CN110852095B

CN110852095B - Statement hot spot extraction method and system

Info

Publication number: CN110852095B
Application number: CN201810871010.0A
Authority: CN
Inventors: 费志军; 邱雪涛; 王宇
Original assignee: China Unionpay Co Ltd
Current assignee: China Unionpay Co Ltd
Priority date: 2018-08-02
Filing date: 2018-08-02
Publication date: 2023-09-19
Anticipated expiration: 2038-08-02
Also published as: CN110852095A

Abstract

The invention relates to a statement hot spot extraction method, which comprises the following steps: extracting at least one emotion keyword from a document; wherein the document includes at least one sentence; performing dependency grammar analysis on the first sentence to generate a dependency syntax tree; pruning the dependency syntax tree based on the first emotion keywords to form a result syntax tree; wherein the first emotion keyword is derived from a first sentence; forming a text vector set based on the result syntax tree; and clustering the set of text vectors to form at least one class of interest. The method is beneficial to improving the accuracy of the clustering result, does not influence the original semantics, and can more accurately express the hot spot concerned by the client.

Description

Statement hot spot extraction method and system

Technical Field

The invention relates to the technical field of text analysis, in particular to a sentence hot spot extraction method and a sentence hot spot extraction system.

Background

The customer service center can receive a large number of customer complaints and consultations, then classify the consultation documents according to actual consultation categories of customers through operators and other related service personnel, import the results into a database, acquire consultation hotspots according to statistical information of the consultation documents, and analyze the acquired consultation hotspots.

However, as the customer consultation contents are gradually diversified and the number of consultation documents is gradually increased, a large number of consultation documents are classified only by means of manual mode, and further classified documents are counted to obtain consultation hotspots and to perform hotspot analysis, so that time and effort are consumed, and subjectivity of analysts is excessively relied on. If the sentences in the consultation documents are extracted and automatically analyzed, the customer evaluation and feedback can be timely obtained, and further the updating/upgrading of the products or services is facilitated.

The prior art provides the following technical proposal: k consultation documents are extracted from a plurality of consultation documents and used as k document categories, namely, used as centers of initial clustering, then similarity between other consultation documents and each document category is calculated respectively, each consultation document is classified into the document category with the highest similarity, and then an automatic classification process is realized.

However, the information involved in the dialogue between the customer and the customer service personnel is quite abundant, and much information is not relevant to the problem that the customer wants to consult, and the method proposed by the prior art is not suitable for filtering irrelevant noise, and the noise may mask valid data, so that the document clustering process is inaccurate, the execution efficiency of the clustering algorithm is low, and the expected analysis effect is quite different from the actual effect.

Disclosure of Invention

The invention aims to provide a statement hot spot extraction method which can be used for more effectively filtering noise and uninteresting information and improving the clustering accuracy and the execution efficiency of a clustering algorithm.

In order to achieve the above purpose, the present invention provides a technical solution as follows.

A sentence hot spot extraction method comprises the following steps: a) Extracting at least one emotion keyword from the document; wherein the document includes at least one sentence; b) Performing dependency grammar analysis on the first sentence to generate a dependency syntax tree; c) Pruning the dependency syntax tree based on the first emotion keywords to form a result syntax tree; wherein the first emotion keyword is derived from a first sentence; d) Forming a text vector set based on the result syntax tree; and e) clustering the set of text vectors to form at least one class of interest.

Preferably, the method further comprises: for each class of interest: extracting at least one hot spot keyword; generating at least one hotspot expression based on word frequency information of each hotspot keyword, wherein the hotspot expression corresponds to an arrangement combination of the at least one hotspot keyword; and determining a hotspot representation based on the support of each hotspot description in the document.

Preferably, step c) comprises: c1 Determining a root node of the dependency syntax tree; c2 And performing breadth-first traversal based on the root node, and pruning nodes and subtrees of the nodes, of which the grammar relationship does not accord with the dependency relationship white list, from the dependency syntax tree.

Preferably, in step c 1) the root node is determined by: and searching the leaf nodes upwards along the dependency syntax tree by taking the emotion keywords as current searching leaf nodes, or determining the node with the current parallel relation reaching the maximum depth limit value as the root node.

Preferably, the dependency white list comprises: COO; VOB; ADV; and POB.

