KR102052823B1 - Topic model automation method and apparatus using latent dirichlet allocation - Google Patents

Abstract

Topic modeling methods and apparatus using latent Dirichlet allocation are disclosed. According to an embodiment, a topic modeling method may include: extracting a topic from an object to optimize a prior probability related to a distribution of the topic, and refining the topic by dividing or merging the extracted topic based on the prior probability And automatically labeling the refined topic by extracting a sentence representative of the refined topic.

Description

TOPIC MODEL AUTOMATION METHOD AND APPARATUS USING LATENT DIRICHLET ALLOCATION

The embodiments below relate to a method and apparatus for topic model automation using latent Dirichlet allocation.

Late Dirichlet Allocation (LDA), one of the active statistical topic modeling techniques, is one of the popular techniques for text mining. LDA is often used to summarize and cluster large documents. In addition, LDA is useful in various fields. It is typically used for technology management, marketing, investment, technology change, and recently, it is also used for bio analysis.

However, the LDA used in various fields has a disadvantage that it is difficult for the general user because it is a method that the statisticians theoretically made. First, since LDA is based on Bayesian statistical methodology, the initial values of α and β, which are prior probabilities, must be determined in advance.

Traditionally, the default value is used to determine the initial value or a domain expert arbitrarily determines the initial value. In this method, it is not easy to find accurate prior probabilities in large documents, and the prior probabilities determined by subjective judgments of domain experts have a negative effect on the results of LDA.

Second, since LDA extracts topics by statistical methods, even if the optimal prior probability is used, the topic results are often inaccurate or unknown. This is because many of the topic results generated by the quantitative and probabilistic methods generate duplicate or ambiguous topics. Therefore, it is necessary to automatically post-process the results of the LDA to clean the topic results.

Third, if there is insufficient domain knowledge of the large text document corpus given as input, it is not easy to determine the meaning of each resulting topic, only an abstract level of understanding is possible. That is, in large text documents, LDA can be used to infer only general or common sense content through the topic. Therefore, there is a need for a topic labeling method that automatically indicates the meaning of the topic.

The embodiments below can provide techniques for automating topic modeling using latent Dirichlet allocation.

According to an embodiment, a topic modeling method may include: extracting a topic from an object to optimize a prior probability related to a distribution of the topic, and refining the topic by dividing or merging the extracted topic based on the prior probability And automatically labeling the refined topic by extracting a sentence representative of the refined topic.

The optimizing may include extracting the topic from the object, calculating an average topic coherence according to the change of the prior probability for the topic, and optimizing the dictionary based on the average topic consistency. Employing a probability.

The calculating may include calculating the average topic consistency based on the following equation.

[Equation]

Here, TC means topic consistency, and means semantic similarity values of the related words w _l and w _{f in} the sim (w _l , w _f ) topic.

The refining may include dividing the topic based on semantic similarity of the related words included in the topic or merging the topic based on the similarity between topics calculated from the related words. .

The dividing may be performed by clustering the related words to generate a plurality of clusters, calculating mean similarity between related words within each cluster, and calculating a within-cluster similarity. Calculating cross-clusters similarity by averaging the semantic similarities between the association terms of the clusters, and dividing the plurality of clusters into different topics based on the internal similarity and the cross similarity. .

The merging may include calculating an association between the topic and a sentence for the topic based on the association words, replacing the association words with the sentence based on the association, and different from the sentence. Calculating a bipartite graph by calculating cosine similarity between sentences included in the topic, calculating a topic similarity between the topic and the different topic, and based on the topic similarity And merging the different topics.

The labeling may include adopting a sentence highly related to the topic as a candidate sentence, and extracting a sentence representing the topic based on a probability distribution of related words included in the candidate sentence to label the topic. It may include a step.

The adopting may include removing association words having a small probability value from the topic based on a probability distribution of the association words included in the topic, and comparing a sentence included in the topic and a pair of association words included in the topic. Generating the number of sentences paired with each associated word based on the pair; and removing the pair based on the number to adopt the candidate sentence.

The labeling may include calculating a labeling similarity between the topic and the candidate sentence based on the probability distribution of the related words included in the candidate sentence and the pair, and representing the topic corresponding to the topic based on the labeling similarity. And extracting the sentence.

The calculating of the labeling similarity may include calculating the labeling similarity based on the following equation.

[Equation]

Here, score means the labeling similarity between the topic T and the candidate sentence s (similarity between T and s), t means a word within the topic, and α means a weight indicating the importance of each term, P (t) denotes a probability value for the topic word t, and wordnet (w ', T) is a wordnet-based semantic similarity between the words constituting s that are not in T and each word in topic T ( semantic similarity based on WordNet) A topic modeling apparatus according to an embodiment includes a receiver for receiving an object, a controller for automatically extracting and refining a topic from the object, and automatically labeling the topic, The controller is configured to extract a topic from the object to optimize a prior probability associated with a distribution of the topic, and to divide or extract the topic extracted based on the prior probability. It may include a topic refiner for refining the topic by merging, and a labeler for automatically labeling the refined topic by extracting sentences representing the refined topic.

