ConceptNet-based Information Retrieval Query Expansion Method

Abstract

The present invention discloses an information retrieval query expansion method based on ConceptNet, comprising the following steps: S1: Use BM25 for retrieval to obtain a set of feedback documents and terms; S2: Obtain a set of semantic descriptions of terms by taking the semantic expansion set of terms from ConceptNet; S3: Measure the semantic relevance between candidate expansion terms and the query; S4: Obtain the weights of candidate expansion terms; S5: Construct a new query and conduct a secondary retrieval. The beneficial effect of the present invention is: It can optimize query expansion to improve the average precision rate of retrieval results.

Description

Technical Field

[0001] The present invention relates to the technical field of information retrieval methods, and particularly to an information retrieval query expansion method based on ConceptNet. Background Art

[0002] The main process of relevance feedback is as follows: When a user wants to query a content, at first, the user doesn't know whether the query to be input can find relevant results. However, it is easy for the user to judge whether the given results are relevant. In the process of the user continuously submitting feedback results, the system optimizes the representation of the query based on the user's feedback. Then, a new query is used for retrieval until satisfactory results for the user are obtained. The process of relevance feedback may require the user to perform multiple feedbacks and repeatedly optimize the query according to the user's feedback. Although the relevance feedback method can significantly improve the retrieval effect, for most users, they may not want the system to collect their information. The pseudo-relevance feedback method automates the above process of user intervention. The pseudo-relevance assumption is that the top N documents in the ranking of the initial retrieval results are relevant to the query, and these documents are called feedback documents or pseudo-relevant documents. Then, expansion terms are extracted from the feedback documents to design a query expansion method to redefine the representation of the original query. Subsequently, the second round of retrieval is carried out. A large number of studies have proved that the pseudo-relevance feedback method can effectively improve the retrieval results through automatic query expansion.

[0003] In the prior art, most pseudo-relevance feedback models select expansion terms from aspects such as frequency, co-occurrence frequency, and proximity. This way assumes that the terms that are close to the query and repeatedly co-occur may have the same semantics as the query. However, these methods all indirectly obtain the expansion terms that may be relevant to the query and do not directly measure the terms with similar meanings to the query expression. A large number of studies have found that if query-related knowledge can be used as one of the sources of the expansion term set, it helps to solve the problem of term mismatch between the query and the document. The main idea of query expansion is that the expanded query can better describe the user's needs and avoid the semantic ambiguity problem caused by polysemy. If the wrong expansion terms are selected, the expanded query may deviate from the semantics of the original query, ultimately resulting in a decrease in the accuracy of the retrieval results. Filtering query terms by measuring the semantic relationship between the expansion terms and the query helps to solve the problem of semantic deviation and thus improve the correctness of the retrieval results. These methods cannot measure the semantic relevance between the query and the expansion terms. These pseudo-relevance feedback methods all select expansion terms from the feedback documents. This method has an obvious disadvantage: if several documents marked as relevant are actually irrelevant, this method will introduce errors in the first step of the query expansion process and may lead to the generation of expansion terms that are not relevant to the query in the query expansion process.

[0004] For example, an "Automatic Image Annotation Method Integrating Pseudo-Relevance Feedback and Retrieval Technology" disclosed in a Chinese patent document, with the publication number: CN101075263A and the application date: June 28, 2007. This invention improves the retrieval performance and the accuracy of annotation, greatly improves the scalability of annotation, but there is a problem of relatively low average correct rate of retrieval results. Summary of the Invention

[0005] Aiming at the deficiency of the relatively low average correct rate of retrieval results in the prior art, the present invention proposes an information retrieval query expansion method based on ConceptNet, which can optimize query expansion to improve the average correct rate of retrieval results.