Preferably, in step b): the syntax relationship and parts of speech of each node of the dependency syntax tree satisfy the LTP specification and the ICTCLAS standard, respectively.

Preferably, step a) specifically comprises: sentences that do not contain any emotion keywords are excluded from the document.

Preferably, the clustering algorithm used in step e) comprises: k-means clustering algorithm.

The invention also discloses a sentence hot spot extraction system, which comprises: the emotion keyword extraction unit is used for extracting at least one emotion keyword from the document; a dependency syntax tree generation unit for performing dependency syntax analysis on the first sentence to generate a dependency syntax tree; wherein the first sentence originates in the document; a result syntax tree generation unit respectively coupled to the emotion keyword extraction unit and the dependency syntax tree generation unit, which prunes the dependency syntax tree based on the first emotion keyword to form a result syntax tree; wherein the first emotion keyword is derived from a first sentence; a vector conversion unit for forming a text vector set based on the result syntax tree; and a vector clustering unit for clustering the text vector sets to form at least one class of interest.

The invention provides a sentence hot spot extraction method, which refines a client consultation document by adopting a method of pruning a dependency syntax tree based on emotion keywords before cluster analysis, thereby improving the accuracy of a clustering result. In addition, when constructing the hot spot expression, the method can take word frequency into consideration to synthesize expression sentences, so that the automatically extracted consultation hot spot is not only representative but also easy to understand.

Drawings

Fig. 1 shows a flowchart of a statement hotspot extraction method provided in a first embodiment of the present invention.

Fig. 2 is a schematic block diagram of a sentence hot spot lifting system according to a second embodiment of the present invention.

Detailed Description

In the following description, specific details are set forth in order to provide a thorough understanding of the present invention. It will be apparent, however, to one skilled in the art that embodiments of the invention can be practiced without these specific details. In the present invention, specific numerical references such as "first element", "second device", etc. may be made. However, a specific numerical reference should not be construed as necessarily subject to its literal order, but rather as a "first element" distinct from a "second element".

The particular details presented herein are exemplary only and the particular details may vary and yet fall within the spirit and scope of the present invention. The term "coupled" is defined as either directly connected to an element or indirectly connected to an element via another element.

Preferred embodiments of methods, systems and apparatus suitable for implementing the present invention are described below with reference to the accompanying drawings. Although the embodiments are described with respect to a single combination of elements, it is to be understood that the invention includes all possible combinations of the disclosed elements. Thus, if one embodiment includes elements A, B and C, while a second embodiment includes elements B and D, the present invention should also be considered to include other remaining combinations of A, B, C or D, even if not explicitly disclosed.

As shown in fig. 1, a first embodiment of the present invention provides a statement hotspot extraction method, which includes the following steps.

And step S10, extracting at least one emotion keyword from the document.

The consultation file comprises a dialogue record of a customer and a customer service center in a period of time or about a certain problem, wherein a plurality of sentences are recorded, and the customer sentences and the customer service sentences are protected.

By way of example, an advisory document is shown in the following table, which records a customer's inquiry dialogue with a customer service about a transaction record.

Firstly, the collected customer service dialogue data can be preprocessed, meaningless information or noise such as 'xxx number customer service personnel' and 'visitor xxx' is removed, and an original corpus is obtained, wherein the content of the original corpus is shown in the following table.

Then, an emotion keyword set is determined from the original corpus. The inventors have found statistically that in the consultation dialogue of the customer and the customer service (visitor), a part of words with emotion colors (e.g.

"none", etc.) are germane to the hotspot of interest to the customer, and thus, these keywords may be determined as emotion keywords.

According to step S10, at least one emotion keyword can be extracted from the consultation document. The sentences shown in the black boxes in the table above contain the emotion keywords "none" and can be used for subsequent analysis. Preferably, sentences which do not contain any emotion keywords can be eliminated from the original corpus, so that the accuracy of subsequent clustering and the execution efficiency of an algorithm are improved.

And S12, performing dependency grammar analysis on the first statement to generate a dependency syntax tree.

The first sentence is derived from an original corpus, and the original corpus is a result obtained by preprocessing an consultation document. The statement "why me did not get the exempt benefit" shown in the black box in the table above is an example of the first statement.