The prior probability extractor extracts the topic from the object, calculates an average topic coherence for the topic according to the change of the prior probability, and adopts an optimized prior probability based on the average topic consistency. Can be.

The prior probability extractor may calculate the average topic consistency based on the following equation.

 [Equation]

Here, TC means topic consistency and semantic similarity value of the related words w _l and w _{f in} the sim (w _l , w _f ) topic.

The topic refiner may split the topic based on semantic similarity of related words included in the topic or merge the topics based on similarity between the topics based on the related words.

The topic refiner clusters the related words to generate a plurality of clusters, calculates internal similarities by averaging semantic similarities between related words within each cluster, and averages similarity between related words of different clusters. Cross similarity may be calculated and a plurality of clusters may be divided into different topics based on the internal similarity and the cross similarity.

The topic refiner calculates an association between the topic and a sentence for the topic based on the association words, replaces the association words with the sentence based on the association, and includes sentences included in a topic different from the sentence. The cosine similarity of the liver may be calculated to generate a bipartite graph, the topic similarity between the topic and the different topic may be calculated, and the topic and the different topic may be merged based on the topic similarity.

The labeler may select a sentence highly related to the topic as a candidate sentence, and extract the sentence representing the topic based on a probability distribution of related words included in the candidate sentence to label the topic.

The labeler removes association words having a small probability value from the topic based on a probability distribution of the association words included in the topic, generates a pair of sentences included in the topic and an association word included in the topic, The candidate sentence may be adopted by calculating the number of sentences paired with each related word based on the pair, and removing the pair based on the number.

The labeler calculates a labeling similarity between the topic and the candidate sentence based on a probability distribution of the related words included in the candidate sentence and the pair, and extracts a representative sentence corresponding to the topic based on the labeling similarity. can do.

The labeler may calculate the labeling similarity based on the following equation.

 [Equation]

Here, score means the labeling similarity between the topic T and the candidate sentence s (similarity between T and s), t means a word within the topic, and α means a weight indicating the importance of each term, P (t) denotes a probability value for the topic word t, and wordnet (w ', T) is a wordnet-based semantic similarity between the words constituting s that are not in T and each word in topic T ( semantic similarity based on WordNet).

1 shows a schematic block diagram of a topic modeling apparatus.
2 shows a schematic block diagram of the controller shown in FIG. 1.
3 shows an example of the overall operation of the topic modeling apparatus shown in FIG. 1.
4 shows an example of the operation of the prior probability extractor shown in FIG.
FIG. 5 shows an example of an operation in which the prior probability extractor shown in FIG. 2 calculates topic consistency.
FIG. 6 shows an example of an operation in which the prior probability extractor shown in FIG. 2 adopts the optimized prior probability.
FIG. 7 illustrates an example of an operation in which the topic refiner illustrated in FIG. 2 divides a topic.
FIG. 8 illustrates an example of an operation in which the topic refiner shown in FIG. 2 clusters association terms.
FIG. 9 illustrates an example of an operation in which the topic refiner illustrated in FIG. 2 calculates internal similarity and cross similarity.
FIG. 10 shows an example of an operation in which the topic refiner illustrated in FIG. 2 merges a topic.
FIG. 11A illustrates an example of an operation in which the topic refiner illustrated in FIG. 2 calculates an association between a topic and a sentence.
FIG. 11B illustrates another example of an operation in which the topic refiner illustrated in FIG. 2 calculates an association between a topic and a sentence.
FIG. 11C illustrates another example in which the topic refiner illustrated in FIG. 2 calculates an association between a topic and a sentence.
FIG. 12 illustrates an example of an operation in which the topic refiner illustrated in FIG. 2 calculates cosine similarity between sentences.
FIG. 13 shows an example of an operation in which the topic refiner illustrated in FIG. 2 generates a bipartite graph.
FIG. 14 shows an example of an operation in which the topic refiner illustrated in FIG. 2 calculates similarity between topics.
FIG. 15 shows an example of the results of the topic refiner shown in FIG. 2 clustering similarities between topics.
FIG. 16 illustrates an example of an operation in which the labeler illustrated in FIG. 2 automatically labels a topic.
FIG. 17 shows an example of an operation in which the labeler illustrated in FIG. 2 adopts a candidate sentence.
FIG. 18 illustrates an example of an operation in which the labeler illustrated in FIG. 2 extracts a sentence representing a topic.
19 is a flowchart of an operation of the topic modeling apparatus illustrated in FIG. 1.

Hereinafter, exemplary embodiments will be described in detail with reference to the accompanying drawings. However, various changes may be made to the embodiments so that the scope of the patent application is not limited or limited by these embodiments. It is to be understood that all changes, equivalents, and substitutes for the embodiments are included in the scope of rights.

The terminology used herein is for the purpose of description and should not be construed as limiting. Singular expressions include plural expressions unless the context clearly indicates otherwise. In this specification, terms such as "comprise" or "have" are intended to indicate that there is a feature, number, step, action, component, part, or combination thereof described on the specification, and one or more other features. It is to be understood that the present invention does not exclude the possibility of the presence or the addition of numbers, steps, operations, components, components, or a combination thereof.