[0006] The following is the technical solution of the present invention. The information retrieval query expansion method based on ConceptNet includes the following steps:

[0007] S1: Use BM25 for retrieval to obtain a set E of feedback documents and terms;

[0008] S2: Obtain a semantic extension set D of terms from ConceptNet to obtain a semantic description set C of terms;

[0009] S3: Measure the semantic relevance between candidate expansion terms and the query;

[0010] S4: Calculate the weights of the candidate expansion terms;

[0011] S5: Construct a new query and perform a secondary retrieval.

[0012] In this solution, the collaborative filtering method of the query semantic description set obtained from the ConceptNet knowledge base and the term set obtained from the feedback documents is used to initially select candidate expansion terms, which helps to reduce the error of query expansion caused by incorrect selection of feedback documents. A method for directly measuring the semantic relevance between the query and the terms is proposed. The expansion terms are selected by combining semantic matching and relevance matching between terms, and a second-round retrieval is performed, which is statistically relevant to the query to ensure the importance of the terms in the document set; and is semantically relevant to the query to ensure consistency with the theme of the query. It is beneficial to find better expansion terms to supplement the semantics of the query and improve the correct rate of the final retrieval results.

[0013] Preferably, in step S1, the calculation formula for the score of the document is as follows:

[0014]

[0015]

[0016]

[0017] In the above formula, Score(Q, d) is the document score, TF is the normalized representation of the document length, avdl is the average length of the documents, k1 and k3 are constants, qtf is the frequency of the query term q, b is the adjustment factor, dl is the document length, N′ is the number of all documents in the index, and df t is the number of documents containing the term t in all documents, and IDF(t) is the importance of t in the feedback documents.

[0018] In this solution, the document scores of all documents are sorted to facilitate screening target documents to form feedback documents, which is convenient for subsequent calculations and improves the accuracy of the final retrieval results.

[0019] Preferably, in step S1, the set E is a candidate pseudo-relevant term set composed of all terms in the top 10 documents with the highest document scores.

[0020] In this solution, all terms of the target document form a candidate pseudo-relevant term set, which is convenient for forming a semantic expansion set and improves the accuracy of the final retrieval results.

[0021] Preferably, in step S2, all relation triples with the query term as the head node or the tail node are taken, and all head and tail nodes of the relation triples form a set D, and the set C is the intersection of the set D and the set E.

[0022] In this solution, the set D is composed of a set of all terms related to the relation triples and the query term, and the set C is the intersection of the set D and the set E. The collaborative filtering method of the query semantic description set obtained from the ConceptNet knowledge base and the term set obtained from the feedback documents preliminarily selects candidate expansion terms, which helps to reduce the error of query expansion caused by the wrong selection of feedback documents and improves the accuracy of the final retrieval results.

[0023] Preferably, in step S3, the calculation formula for measuring the semantic relevance matching score between the candidate expansion term and the query is as follows:

[0024] S(t, Q) = sim(t, Q) × IDF(t);

[0025]

[0026] In the above formula, S(t, Q) is the matching score, sim(t, Q) is the semantic similarity between t and Q, and IDF(t) is the importance of t in the feedback documents. represent the word vectors with the term t as the head node and the term q as the head node respectively, and μ t is the mean of the semantic similarities between all query terms in the query Q and the term t, and σ represents the variance of the semantic similarities between all query terms in the query Q and the term t.

[0027] In this solution, the semantic relevance between the direct measurement query and the extended term is measured. By measuring the similarity of the word vectors of the query and the extended term, the semantic relevance between them is obtained. The extended term is selected by combining semantic matching and relevance matching between terms, and the second-round retrieval is performed, which can identify terms that do not have high frequency or co-occurrence frequency but may be semantically relevant, improving the accuracy of the final retrieval result.

[0028] Preferably, the dimension of the word vector is 300.

[0029] Preferably, in step S4, the weight calculation formula for the candidate extended term is as follows:

[0030]

[0031] λ ∈ {0, 0.1, 0.2,..., 1.0};

[0032] In the above formula, tf(t, d) is the number of occurrences of term t in document d, N is the number of feedback documents, λ is the balance factor, and w t is the weight of the candidate extended term.

