the list defined above.
b. Repeat section 11a for all subject documents, with the following exceptions:
i. Whenever the current word can be found in the previous x-word index, all dots are recorded in the previous index. For example, if the word “dog” happens to be the word index 4 and also appears in the index 25 as well, all the connectivity dots in the index 25 are moved to the x value in the index 4. After the dots are moved, they are deleted from their original positions on the graph. This rule applies to all other instances of the “dog” instance that occurred after the first instance.
ii. Duplicate words found on the y word index may be incorporated into the original y word index in the same or similar manner as was done for the x word index. If they are not combined, the effect of each word will be seen as a function of the subject because each subject is within a certain range of the y-index.
c. Examine the vertical stripes on the graph and the most prevalent stripe corresponds to the W1 index of the word selected as the tag word or subject group. The W1 word with the most links or dots will be selected as the tag word or subject group. Additional filtering can be applied to these words to further refine the list.

  E. The “tag word” or “theme group” is combined with the seed word list to create a working word list.

  F. The actual word list is obtained from another document (for example, through crawling, obtaining a document from an alligator, etc.), and a predetermined number of additional documents (additional list extension number) by the above C and D stages. It is reinforced by obtaining more documents.

  G. The F stage is repeated until the growth rate of the harvested N-gram falls to a satisfactory rate.

  H. For the sentiment analysis engine, a list of topics, adjectives, and proper nouns about the subject is created. For example, if the subject is “Hotel”, the topics / keywords include concierge, bathroom, bed, bedroom, front desk, TV, room service, maid service, cleanliness, elevator, restaurant, reservation, and checkout It will be. Adjectives include crisp, pleasant, sloppy, rude, clean, dirty, kind, impressive, terrible, nice, confused, poor, filthy, unfriendly That would be true. Proper nouns would include Marriott, Hyatt, Four Seasons, Motel 6, Intercontinental, Quality Inn, Howard Johnson, etc.

  I. Sentences containing both keywords and adjectives are identified for further analysis.

  J. et al. A keyword, a group of adjectives, and a subject matter of interest (eg, the name of a hotel chain) are correlated by looking for proper names before and after the keyword and adjective group. Proper nouns and keyword / adjective groups are saved for future analysis and presentation.

  K. For review sites, the date is obtained and the site is revisited periodically. The period identifies the date of the blog and / or review entry, calculates the time period from one blog entry to the next, and how often the review site is revisited based on the blog entry time period Can be determined by determining. An invariant time period (eg, a minimum time period or a third of the average time period) may be selected. Instead, a variable time period that depends on a predetermined factor (such as the season) may be selected. For example, in the case of a hotel, we will post more entries around holidays like July 4th, so the time from crawling to the next crawling will be shorter around this time. The variable period may be selected based on local events (eg, Indianapolis 500 Mile Race). For example, in the case of a hotel, the crawling to crawling period for a hotel offered in the Indianapolis 500 mile race will be reduced around the time of the event. Once the sampling time period is established, the site is crawled for the next text.

  L. A time dependency may be created for each category from each website.

  M.M. For a subject, a list of categories is obtained. A list of seed categories is obtained by looking at the headlines on the subject website. Word frequency on other subject websites is identified.

N. The importance of the review site is calculated as follows.
a. The incoming and outgoing links of the review site are analyzed.
i. The validity of the linked sites is checked to make sure they are not SPAM links.
ii. The vocabulary distribution of words on the linked site is analyzed.
iii. If the vocabulary spectrum for the linked site is not within the predetermined threshold of the reference vocabulary spectrum, the link is discarded.
b. The update frequency of a blog entry or comment / evaluation for a site vs. a timestamp delta at a competing site for a given subject is identified.
i. A short timestamp delta means that many people in the community are actively reading and updating that particular website on a regular basis. The site with the fastest legitimate comments is considered the most important site in the community.

O. Proposals from community members are identified as follows:
a. A list of patterns that identify recommendations is created. For example, a pattern might include “Recommended”, “There is room for improvement”, “Teach all friends”, and so on.
b. Thematic keywords or N-grams in these sentences are identified.
c. Once keyword patterns are found, these sentences are saved for presentation as suggestions for their subject matter.