Terms such as first or second may be used to describe various components, but the components should not be limited by the terms. The terms are for the purpose of distinguishing one component from another component only, for example, without departing from the scope of the rights according to the concepts of the embodiment, the first component may be named a second component, and similarly The second component may also be referred to as the first component.

Unless defined otherwise, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art. Terms such as those defined in the commonly used dictionaries should be construed as having meanings consistent with the meanings in the context of the related art and shall not be construed in ideal or excessively formal meanings unless expressly defined in this application. Do not.

In addition, in the description with reference to the accompanying drawings, the same components regardless of reference numerals will be given the same reference numerals and duplicate description thereof will be omitted. In the following description of the embodiment, when it is determined that the detailed description of the related known technology may unnecessarily obscure the gist of the embodiment, the detailed description thereof will be omitted.

1 shows a schematic block diagram of a topic modeling apparatus.

Referring to FIG. 1, a topic modeling apparatus 10 may extract and refine a topic from an object. The topic modeling apparatus 10 may automatically label a topic by extracting representative sentences of the refined topic. The topic modeling apparatus 10 may extract a topic by using latent Dirichlet Allocation (LDA).

An object may mean a collection of information consisting of letters. Objects can include documents such as books and articles on the Internet. The object may include a plurality of sentences. Topics can mean specific objects that can be discussed or written. Topics may include a plurality of sentences. The sentence may include a plurality of words.

The topic modeling apparatus 10 may provide a method for easily using the LDA by a general user who lacks statistical knowledge or domain knowledge. The topic modeling apparatus 10 may obtain an optimal prior probability initial value for a large text document set given as an input, and automatically refine the unrefined topic sets to form an optimal set of topics. The topic modeling apparatus 10 may automatically perform labeling to easily know the meaning of the extracted topic.

The topic modeling apparatus 10 may perform automated topic modeling so that LDA can be quickly executed on the entire population rather than a sample by using big data technology, and an ordinary user who is not a domain expert can easily execute LDA. .

The topic modeling apparatus 10 includes a receiver 100 and a controller 200.

The receiver 100 may receive an object. The receiver 100 may output the received object to the controller 200.

The controller 200 may automatically extract the topic from the object and refine the label to automatically label the topic.

2 shows a schematic block diagram of the controller shown in FIG. 1.

Referring to FIG. 2, the controller 200 may include a prior probability extractor 210, a topic refiner 230, and a labeler 250.

The prior probability extractor 210 may extract the topic from the object to optimize the prior probability associated with the distribution of the topic. The prior probability extractor 210 may extract a topic from the object and calculate an average topic coherence according to the change of the prior probability for the topic.

The prior probability extractor 210 can adapt the optimized prior probability based on the average topic consistency. The prior probability extractor 210 may output the optimized prior probability to the topic refiner 230. The operation of the prior probability extractor 210 will be described in detail with reference to FIGS. 4 to 6.

The topic refiner 230 may refine the topic by splitting or merging the extracted topics based on prior probabilities. The topic refiner 230 may split a topic based on semantic similarity of the related words included in the topic or merge the topics based on the similarity between the topics based on the related words.

The topic refiner 230 may output the purified topic to the labeler 250. The operation of topic refiner 230 will be described in detail with reference to FIGS. 7-15.

The labeler 250 may automatically label the refined topic by extracting sentences representing the refined topic. The labeler 250 may select a sentence highly related to the topic as a candidate sentence, and extract a sentence representing the topic based on a probability distribution of related words included in the candidate sentence to label the topic. The operation of the labeler 250 will be described in detail with reference to FIGS. 16-18.

3 shows an example of the overall operation of the topic modeling apparatus shown in FIG. 1.

Referring to FIG. 3, the prior probability extractor 210 may extract a topic using an LDA from an object. The prior probability extractor 210 may extract a topic while changing the prior probability using the LDA. Prior probability extractor 210 may extract (and obtain) an optimized prior probability by calculating the average topic consistency. The prior probability extractor 210 may output the optimized prior probability to the topic refiner 210 or the labeler 270.

Topic refiner 230 may refine the topic based on the optimized prior probabilities. Topic refiner 230 may refine by splitting or merging topics. The topic refiner 230 may output the purified topic to the labeler 270.

The labeler 270 may automatically label the refined topic. The labeler 270 may automatically label the topic by adopting the candidate sentence from the refined topic and extracting the representative sentence from the candidate sentence.

Hereinafter, the operation of the prior probability extractor will be described in detail with reference to FIGS. 4 to 6.

4 shows an example of the operation of the prior probability extractor shown in FIG.

Referring to FIG. 4, the prior probability extractor 210 may extract (eg, obtain) an optimized prior probability by extracting a topic while changing prior probabilities, and calculating topic consistency.

The dictionary probability extractor 210 may extract the topic by calculating a probability that words associated with the topic belong to the topic, respectively, using the LDA method. Since LDA uses a Bayesian probability model, initial values of prior probability α and β may be required.

When the accurate prior probability initial value is used, the execution time of the topic modeling apparatus 10 may be shortened and the topic result may be improved. α may indicate a topic distribution in the document, and β may indicate how probable the words in the topic are distributed.