[0033] In this solution, the extended term weight calculation method takes into account both the term frequency (or term distribution) and the semantic relevance, and it consists of two parts. The first part represents the importance of the candidate term, which is measured using term frequency-inverse document frequency or language model. It is statistically relevant to the query, ensuring the importance of the term in the document collection; and it is semantically relevant to the query, ensuring consistency with the query topic and improving the accuracy of the final retrieval result.

[0034] Preferably, the values of the relevance matching method and the semantic matching method between terms are normalized, and the balance factor is used to adjust the contributions of the two parts to rank the candidate extended terms.

[0035] In this solution, the values of the two parts of the relevance matching method and the semantic matching method between terms are subjected to MIN-MAX normalization. The balance factor λ is used to adjust the contributions of the two parts, and the candidate extended terms are ranked according to the final values, improving the accuracy of the final retrieval result.

[0036] Preferably, in step S5, the new query calculation formula is as follows:

[0037] Q′ = (1 - α) × Q + α × Q1;

[0038] α ∈ {0, 0.1, 0.2,..., 1.0};

[0039] |Tf| ∈ {10, 20, 30, 50};

[0040] In the above formula, α is the balance factor, Q1 is the vector composed of the top |Tf| expansion terms with the largest weights, Q is the original query vector, and Q′ is the new query vector formed by Q1 and Q.

[0041] In this solution, the balance factor α is used to adjust the contributions of the two parts to construct a new query vector. The new query vector is statistically relevant to the query, ensuring the importance of terms in the document collection; and semantically relevant to the query, ensuring consistency with the query topic. By the semantic relevance matching score and the weights of candidate expansion terms, the accuracy of the final retrieval results is improved.

[0042] Preferably, the new query vector is used for secondary retrieval, the document scores are calculated, and the results of the document ranking are sorted and displayed according to the magnitudes of the document scores.

[0043] In this solution, the new query vector is used to improve the accuracy of the retrieval results, the document scores are recalculated and sorted, and the more optimal search ranking results are intuitively displayed.

[0044] The beneficial effect of the present invention is that it can optimize query expansion to improve the average accuracy of retrieval results. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 Flowchart of the information retrieval query expansion method based on ConceptNet of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0046] The technical solution of the present invention will be further specifically described below through embodiments in conjunction with the drawings.

[0047] Embodiment: As Figure 1 shown, the information retrieval query expansion method based on ConceptNet includes the following steps:

[0048] Step 1: Use BM25 (a retrieval algorithm) for the first-round retrieval to obtain a set E of feedback documents and terms;

[0049] Step 2: Obtain the semantic description set C of terms by taking the semantic expansion set D of terms from ConceptNet (a semantic network knowledge base);

[0050] Step 3: Measure the semantic relevance between candidate expansion terms and the query;

[0051] Step 4: Calculate the weights of all candidate expansion terms;

[0052] Step 5: Construct a new query and perform secondary retrieval.

[0053] In Step 1, BM25 is used for the first-round retrieval to obtain feedback documents, and a set E of feedback documents and terms is obtained. The BM25 model is a simple and effective retrieval method. BM25 sorts documents based on factors such as the importance of query terms in all datasets, the length of documents, and the frequency of terms. The calculation method is as follows:

[0054]

[0055]

[0056]

[0057] In the above formula, Score(Q, d) is the document score, TF is the normalized representation of the document length, avdl is the average length of documents, k1 and k3 are constants, qtf is the frequency of query term q, b is the adjustment factor, dl is the document length, N′ is the number of all documents in the index, df t is the number of documents containing term t in all documents, and IDF(t) is the importance of t in the feedback documents.

[0058] Sort the document scores of all documents from high to low to obtain the top 10 documents, which are called feedback documents or pseudo-relevant documents. Take all the terms in these feedback documents to form the candidate pseudo-relevant term set E.