Because the emotion keywords can better represent the hotspots focused by the clients, according to the embodiment of the invention, sentences containing the emotion keywords in the original corpus can be respectively used as the first sentences in the invention, which is the preferred embodiment of the invention. In other words, steps S12 and S14 may be performed for each sentence including emotion keywords in the original corpus.

As an alternative embodiment, in the case that emotion keywords extracted from an original corpus have a plurality of emotion keywords and can be ranked according to priorities, one sentence containing the emotion keywords with the highest priority is extracted from the original corpus as a first sentence, or a sentence containing more emotion keywords is used as the first sentence, so that the way does not need to analyze all sentences containing emotion keywords in the original corpus, and therefore, the efficiency of sentence analysis can be improved.

Dependency parsing (Dependency Pars ing, abbreviated as DP) reveals the syntactic structure by analyzing the dependencies between components within a language unit. Intuitively, dependency parsing recognizes grammar components such as "main predicate" and "definite form complement" in sentences, and analyzes the relationship between the components. By way of example, the client in the consultation document says "why I did not get a exemption offer", whose dependency syntax analysis result is as follows, and a dependency syntax tree can be generated from the dependency syntax analysis result.

Preferably, in generating the dependency syntax tree based on the dependency syntax analysis, part-of-speech tags of nodes of the dependency syntax tree conform to the ICTCLAS standard, and syntax relationships between the nodes satisfy the LTP specification. Alternatively, the parsing may be performed using the Stanford Parser work.

Part-of-speech Tagging (POS) is a task that gives each word in a sentence a Part-of-speech category. The part-of-speech category here may be nouns, verbs, adjectives, or others. Specifically, v represents a verb, n represents a noun, r represents a pronoun, c represents an conjunctive, a represents an adjective, d represents an adverb, wp represents a punctuation mark. According to the LTP specification, grammatical relations between words include a variety of SBV main-predicate relations, VOB move-guest relations, IOB mediate-guest relations, FOB pre-object, DBL doublets, ATT centering relations, ADV-like intermediate relations, CMP move-complement relations, COO parallel relations, POB mediate-guest relations, and the like.

And S14, pruning the dependency syntax tree based on the first emotion keywords to form a result syntax tree.

This step is one of the key steps of the present invention, according to which the first emotion keyword is derived from the first sentence. As an example, the first sentence is "why i did not get a reduced preference", and the corresponding first emotion keyword is the word "none" in the first sentence.

Specifically, this step S14 may be performed as follows. First, a root node of each dependency syntax tree is determined. Secondly, breadth-first traversal is performed from the root node, and those nodes and corresponding subtrees of which the grammar relationships do not accord with the dependency relationship white list are pruned from the dependency syntax tree.

As one example, a leaf node is currently searched for emotion keywords, a Root node is searched for along a dependency syntax tree until the Root node (Root) or the current parallel relationship ("COO") reaches a maximum depth limit (e.g., set to 2), the node is determined to be a syntax tree Root node, and the node is added to the resulting syntax tree. For example, the first sentence is "why I did not get a reduced preference," its emotion keyword is "none," and its parent node is

The parent node of "get", "get" is "Root", and thus "get" is determined to be the Root node of the current syntax tree.

Then, starting from the obtained root node of the dependency syntax tree, performing a restricted breadth-first traversal. According to one implementation, only those nodes and corresponding subtrees whose grammatical relations between parent and child nodes do not conform to the dependency whitelist are pruned from the syntax tree in the breadth-first traversal. As another implementation, those nodes whose grammatical relations and parts of speech between parent and child nodes satisfy the dependency whitelist are retained, while other nodes, along with the corresponding subtrees, are all pruned from the syntax tree until traversal is complete for all leaf nodes or the current parallel relationship ("COO") reaches a maximum depth.

Preferably, the dependency whitelist is shown in the following table, wherein only part of the preferred whitelist is shown, e.g. grammatical relations COO, VOB, ADV and POB. It will be appreciated that the whitelist may be modified depending on the particular application scenario.

For example, in the first sentence "why I did not get a reduced preference," get "

The child node (part of speech v) is "why" (part of speech r), "I" (part of speech r),

"none" (part of speech d) and "exempt" (part of speech v); "get" and "why" are related to ADV, but parts of speech do not conform to the dependency whitelist, so "why" is discarded; the "none" relationship is that "ADV" conforms to the dependency white list, so "none" joins the result syntax tree, and similarly "get" also joins the result syntax tree, and "me" is also discarded.