The prior probability extractor 210 may experimentally find a value optimizing the prior probabilities α and β through topical coherence. Since the prior probability extractor 210 finds the prior probability experimentally, it is possible to obtain an optimal prior probability for various data, and to accurately estimate the population through big data technology to derive an accurate result.

The prior probability extractor 210 may extract the topic from the object. In the example of FIG. 4, the topic may include performance, driving, design, and service.

The prior probability extractor 210 may calculate the average topic consistency for each topic while changing the prior probabilities to a predetermined value for the extracted topic. For example, the average topic consistency can be calculated by changing the α or β value by 0.05.

The prior probability extractor 210 may adopt a prior probability having the highest average topic consistency as the optimized prior probability.

The prior probability extractor 210 may optimize the prior probabilities using a software framework. For example, the prior probability extractor 210 may optimize the prior probability using a map reduce framework.

The MapReduce framework can support parallel processing of large data of more than petabytes in a cluster environment. The MapReduce framework can be configured based on a function called Map and Reduce.

The map and reduce steps have key-value pairs as inputs and outputs, and the type can be selected directly by the programmer. Map and reduce functions can also be written directly by the programmer.

The map step is a task of grouping scattered data into related data classifications in key-value form, and the reduce step may be a task of removing redundant data and extracting desired data from the output values of the map.

The prior probability extractor 210 executes the map reduce program, and simultaneously divides the input data by the number of data nodes (the number of mappers) and reduces the execution time by the number of 1 / data nodes. MapReduce is used as open source software in Apache Hadoop and can be written to be applicable to Java, C ++, and other languages. For example, the MapReduce program may be implemented based on the Java language.

FIG. 5 shows an example of an operation in which the prior probability extractor shown in FIG. 2 calculates topic consistency.

Referring to FIG. 5, the prior probability extractor 210 may calculate the average topic consistency while applying different prior probabilities to the topics. The prior probability extractor 210 may calculate average topic consistency using data nodes and mappers corresponding to different prior probabilities.

The prior probability extractor 210 may consider topic consistency to extract the optimal prior probability. Topic consistency may refer to an evaluation measure used to evaluate the effectiveness of a modeling technique, and may indicate how similarly the related words (words) making up a topic have meaning.

The prior probability extractor 210 may calculate the average topic consistency based on Equation 1.

Here, TC may mean topic consistency, and may mean semantic similarity value of the related words wl and wf in the sim (wl, wf) topic. For example, semantic similarity may include similarity based on WordNet.

The numerator of Equation 1 refers to the total similarity of each pair of words in a given topic, and the denominator may be a value for the mean. WordNet may mean a database for identifying semantic relationships between words.

As the meaning similarity value approaches 0, the meanings of the two words may not be relevant. For example, "calm" and "peaceful" have different spellings, but they have similar meanings. The similarity between "calm" and "peaceful" using WordNet can be close to one.

That is, the higher the WordNet semantic similarity between the words constituting the topic, the higher the topic consistency of the corresponding topic.

In the example of FIG. 5, the prior probability extractor 210 increases each parameter by 0.05 from prior probability α = 0, β = 0 to α = 1, β = 1 based on the input text document, and increases the dictionary. In each case of probability, a MapReduce based topic extraction algorithm (Mahout LDA) may be executed. In the example of FIG. 5, the total number of cases of prior probabilities is 441 in total.

The prior probability extractor 210 executes a MapReduce based topic extraction algorithm to extract 100 topics for each pre-probability case, and the extracted topics may have relevant words and probability values for each topic. .

Prior probability extractor 210 may calculate the average topic consistency for each prior probability. The dictionary probability extractor 210 may calculate topic consistency of all topics by using WordNet semantic similarity between words in each mapper.

The prior probability extractor 210 may calculate an average of the 100 identified topic consistency values when the 100 topic consistency values corresponding to one case of the prior probability are identified. The prior probability extractor 210 may return the average topic consistency value along with the prior probability case.

The prior probability extractor 210 may collect all the output results and use them as an input of the reducer after the operation of the mappers calculating the average topic consistency value is completed.

In the example of FIG. 5, the prior probability extractor 210 may perform the calculation by dividing by the number of mappers when calculating the average topic consistency value of all 441 prior probability cases. Since the prior probability extractor 210 executes the operations of the respective mappers simultaneously in parallel, the execution speed may be faster than that of the single program based prior probability optimization algorithm.

FIG. 6 shows an example of an operation in which the prior probability extractor shown in FIG. 2 adopts the optimized prior probability.

Referring to FIG. 6, the prior probability extractor 210 may adopt a prior probability having the highest average topic consistency as an optimized prior probability.

The prior probability extractor 210 may compare all the average topic consistency values of all prior probability cases in the reducer and adopt the prior probability having the highest average topic consistency value as the optimized prior probability. The prior probability extractor 210 may output the optimized prior probability to the topic refiner 230.

Hereinafter, the operation of the topic refiner 230 will be described in detail with reference to FIGS. 7 to 15.

FIG. 7 illustrates an example of an operation in which the topic refiner illustrated in FIG. 2 divides a topic.