P. User interface for mobile devices.
a. User interface designs for viewing consumer sentiment ratings for various products and services, such as hotel services, must be formatted to facilitate browsing, data entry, and scoring searches. Various parameters versus time graphs in the form of XY diagrams, or ranked issues in the form of Pareto diagrams, need to fit within the screen of the smartphone. Horizontal and vertical scrolling to view various data can be operated by swiping your finger on the screen, by mini trackball or thumb joystick, or by toggle user key do it. The data may be list fragments, time-dependent plots, and / or Pareto charts that contain source descriptions from blogs that expose sentiment.

  The above description is described in association with a specific embodiment for explanation. However, the foregoing discussion for purposes of explanation is not intended to be exhaustive or to limit the embodiments to the precise forms disclosed. Many modifications or variations are possible in light of the above teaching. The embodiments describe the principles and their practical application in the most understandable way so that those skilled in the art can best utilize the various embodiments with various modifications commensurate with those embodiments and the specific applications contemplated. So that it can be selected.

100 server 102 document contained in server 104 aggregator 106 document contained in aggregator 108 sentiment server 120 network 202 sentiment module 204 crawling module 206 keyword module 208 filtering module 210 classification module, classification engine 1000 computer system 1002 processor 1004 Memory 1006 Video display unit 1008 Bus 1010 Input device 1012 Output device 1016 Network interface device 1020 Machine-readable medium 1022 Data structure and instruction

Claims (20)

In a computer-implemented method for seeking sentiment (sentiment) expressed in a document,
Receiving a document from a plurality of documents;
Using at least one processor to identify sentences in the document that contain at least one sentiment signature within a predetermined distance of at least one keyword from the keyword list, the list of keywords Are extracted from the plurality of documents and filtered using a phase transition equation, and the at least one sentiment signature corresponds to a representation of at least one sentiment in the sentence, Identifying, and
Determining at least one category corresponding to the at least one keyword of the sentence, wherein the at least one category is included in a list of categories generated using the list of keywords; Seeking a category,
Determining at least one sentiment corresponding to the at least one category based on the at least one sentiment signature. Prior to identifying the sentence that includes the at least one sentiment signature within a predetermined distance of the at least one keyword from the list of keywords,
Extracting a keyword from each of the plurality of documents;
For each keyword,
Calculating the frequency f of the keywords in the plurality of documents and the number N of documents including the keywords;
Using the phase transition formula to calculate the relevance of the keyword based on the frequency of the keyword in the plurality of documents and the number of documents containing the keyword;
The computer-implemented computer of claim 1, further comprising: extracting the keyword list by adding the keyword to the keyword list if the relevance of the keyword exceeds a predetermined threshold. the method of. The phase transition equation is

The computer-implemented method of claim 2, wherein x ≧ 1.   The computer-implemented method of claim 3, wherein x is three. Prior to determining at least one category corresponding to the at least one keyword of the sentence,
Identifying a first set of documents among the plurality of documents that includes at least one keyword from the list of keywords;
Identifying a set of keywords contained in at least a predetermined number of documents in the first set of documents;
The computer-implemented method of claim 1, further comprising generating the list of categories by adding the set of keywords to the list of categories, each category including each keyword set. Method. Prior to determining at least one category corresponding to the at least one keyword of the sentence,
In the keyword list, a keyword pair that is a pair of keywords that are related to each other and that is a unique keyword pair is obtained.
A set of keyword pairs, each set identifying a set of keyword pairs that includes at least one keyword that is common to all of the keyword pairs in the set;
The set of keyword pairs are iteratively combined until a predetermined termination condition is achieved, and each combined set is common to all of the keyword pairs in the combined set. The computer-implemented method of claim 1, further comprising generating the list of categories by combining in such a manner that at least one keyword is included.   Determining at least one category corresponding to the at least one keyword of the sentence includes using a support vector machine to determine the at least one category corresponding to the at least one keyword of the sentence. The computer-implemented method of claim 1.   Determining at least one category corresponding to the at least one keyword of the sentence includes determining the at least one category corresponding to the at least one keyword of the sentence using a neural network. The computer-implemented method of claim 1. Determining at least one category corresponding to the at least one keyword of the sentence;
Obtaining a category spectrum comprising a plurality of category spectra, each category spectrum including the frequency of occurrence of a keyword in the list of keywords corresponding to each category;
Determining a category spectrum for the sentence based on the at least one keyword;
Calculating a dot product of a category spectrum of the sentence and each category spectrum in the plurality of category spectra;
The computer-implemented method of claim 1, comprising: determining the at least one category as a category corresponding to at least one dot product that exceeds a predetermined threshold. Prior to obtaining the plurality of category spectra, for each category,
Obtain a collection of documents corresponding to the category,
Extracting keywords from each document in the document collection;
Filter the keyword using the phase transition equation to derive a filtered keyword,
Determining the frequency of occurrence of the filtered keyword in the collection of documents;
The computer-implemented method of claim 9, further comprising: determining the category spectrum for the category by normalizing the frequency of occurrence of the filtered keyword to derive a category spectrum for the category. Method. Prior to receiving a document from a plurality of documents,
For each document in the document collection,
Extract n-grams from the document;
For the document, based on the extracted n-gram, determine an n-gram spectrum indicating the frequency of occurrence of the n-gram as a function of the size of the n-gram;
Determining whether the n-gram spectrum of the document is matched within a predetermined threshold to a reference n-gram spectrum defined by a predetermined equation;
If the n-gram spectrum of the document matches the reference n-gram spectrum within the predetermined threshold, add the document to the plurality of documents;
If the n-gram spectrum of the document does not match the reference n-gram spectrum within the predetermined threshold, selecting the plurality of documents from the collection of documents by discarding the document The computer-implemented method of claim 1, further comprising: The predetermined equation is