In this way, pruning of the dependency syntax tree is completed, and the resulting syntax tree corresponds to the statement "no mitigation offers" and has a significantly simplified syntax structure relative to the dependency syntax tree. Typically, a corresponding result syntax tree can be obtained from each first sentence containing emotion keywords; it is also possible that the dependency syntax tree of the first statement is pruned entirely during pruning.

And S16, forming a text vector set based on the result syntax tree.

In this step, a text vector set can be formed from at least one dependency syntax tree using text-to-digitize techniques such as one_hot, tf-idf, wordbec, etc., which contains a plurality of text vectors. Text vectors are units of language that a computer can calculate and thereby use to cluster. Each sentence includes one or more text vectors. It will be appreciated that two sentences that are similar may have a plurality of identical text vectors, but also necessarily differ, for example, by including text vectors that differ from each other, respectively; alternatively, one sentence has more text vectors and the other sentence has fewer text vectors.

Step S18, clustering the text vector sets to form at least one interest class.

In this step, the text vector set is clustered using, for example, a K-means clustering algorithm, and the clustering results may be integrated using a cluster optimization (merging) algorithm known in the art, resulting in one or more interest classes, each of which is significantly different from the other classes and contains at least one clause.

On the basis of the syntax tree pruning, the execution efficiency of the clustering algorithm and the accuracy of the clustering result are improved, so that the hot spots concerned by the clients can be more accurately reflected.

According to a further improved embodiment of the present invention, following step S18, the following subsequent steps may also be performed. Specifically, the processing is performed separately for each class of interest: a1, extracting at least one hot spot keyword; a2, generating at least one hotspot expression based on word frequency information of each hotspot keyword, wherein the hotspot expression corresponds to an arrangement combination of the at least one hotspot keyword; a3, determining a hotspot expression based on the support degree of each hotspot description in the document.

As an example, an algorithm (e.g., textRank algorithm or TF-IDF algorithm) well known to those skilled in the art may be employed in A1 to extract the hotspot keywords. In A2, selecting hot spot keywords with higher word frequency, and forming different hot spot descriptions by utilizing various arrangement combinations of the hot spot keywords; it will be appreciated that these hotspot descriptions are often not easily understood or do not conform to grammatical practices. Wherein the permutation and combination of the keywords may not be arbitrary, but conform to some kind of abstract template, for example dvn (adverb-verb-noun) template. In A3, determining the support degree or the occurrence probability of various hotspot descriptions in the consultation document, and taking a plurality of hotspot descriptions with the highest support degree or occurrence probability as hotspot expressions finally displayed to the user; it will be appreciated that the hotspot representation so formed best conforms to the language habits of the client in the consultation document.

The second embodiment of the present invention provides a sentence hotspot extraction system 20, which includes an emotion keyword extraction unit 201, a dependency syntax tree generation unit 202, a result syntax tree generation unit 211, a vector conversion unit 220, and a vector clustering unit 231.

The emotion keyword extraction unit 201 extracts at least one emotion keyword from the consultation document. Emotional keywords include, for example, "none", "many", and

"less".

The dependency syntax tree generation unit 202 performs dependency syntax analysis on a first sentence selected from the advisory document to generate a dependency syntax tree. The first sentence may be any sentence in the advisory document, or a sentence in the advisory document containing only emotion keywords.

Result syntax tree generation unit 211 is coupled to emotion keyword extraction unit 201 and dependency syntax tree generation unit 202, respectively, which prunes the dependency syntax tree based on the first emotion keywords to form a result syntax tree. The first emotion keyword is derived from the first sentence, that is, when the first sentence contains the emotion keyword, the first emotion keyword corresponds to the first sentence.

The vector conversion unit 220 forms a text vector set based on the result syntax tree.

The vector clustering unit 231 clusters the set of text vectors to form at least one class of interest, wherein each class of interest contains at least one clause and is significantly different from other classes.

As a preferred embodiment, the sentence hot spot extraction system 20 further includes a hot spot extraction unit (not shown in the drawings) that extracts at least one hot spot keyword for each class of interest, respectively; generating at least one hotspot expression based on word frequency information of each hotspot keyword, wherein the hotspot expression is obtained by arranging and combining different hotspot keywords. The hotspot extraction unit may further determine a hotspot representation based on the support of each hotspot description in the consultation document and present to a user of the system.