Referring to FIG. 7, optimizing the prior probabilities of LDA and performing topic modeling generally show excellent results, but since the topic is extracted by a mechanical and statistical algorithm, there may be a case in which a result that is different from the actual one is output. have.

There are two main problems, the first being a group of related words having two or more different meanings in one topic, and the second being a group containing related words with similar meanings in several topics.

The topic refiner 230 may divide a topic into two or more topics when there are groups of related words having two or more different meanings in one topic. In addition, the topic refiner 230 may merge a set of related words constituting topics having similar meanings.

The topic refiner 230 clusters the related words to generate a plurality of clusters, and calculates the internal similarity by averaging the semantic similarities between the related words in each cluster.

The topic refiner 230 may calculate cross similarity by averaging semantic similarities between related words of different clusters, and divide the plurality of clusters into different topics based on the internal similarity and the cross similarity.

FIG. 8 illustrates an example of an operation in which the topic refiner shown in FIG. 2 clusters association terms.

Referring to FIG. 8, the topic refiner 230 may measure semantic similarity based on WordNet between related words in a topic. The topic refiner 230 may automatically cut-off the associations through hierarchical clustering. The example of FIG. 8 may represent an example of automatic cutoff according to hierarchical clustering.

In the case of (1), it may mean that there are different topics having two meanings in a topic, and in this case, the topic refiner 230 may cluster the related words into two. In the case of (2), the topic refiner 230 may cluster the related words into four groups.

FIG. 9 illustrates an example of an operation in which the topic refiner illustrated in FIG. 2 calculates internal similarity and cross similarity.

Referring to FIG. 9, the topic purifier 230 may calculate an inner similarity and an outer similarity to perform clustering with automatic cutoff.

The internal similarity may mean an average of similarity similarities of a within-cluster, and the cross similarity may mean an average of similarity between cross-clusters.

That is, the topic refiner 230 may calculate internal similarity by measuring semantic similarities between related words included in each cluster and calculating averages thereof. The topic refiner 230 may calculate cross similarity by measuring semantic similarities between related words included in different clusters and calculating their averages.

In the example of FIG. 9, the topic purifier 230 measures the semantic similarity of w1 and w2 included in cluster 1, and calculates the meaning of w3 and w4 included in cluster 2, in order to calculate the internal similarity between cluster 1 and cluster 2. The average can be calculated by measuring the similarity. In FIG. 9, the internal similarity may be 0.85.

The topic refiner 230 measures cross-similarity by measuring semantic similarities of w1 included in cluster 1 and w3 included in cluster 2, and measuring and averaging the semantic similarities of w1-w4, w2-w3 and w2-w4 in the same manner. Can be calculated. In FIG. 9, the cross similarity may be 0.25.

The higher internal similarity and lower cross similarity means that the two clusters are well divided. That is, the topic purifier 230 may find a cutoff having a large internal similarity value and a small cross similarity value for the best automatic cutoff.

The topic refiner 230 may calculate probability values for each cutoff as shown in Table 1.

Cut off Internal Similarity (A) Cross similarity (B)

(One) 0.76 0.2 3.8 0.52 (2) 0.88 0.25 3.52 0.48

The topic refiner 230 can take the inverse of the cross similarity and multiply it by the internal similarity in order to probabilistically calculate the cut off. Taking the inverse of the cross similarity is because the smaller the value of the cross similarity, the more the topic should be divided. The topic refiner 230 may divide a topic by adopting a cutoff such that the rightmost probability value of Table 1 has the largest value.

FIG. 10 shows an example of an operation in which the topic refiner illustrated in FIG. 2 merges a topic.

Referring to FIG. 10, the topic refiner 230 may merge a plurality of topics into one topic when different topics include similar association words. In the example of FIG. 10, Topic 0 and Topic 1 are topics having similar meanings, but may be divided into separate topics.

The topic refiner 230 may merge topics having similar meanings into one topic through a topic merging algorithm.

The topic refiner 230 may calculate an association between a topic and a sentence included in the topic based on the related words. The topic refiner 230 replaces the relevant words with the sentences most relevant to the topic based on the associations, and calculates a cosine similarity between the sentences associated with the topic and the sentences associated with the different topics to generate a bipartite graph. can do.

Topic refiner 230 may calculate topic similarity between the topic and the different topics, and merge topics and different topics based on topic similarity.

FIG. 11A illustrates an example of an operation in which the topic refiner illustrated in FIG. 2 calculates an association between a topic and a sentence, and FIG. 11B illustrates another example of an operation in which the topic refiner illustrated in FIG. 2 calculates an association between a topic and a sentence. 11C illustrates another example of an operation in which the topic refiner illustrated in FIG. 2 calculates an association between a topic and a sentence.

Referring to FIGS. 11A to 11C, the topic refiner 230 may extract the top k sentences by extracting a topic and calculating a probability value for each sentence used in the topic through the related words of each topic.

The topic refiner 230 may add probabilities of related words appearing in sentences to be analyzed for relevance among related words included in the topic for each topic, and divide them by the number of related words included in the topic.