Where x is the size of the n-gram, and a, b, and c are predetermined values where the peak of the predetermined equation is between an n-gram of size 2 and an n-gram of size 3. The computer-implemented method of claim 11, wherein   The computer-implemented method of claim 1, wherein the at least one sentiment signature includes at least one word that indicates that an expression of the at least one sentiment is present in the sentence.   The computer-implemented method of claim 1, wherein the at least one sentiment is a representation of an opinion related to at least one category.   The computer-implemented method of claim 1, wherein the at least one category is associated with a product.   The computer-implemented method of claim 1, wherein the at least one category is associated with a service. In a system for seeking sentiment expressed in a document,
At least one processor;
Memory,
At least one program stored in the memory,
Instructions for receiving a document from a plurality of documents;
Instructions for identifying a sentence in the document that includes at least one sentiment signature within a predetermined distance of at least one keyword from the keyword list, the keyword list comprising the plurality of documents Instructions for identifying a sentence, wherein the at least one sentiment signature corresponds to a representation of at least one sentiment in the sentence, and is filtered using a phase transition equation ,
Instructions for determining at least one category corresponding to the at least one keyword of the sentence, wherein the at least one category is included in a list of categories generated using the list of keywords An instruction to ask for a category,
And at least one program comprising: instructions for determining at least one sentiment corresponding to the at least one category based on the at least one sentiment signature. An instruction for extracting the keyword list; and an instruction for extracting the keyword list from the document,
Extracting a keyword from each of the plurality of documents;
For each keyword,
Calculating the frequency f of the keywords in the plurality of documents and the number N of documents including the keywords;
Using the phase transition formula to calculate the relevance of the keyword based on the frequency of the keyword in the plurality of documents and the number of documents containing the keyword;
The system according to claim 17, further comprising: an instruction for adding the keyword to the keyword list if the relevance of the keyword exceeds a predetermined threshold. The phase transition equation is

The system of claim 18, wherein x ≧ 1. In a computer readable storage medium storing at least one program configured to be executed by a computer, the at least one program comprises:
Instructions for receiving a document from a plurality of documents;
Instructions for identifying a sentence in the document that includes at least one sentiment signature within a predetermined distance of at least one keyword from the keyword list, the keyword list comprising the plurality of documents Instructions for identifying a sentence, wherein the at least one sentiment signature corresponds to a representation of at least one sentiment in the sentence, and is filtered using a phase transition equation ,
Instructions for determining at least one category corresponding to the at least one keyword of the sentence, wherein the at least one category is included in a list of categories generated using the list of keywords An instruction to ask for a category,
Instructions for determining at least one sentiment corresponding to the at least one category based on the at least one sentiment signature.

Images