[0059] In Step 2, take the semantic expansion set D of terms from ConceptNet. ConceptNet expresses the relationships of common-sense knowledge in the form of triples. For example, given a concept term h, its triple relationship in ConceptNet is expressed as <h, relation, d>. Take all the relationship triples with h as the head node or tail node. Then the set of all terms related to each query term constitutes the semantic expansion set D of the query.

[0060] Apply the collaborative filtering method of the query semantic expansion set obtained from the ConceptNet knowledge base and the term set obtained from the feedback documents to initially select candidate expansion terms, which helps to reduce the error of query expansion caused by the incorrect selection of feedback documents. In actual situations, there may also be some members in the term semantic description set that have low semantic relevance to the query. If the semantic relevance between the query term and the candidate term is very low and it does not appear in the feedback documents, it can be ignored.

[0061] In this way, a term semantic description set C = {t1, t2,..., t n} can be obtained;

[0062] Its calculation method is as follows:

[0063] C = D ∩ E;

[0064] In the above formula, D is the set of candidate terms of the filtered query Q, and E is the set of all terms in the feedback document.

[0065] The collaborative filtering method of the set of query semantic descriptions obtained from the ConceptNet knowledge base and the set of terms obtained from the feedback document initially selects candidate expansion terms, which helps to reduce the error of query expansion caused by the wrong selection of the feedback document.

[0066] In step 3, to measure the semantic relevance between the candidate expansion term and the query, the pseudo-relevance feedback model only considers information such as the frequency of terms and the inverse document frequency, lacking the measurement of the semantics between terms. Measure the semantic matching score S(t,Q) between the candidate term and the query term. The calculation formula is as follows:

[0067] S(t,Q) = sim(t,Q) × IDF(t);

[0068] In the above formula, S(t,Q) is the matching score, sim(t,Q) is the semantic similarity between t and Q, and IDF(t) is the importance of t in the feedback document.

[0069]

[0070] In the above formula, represent the word vectors with term t as the head node and term q as the head node respectively, and μ t is the mean of the semantic similarities between all query terms in query Q and term t, and σ represents the variance of the semantic similarities between all query terms in query Q and term t.

[0071] Use the TransE (an algorithm) model to obtain 300-dimensional word vectors.

[0072] In step 4, according to the combined relevance matching and semantic matching between terms, obtain the weights of all candidate expansion terms. The weight w t of the candidate expansion term is calculated as follows:

[0073]

[0074] λ ∈ {0, 0.1, 0.2,..., 1.0};

[0075] In the above formula, tf(t,d) is the number of term t in document d, N is the number of feedback documents, λ is the balance factor, and w t is the weight of the candidate expansion term.

[0076] The extended term weight calculation method takes into account both term frequency (or term distribution) and semantic relevance. It consists of two parts. The first part represents the importance of candidate terms, which is measured using term frequency-inverse document frequency or a language model. MIN-MAX normalization is performed on the values of the two parts of the relevance matching method and the semantic matching method between terms. The balance factor λ is used to adjust the contributions of the two parts, and the candidate extended terms are sorted according to the final values.

[0077] It directly measures the semantic relevance between the query and the terms, selects extended terms by combining semantic matching between terms and relevance matching, and performs a second-round retrieval. The proposed method can identify terms that may be semantically relevant but do not have a high frequency or co-occurrence frequency.

[0078] In step 5, a new query is constructed and a secondary retrieval is performed. According to the weights of all candidate extended terms calculated in step 4, they are sorted in descending order, and the top |Tf| extended terms with the largest weights are selected to form a vector Q1, which together with the original query vector Q forms a new query vector Q′:

[0079] Q′ = (1 - α) × Q + α × Q1;

[0080] α ∈ {0, 0.1, 0.2,..., 1.0};

[0081] |Tf| ∈ {10, 20, 30, 50};

[0082] In the above formula, α is the balance factor, Q1 is the vector composed of the top |Tf| extended terms with the largest weights, Q is the original query vector, and Q′ is the new query vector formed by Q1 and Q. The balance factor α is used to adjust the contributions of the two parts.