For example, in order to measure the correlation between the topic 0 and the sentence S1 in FIG. 11A, the topic refiner 230 includes a probability distribution of the related words w1 and w3 included in the sentence S1 among the related words w1, w2, and w3 included in the topic 0. The value can be added and divided by the number of related words (or related words) of topic 0.

Similarly, the topic refiner 230 may measure an association between topic 0 and sentences S2 and S3. The topic refiner 230 may perform the same process for other topics. The topic refiner 230 may replace the related words representing each topic with k highly related sentences based on the measured correlation.

12 illustrates an example of an operation in which the topic refiner illustrated in FIG. 2 calculates cosine similarity between sentences, and FIG. 13 illustrates an example of an operation in which the topic refiner illustrated in FIG. 2 generates a bipartite graph.

12 and 13, the topic refiner 230 may calculate a cosine similarity between representative sentences of different topics. Topic purifier 230 may generate a bipartite graph based on cosine similarity.

In the example of FIG. 12, the topic refiner 230 may combine the related word of topic 0 and the sentence S1 replaced with the related word of topic 1 and the word included in the replaced S2. The topic refiner 230 may calculate the cosine similarity after vectorizing the combined association word and each sentence by inserting 1 if the included words overlap and inserting 0 if not included.

The topic refiner 230 may display two topics to be compared in a bipartite graph based on the calculated cosine similarity. The topic refiner 230 may calculate the cosine similarity for all extracted topics and display the results in a bipartite graph.

FIG. 14 illustrates an example of an operation in which the topic refiner illustrated in FIG. 2 calculates similarity between topics, and FIG. 15 illustrates an example of a result of the topic refiner illustrated in FIG. 2 clustering similarities between topics.

The topic refiner 230 may calculate topic similarity of all topic pairs using the Maximum Weighted Bipartite Matching (MWWM) method, which is a graph optimization algorithm that allocates the highest value based on the bipartite graph described above.

The topic refiner 230 may determine whether a topic pair is merged by clustering topic similarity values of each topic pair by a k-average method.

In the example of FIG. 15, topic refiner 230 shows the result of applying topic similarity values of all topic pairs to k-mean clustering after applying MWBM to a bipartite graph, and not merging with topic pairs that are merged (Tx = Ty). The pair Tx ≠ Ty can be determined.

FIG. 16 illustrates an example of an operation in which the labeler illustrated in FIG. 2 automatically labels a topic.

Referring to FIG. 16, the labeler 270 may label a topic by determining the meanings of main association terms constituting the extracted topic and naming the topic.

Topic labeling may require sufficient domain knowledge of large text document sets that are given as input to associations and topics.

However, because end users do not have enough knowledge of the domain, it may be difficult to do topic labeling manually. The labeler 270 may perform topic labeling automatically, thereby enabling fast and accurate topic labeling.

The labeler 270 may calculate a score through filtering and blocking methods, and use the Top-K sentences as a topic label. The score may mean labeling similarity.

When labeling all the related words and sentences included in the topic, it may take considerable time to calculate the labeling similarity. For example, assuming 100 topics, 500 associative words, and 100,000,000 sentences exist for each topic, 5,000,000,000,000 calculations must be performed to perform a significant amount of time.

The labeler 270 may reduce the time required for labeling by adopting the candidate sentence through filtering to reduce the time required, and extracting the representative sentence representing the topic through blocking.

FIG. 17 shows an example of an operation in which the labeler illustrated in FIG. 2 adopts a candidate sentence.

Referring to FIG. 17, the labeler 270 removes association words having a small probability value from a topic based on a probability distribution of association words included in a topic, and compares a sentence included in the topic and a pair of related words included in the topic. Can be generated.

The labeler 270 may calculate a number of sentences paired with each related word based on a pair of sentences and related words, and adopt a candidate sentence by removing a pair of sentences and related words based on the number of sentences. have.

The labeler 270 may perform filtering that selects and provides sentences related to a topic in the entire sentence as candidate sentences. The labeler 270 may minimize unnecessary operations through filtering.

Since the association word having a low probability distribution is low in relation to the topic, the labeler 270 may drop an association word having a probability distribution value lower than a specific value in the topic.

The labeler 270 can quickly scan a sentence and generate a pair of sentences and associated words. The labeler 270 may build an inverted index of each related word through a pair of sentences and words, and count the number. At this time, if one sentence overlaps a plurality of related words, the labeler 270 may build only the inverse index of the upper related word, and count may be normally performed.

The labeler 270 may drop (remove) the related word and the sentence having a small number of counts. The labeler 270 may select sentences that are highly related to the topic by removing sentences that are not related, and minimize unnecessary operations.

The candidate sentence may mean a sentence remaining in the topic after the labeler 270 is dropped.

FIG. 18 illustrates an example of an operation in which the labeler illustrated in FIG. 2 extracts a sentence representing a topic. Referring to FIG. 18, the labeler 270 may calculate a labeling similarity between a topic and a candidate sentence based on a probability distribution of association words included in the candidate sentence and a pair of sentences and association words.

The labeler 270 may extract a representative sentence corresponding to the topic based on the labeling similarity.

The labeler 270 may calculate labeling similarity based on Equation 2.

