CN110309255A - A kind of entity search method for incorporating entity description distribution and indicating - Google Patents
A kind of entity search method for incorporating entity description distribution and indicating Download PDFInfo
- Publication number
- CN110309255A CN110309255A CN201810187938.7A CN201810187938A CN110309255A CN 110309255 A CN110309255 A CN 110309255A CN 201810187938 A CN201810187938 A CN 201810187938A CN 110309255 A CN110309255 A CN 110309255A
- Authority
- CN
- China
- Prior art keywords
- entity
- text
- term vector
- entity search
- vector matrix
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/33—Querying
- G06F16/3331—Query processing
- G06F16/334—Query execution
- G06F16/3347—Query execution using vector based model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/36—Creation of semantic tools, e.g. ontology or thesauri
- G06F16/367—Ontology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/279—Recognition of textual entities
- G06F40/284—Lexical analysis, e.g. tokenisation or collocates
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Computational Linguistics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Databases & Information Systems (AREA)
- Audiology, Speech & Language Pathology (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Animal Behavior & Ethology (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The present invention relates to a kind of entity search methods that involvement entity description distribution indicates, the method includes the following steps: effective term vector matrix training step is trained according to existing training sample, obtains effective term vector matrix；Entity search sequence step, entity search related text is embedded into effective term vector matrix, obtain the distributed nature expression of entity search related text, the correlation calculations and sequence that entity search result is carried out according to the result of distributed nature expression, obtain the output listing of entity search result.Compared with prior art, the present invention has many advantages, such as the accuracy rate for reducing manual intervention, reducing Feature Engineering work and effectively promoting entity search.
Description
Technical field
The present invention relates to Computer Science and Technology fields, indicate more particularly, to a kind of involvement entity description distribution
Entity search method.
Background technique
In entity search engine, search system how to be enabled effectively to understand the intention of user subject search inquiry,
And then it is highly important for returning to accurate list of entities.The entity of return is not only merely between query text with literal
On matching, should also have certain semantic dependency.One entity search query text is usually one of user's input
Short text, and the query result that system returns consists of two parts, a part is entity text, and another part is the entity
Entity description.Entity search method attempts to be ranked up all candidate answers, to will meet entity search query demand
Candidate answers come list forefront as far as possible.
The rule-based entity search method of tradition needs a large amount of Feature Engineering to obtain the semantic letter of word or sentence
Breath.Due to the flexibility of short text, the rule of Manual definition cannot cover all features.This also causes conventional method to need
More manual interventions are to obtain better query result.
Another problem of existing method is, in entity search system, the input of user may includes a variety of languages
Speech, since the syntax rule of different language is different, we can not parse the knot of short text using same syntax analyzer
Structure and semantic information, this problem also result in conventional method and need a large amount of manual intervention.
Summary of the invention
The purpose of the present invention is provide a kind of entity search side that involvement entity description distribution indicates regarding to the issue above
Method.
The purpose of the present invention can be achieved through the following technical solutions:
A kind of entity search method for incorporating entity description distribution and indicating, the method includes the following steps:
Effective term vector matrix training step, is trained according to existing training sample, obtains effective term vector square
Battle array；
Entity search related text is embedded into effective term vector matrix, obtains entity and search by entity search sequence step
The distributed nature of rope related text is expressed, and the correlation of entity search result is carried out according to the result of distributed nature expression
It calculates and sorts, obtain the output listing of entity search result.
Preferably, effective term vector matrix training step includes the following steps:
A1) all training samples are initialized, obtain all term vector matrixes；
A2 stochastical sampling) is carried out to all term vector matrixes that step A1) is obtained, and constructs and obtains multiple term vector squares
Battle array training sample；
A3 all term vector matrix training samples) are traversed, the loss of each term vector matrix training sample is calculated
Value, and all term vector matrixes are updated according to gradient descent algorithm method；
A4) judgment step A3) whether the sum of all penalty values for being calculated tend to restrain, if then entering step
A5), if otherwise return step A2)；
A5) current all term vector matrixes are exported as effective term vector matrix.
Preferably, the term vector matrix training sample includes entity search query text sample, is positively correlated entity text
Sample, is positively correlated entity description samples of text and negatively correlated entity description samples of text at negatively correlated entity samples of text.