[0083] The new query vector Q′ is used for secondary retrieval, the document score Score(Q, d) is calculated, and the results are sorted according to the size of the document scores and presented to the user.

[0084] It selects extended terms by considering the contributions of both semantic similarity between terms and the importance of terms, and can overcome the problems of traditional pseudo-relevance feedback models that only consider information such as term frequency and inverse document frequency and lack the measurement of semantic relationships between terms. It can also overcome the influence of the first-round retrieval method on query expansion in traditional pseudo-relevance methods. It directly measures the semantic relevance between the query and the extended terms, and obtains their semantic relevance by measuring the similarity of the word vectors of the query and the extended terms. It is statistically relevant to the query to ensure the importance of terms in the document collection; it is semantically relevant to the query to ensure consistency with the query topic. It is beneficial to find better extended terms to supplement the semantics of the query and improve the accuracy of the final retrieval results.

Claims

1. An information retrieval query expansion method based on ConceptNet, characterized in that It includes the following steps: S1: Use BM25 for retrieval to obtain a set E of feedback documents and terms; S2: Obtain the semantic description set C of terms by taking the semantic extension set D of terms from ConceptNet; S3: Measure the semantic matching score between the candidate extended terms and the query: S(t,Q) = sim(t,Q) × IDF(t); sim(t, Q) is the semantic similarity between t and Q, and IDF(t) is the importance degree of t in the feedback documents. denote the word vectors with the term t and the term q as the head nodes respectively, μ t is the mean of the semantic similarities between all query terms in the query Q and the term t, and σ represents the variance of the semantic similarities between all query terms in the query Q and the term t; S4: Calculate the weights of the candidate extended terms; where, tf(t, d i ) is the number of occurrences of term t in document d i , N is the number of feedback documents, and λ is a balancing factor; S5: Construct a new query and perform a secondary retrieval.

2. The information retrieval query expansion method based on ConceptNet according to claim 1, wherein In step S1, the formula for calculating the score of a document is as follows: In the above formula, Score(Q, d) is the document score, TF is the normalized representation of the document length, avdl is the average length of the documents, k1 and k3 are constants, tf is the frequency of the term q, qtf is the frequency of the query term q in the query, b is a tuning factor, dl is the document length, N′ is the number of all documents in the index, df t is the number of documents containing the term t in all documents, and IDF(t) is the importance of t in the feedback documents.

3. The information retrieval query expansion method based on ConceptNet according to claim 2, characterized in that, In step S1, the set E is a candidate pseudo-relevant term set composed of all terms in the top 10 documents with the highest document scores.

4. The information retrieval query expansion method based on ConceptNet according to claim 1, wherein Take all relationship triples with the query term as the head node or the tail node, and form a set D with all the head and tail nodes of the relationship triples. The set C is the intersection of the set D and the set E.

5. The information retrieval query expansion method based on ConceptNet according to claim 1, characterized in that, The dimension of the word vector is 300.

6. The information retrieval query expansion method based on ConceptNet according to claim 1, wherein Normalize the values of the relevance matching method and the semantic matching method between terms. The balance factor is used to adjust the contributions of the two parts, and the candidate extended terms are sorted.

7. The information retrieval query expansion method based on ConceptNet according to claim 1, characterized in that In step S5, the formula for calculating the new query is as follows: Q′ = (1 - α) × Q + α × Q1; α∈{0,0.1,0.2,...,1.0}； |Tf| ∈ {10, 20, 30, 50}; In the above formula, α is the balance factor, Q1 is a vector composed of the top |Tf| extended terms with the largest weights, Q is the original query vector, and Q′ is a new query vector formed by Q1 and Q.

8. The method for expanding an information retrieval query based on ConceptNet according to claim 1 or 7, wherein Use the new query vector to perform a secondary retrieval, calculate the document scores, sort them according to the magnitudes of the document scores, and display the result of the document sorting.

Images