Here, score may mean the similarity between T and s of the topic T and the candidate sentence s, t may mean a word within the topic, and α may mean a weight indicating the importance of each term. .

P (t) is the probability value for the topic word t, and wordnet (w ', T) is the wordnet-based semantic similarity between the words w' that are not in T and the words in topic T that make up s. (semantic similarity based on WordNet).

The labeler 270 may calculate the labeling similarity through map reduction based parallel processing by performing a mapper by inputting the record of the inverted index described above.

The labeler 270 may use the records of each inverted index as input to calculate the labeling similarity. In mapper n, you can calculate labeling similarity based on sentences and associations for topic n.

Thereafter, the labeler 270 may sort the sentences in descending order through shuffling. Finally, the labeler 270 may output a sentence of Top-K for each topic n in the reducer. The Top-K sentence may mean a representative sentence for topic n.

The labeler 270 may extract Top-k sentences having a high labeling similarity for each topic through the above-described process, and may extract the sentences as sentences representing a topic that is highly related to the topic.

The labeler 270 automatically performs topic labeling using the extracted representative sentence, and can easily show at a glance what the topic means.

19 is a flowchart of an operation of the topic modeling apparatus illustrated in FIG. 1.

Referring to FIG. 19, the receiver 100 may receive an object. In the example of FIG. 19, the object received by the receiver 100 may include a post of a website. In the example of FIG. 19, the receiver may receive a post of a website (for example, a baby dream) corresponding to a domain of a car.

The prior probability extractor 210 may collect, refine, and store the object in a database. For example, the dictionary probability extractor 210 may collect, refine, and store a large amount of text documents on a website called baby dream.

The prior probability extractor 210 may extract the topic from the object and calculate the topic consistency according to the change of the prior probability. The prior probability extractor 210 may extract (obtain) an optimized prior probability α, β based on topic consistency. α may indicate a topic distribution in the document, and β may indicate how probable the words in the topic are distributed.

The topic refiner 250 may refine topics by merging topics with overlapping meanings into one topic and dividing related terms having different meanings into one topic into a plurality of topics.

The labeler 270 may adopt the candidate sentence based on the probability distribution of the related words and the similarity between the related word and the sentence, and extract the representative sentence representing the topic from the candidate sentence. The labeler 270 may automatically perform topic labeling by extracting K sentences having the highest labeling similarity as representative sentences.

Through this, the topics that are frequently dealt with in a large document set given by the topic modeling apparatus 10 and the meanings of the topics may be easily grasped.

The method according to the embodiment may be embodied in the form of program instructions that can be executed by various computer means and recorded in a computer readable medium. The computer readable medium may include program instructions, data files, data structures, and the like, alone or in combination. The program instructions recorded on the media may be those specially designed and constructed for the purposes of the embodiments, or they may be of the kind well-known and available to those having skill in the computer software arts. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tape, optical media such as CD-ROMs, DVDs, and magnetic disks, such as floppy disks. Magneto-optical media, and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, flash memory, and the like. Examples of program instructions include machine code, such as produced by a compiler, as well as high-level language code that can be executed by a computer using an interpreter or the like. The hardware device described above may be configured to operate as one or more software modules to perform the operations of the embodiments, and vice versa.

The software may include a computer program, code, instructions, or a combination of one or more of the above, and configure the processing device to operate as desired, or process it independently or collectively. You can command the device. Software and / or data may be any type of machine, component, physical device, virtual equipment, computer storage medium or device in order to be interpreted by or to provide instructions or data to the processing device. Or may be permanently or temporarily embodied in a signal wave to be transmitted. The software may be distributed over networked computer systems so that they may be stored or executed in a distributed manner. Software and data may be stored on one or more computer readable recording media.

Although the embodiments have been described with reference to the accompanying drawings as described above, those skilled in the art may apply various technical modifications and variations based on the above. For example, the described techniques may be performed in a different order than the described method, and / or components of the described systems, structures, devices, circuits, etc. may be combined or combined in a different form than the described method, or other components. Or even if replaced or substituted by equivalents, an appropriate result can be achieved.

Therefore, other implementations, other embodiments, and equivalents to the claims are within the scope of the following claims.

Claims (20)

Extracting a topic from an object to optimize prior probabilities associated with the distribution of the topic;
Refining the topic by dividing or merging the extracted topic based on the prior probability; And
Automatically labeling the refined topic by extracting a sentence representative of the refined topic
Including,
The optimizing step,
Extracting the topic from the object;
Calculating an average topic coherence for the topic according to a change in the prior probability; And
Adopting an optimized prior probability based on the average topic consistency
Topic modeling method comprising a.

delete The method of claim 1,
The calculating step,
Calculating the average topic consistency based on the following equation
Topic modeling method comprising a.
[Equation]