The sentence hot spot extraction system 20 is not only suitable for carrying out hot spot extraction on client consultation documents, but also suitable for other sentence analysis occasions so as to efficiently and accurately acquire hot spots concerned by clients, and further can start, for example, an intelligent response function to respond. Alternatively, such a sentence hotspot extraction system may be integrated therein as a subsystem of an intelligent answer system or a game system or a speech recognition system, implementing rich functionality, facilitating the user's use experience.

In some embodiments of the invention, at least a portion of the system may be implemented using a set of distributed computing devices connected by a communications network, or based on a "cloud". In such a system, multiple computing devices operate together to provide services by using their shared resources.

The "cloud" based implementation may provide one or more advantages, including: openness, flexibility and extensibility, centralness management, reliability, scalability, optimization of computing resources, the ability to aggregate and analyze information across multiple users, the ability to connect across multiple geographic areas, and the use of multiple mobile or data network operators for network connectivity.

The above description is only for the preferred embodiments of the invention and is not intended to limit the scope of the invention. Numerous variations and modifications can be made by those skilled in the art without departing from the spirit of the invention and the appended claims.

Claims

1. A sentence hot spot extraction method comprises the following steps:

a) Extracting at least one emotion keyword from the document; wherein the document includes at least one sentence;

b) Performing dependency grammar analysis on the first sentence to generate a dependency syntax tree;

c) Pruning the dependency syntax tree based on the first emotion keywords to form a result syntax tree; wherein the first emotion keyword is derived from the first sentence;

d) Forming a text vector set based on the result syntax tree; and

e) Clustering the set of text vectors to form at least one class of interest,

wherein said step c) comprises:

c1 Determining a root node of the dependency syntax tree;

c2 Performing breadth-first traversal based on the root node, pruning nodes whose grammar relationships do not conform to the dependency relationship white list and subtrees of the nodes from the dependency syntax tree,

wherein in step c 1) the root node is determined by:

searching up to the root node along the dependency syntax tree by taking the emotion keywords as current searching leaf nodes, or determining the node with the current parallel relation reaching the maximum depth limit value as the root node,

wherein, the emotion keyword includes:

"not"; "none"; "none".

2. The method according to claim 1, wherein the method further comprises:

for each of the classes of interest:

extracting at least one hot spot keyword;

generating at least one hotspot description based on word frequency information of each hotspot keyword, wherein the hotspot description corresponds to an arrangement combination of the at least one hotspot keyword; and

and determining a hotspot expression based on the support degree of each hotspot description in the document.

3. The method of claim 1, wherein the dependency whitelist comprises:

COO; VOB; ADV; and POB.

4. The method according to claim 1, characterized in that in said step b):

the syntax relationship and the part of speech of each node of the dependency syntax tree respectively meet the LTP specification and the ICTCLAS standard.

5. The method according to claim 1, wherein step a) comprises in particular:

and excluding the sentence which does not contain any emotion keyword from the document.

6. The method of claim 1, wherein the clustering algorithm employed in step e) comprises:

k-means clustering algorithm.

7. A statement hotspot extraction system, comprising:

the emotion keyword extraction unit is used for extracting at least one emotion keyword from the document;

a dependency syntax tree generation unit for performing dependency syntax analysis on the first sentence to generate a dependency syntax tree; wherein the first sentence originates in the document;

a result syntax tree generation unit, coupled to the emotion keyword extraction unit and the dependency syntax tree generation unit, respectively, that prunes the dependency syntax tree based on a first emotion keyword to form a result syntax tree; wherein the first emotion keyword is derived from the first sentence;

a vector conversion unit for forming a text vector set based on the result syntax tree; and

a vector clustering unit for clustering the set of text vectors to form at least one class of interest,

wherein the result syntax tree generation unit performs the following actions:

c1 Determining a root node of the dependency syntax tree;

c2 Performing breadth-first traversal based on the root node, and pruning nodes with grammar relationships not conforming to a dependency relationship white list and subtrees of the nodes from the dependency syntax tree, wherein in c 1), the root node is determined by the following method:

wherein, the emotion keyword includes:

"not"; "none"; "none".