Here, TC means topic consistency, and means semantic similarity values of the related words w _l and w _{f in} the sim (w _l , w _f ) topic.
The method of claim 1,
The purification step,
Dividing the topic based on semantic similarity of the related words included in the topic; or
Merging the topics based on the similarity between the topics calculated from the association terms
Topic modeling method comprising a.
The method of claim 4, wherein
The dividing step,
Clustering the association words to generate a plurality of clusters;
Calculating a within-cluster similarity by averaging the semantic similarities between the related words in each cluster;
Calculating cross-clusters similarity by averaging semantic similarities between associations of different clusters; and
Dividing a plurality of clusters into different topics based on the internal similarity and the cross similarity
Topic modeling method comprising a.
The method of claim 4, wherein
Merging step,
Calculating an association between the topic and a sentence for the topic based on the association terms;
Replacing the association words with the sentence based on the association;
Generating a bipartite graph by calculating cosine similarity between the sentences and sentences included in different topics;
Calculating a topic similarity between the topic and the different topic; And
Merging the topic with the different topic based on the topic similarity
Topic modeling method comprising a.
The method of claim 1,
The labeling step,
Adopting a sentence highly related to the topic as a candidate sentence; And
Labeling the topic by extracting a sentence representative of the topic based on a probability distribution of related words included in the candidate sentence
Topic modeling method comprising a.
The method of claim 7, wherein
The step of adopting,
Removing association words having a small probability value from the topic based on a probability distribution of the association words included in the topic;
Generating a pair of sentences included in the topic and a related word included in the topic;
Calculating the number of sentences paired with each association word based on the pair; And
Removing the pair based on the number to adopt the candidate sentence
Topic modeling method comprising a.
The method of claim 8,
The labeling step,
Calculating a labeling similarity between the topic and the candidate sentence based on a probability distribution of the related words included in the candidate sentence and the pair;
Extracting a representative sentence corresponding to the topic based on the labeling similarity
Topic modeling method comprising a.
The method of claim 9,
Computing the labeling similarity,
Calculating the labeling similarity based on the following equation
Topic modeling method comprising a.
[Equation]

Here, score means the labeling similarity between the topic T and the candidate sentence s (similarity between T and s), t means a word within the topic, and α means a weight indicating the importance of each term, P (t) denotes a probability value for the topic word t, and wordnet (w ', T) is a wordnet-based semantic similarity between the words constituting s that are not in T and each word in topic T ( semantic similarity based on WordNet).
A receiver receiving the object; And
A controller that extracts and refines a topic from the object to automatically label the topic;
Including,
The controller,
A prior probability extractor for extracting the topic from the object to optimize a prior probability associated with the distribution of the topic;
A topic refiner for refining the topic by dividing or merging the extracted topic based on the prior probability; And
A labeler that automatically labels the refined topic by extracting sentences representing the refined topic
Including,
The prior probability extractor,
Extract the topic from the object, calculate an average topic coherence for the topic according to the change of the prior probability, and adopt an optimized prior probability based on the average topic consistency
Topic modeling device.

delete The method of claim 11,
The prior probability extractor,
Computing the average topic consistency based on the following equation
Topic modeling device.
[Equation]

Here, TC means topic consistency and semantic similarity value of the related words w _l and w _{f in} the sim (w _l , w _f ) topic.
The method of claim 11,
The topic purifier,
Dividing the topic based on semantic similarity of association terms included in the topic or merging the topic based on similarity between the topics based on the association terms.
Topic modeling device.
The method of claim 14,
The topic purifier,
Clustering the related words to generate a plurality of clusters, calculating the internal similarity by averaging the semantic similarities between the related words in each cluster, calculating the cross similarity by averaging the semantic similarities between the related words of the different clusters, Dividing the plurality of clusters into different topics based on the internal similarity and the cross similarity
Topic modeling device.
The method of claim 14,
The topic purifier,
Calculate an association between the topic and a sentence for the topic based on the association words, replace the association words with the sentence based on the association, and calculate cosine similarity between the sentence and sentences contained in a different topic Generate a Bipartite Graph, calculate topic similarity between the topic and the different topics, and merge the topic and the different topics based on the topic similarity.
Topic modeling device.
The method of claim 11,
The labeler is,
Labeling the topic by adopting a sentence highly related to the topic as a candidate sentence, extracting a sentence representing the topic based on a probability distribution of related words included in the candidate sentence
Topic modeling device.
The method of claim 17,
The labeler is,
Based on a probability distribution of related words included in the topic, removing related words having a small probability value from the topic, generating a sentence included in the topic and a pair of related words included in the topic, and based on the pair Calculating the number of sentences paired with each association word, and removing the pair based on the number to adopt the candidate sentence.
Topic modeling device.
The method of claim 18,
The labeler is,
Calculating a labeling similarity between the topic and the candidate sentence based on a probability distribution of the related words included in the candidate sentence and the pair, and extracting a representative sentence corresponding to the topic based on the labeling similarity;
Topic modeling device.
The method of claim 19,
The labeler is,
Computing the labeling similarity based on the following equation
Topic modeling device.
[Equation]

Here, score means the labeling similarity between the topic T and the candidate sentence s (similarity between T and s), t means a word within the topic, and α means a weight indicating the importance of each term, P (t) denotes a probability value for the topic word t, and wordnet (w ', T) is a wordnet-based semantic similarity between the words constituting s that are not in T and each word in topic T ( semantic similarity based on WordNet).

Images