It is preferably, described that all term vector matrixes are updated according to gradient descent algorithm method specifically:
Wherein, WnewFor updated all term vector matrixes, WoldFor all term vector matrixes before update, η is to fix
Learning rate, loss be term vector matrix training sample penalty values, ▽ loss be term vector matrix training sample penalty values
Gradient.
Preferably, the penalty values of the term vector matrix training sample specifically:
Loss=max { 0, m-cos (Rent ++Rdes +,Rq)+cos(Rent -+Rdes -,Rq)}
Wherein, loss is the penalty values of term vector matrix training sample, and m is the hyper parameter of model, Rent +It is real to be positively correlated
The feature vector of body samples of text, Rdes +The related entities that are positive describe the feature vector of samples of text, RqFor entity search inquiry
The feature vector of samples of text, Rent -The feature vector of the related entities that are negative samples of text, Rdes -The related entities that are negative describe text
The feature vector of sample.
Preferably, the entity search sequence step includes the following steps:
B1 entity search query text, entity text and entity description text) are embedded into effective term vector matrix respectively
In, obtain the distributed nature expression of entity search related text；
B2 feature selecting) is carried out to the distributed nature expression of the obtained entity search related text of step B1), respectively
Obtain the distributed nature vector of entity search query text, entity text and entity description text；
B3) according to the obtained distributed nature of entity search query text, entity text and entity description text to
Amount, carries out the correlation calculations of entity search result；
B4) entity search result is ranked up according to correlativity calculation result, obtains the output column of entity search result
Table.
Preferably, the step B1) include the following steps:
B11) entity search query text is segmented, is inquired in effective term vector matrix according to word segmentation result pair
The vector answered, arrangement obtain the term vector matrix W of entity search query textq；
B12 corresponding vector) is inquired in effective term vector matrix using entity text as a word, arrangement obtains reality
The term vector W of body textent；
B13) entity description text is segmented, is inquired in effective term vector matrix according to word segmentation result corresponding
Vector, arrangement obtain the term vector matrix W of entity description textdes。
Preferably, the feature selecting specifically: all distributed nature vectors are subjected to 1- maximum pond, respectively
Obtain the distributed nature vector of entity search query text, entity text and entity description text.
Preferably, the step B3) include:
B31) to the distributed nature vector of the corresponding entity description text of the distributed nature vector sum of entity text into
Row combination, obtains candidate answers feature vector；
B32) the distributed nature vector sum candidate answers feature vector of computational entity query text carries out cosine phase
Like degree matching primitives, matching score is obtained.
Preferably, the cosine similarity matching primitives specifically:
Wherein, RqFor the distributed nature vector of entity search query text, RcandFor candidate answers feature vector.
Compared with prior art, the invention has the following advantages:
(1) method proposed by the present invention obtains effective term vector matrix by the existing training text of training, in reality
In body search process, entity search query text, entity text and entity description text need to be only embedded into trained effective
In term vector matrix, the result after being then embedded in carries out correlation calculations, can judge entity search query text and entity
The degree of correlation size of search result (i.e. entity text and entity description text), thus obtain the output of the high degree of correlation as a result,
This searching method is that one kind searching method, manual intervention realized to improve precision end to end is only needed effective
It is realized in term vector matrix training step, in actual search process, search result can be realized without manual intervention
Sequence, had both reduced the work of Feature Engineering, and had also effectively improved the order of accuarcy of entity search, was suitble to popularity.
(2) it during the training of effective term vector matrix, changes by building term vector matrix training sample and constantly
In generation, can greatly promote the order of accuarcy of effective term vector matrix whether according to the convergence of penalty values, and with searching times
Increase, the quantity of training sample can also promote therewith, can make the order of accuarcy of effective term vector matrix higher in this way,
To greatly promote the order of accuarcy of entire search result.
(3) in entity search sequencer procedure, by by the feature vector of entity text and entity description text merge after again
Degree of correlation matching is carried out with entity search query text, judges the degree of correlation of search result by calculating cosine similarity,
The entity that this matching had both considered search result has also contemplated the associated description of search result, therefore order of accuarcy is high, more can
Accurately correspond to the answer that searchers wants.
Detailed description of the invention
Fig. 1 is the method flow diagram of effective term vector matrix training step；
Fig. 2 is the method flow diagram of entity search sequence step.
Specific embodiment
The present invention is described in detail with specific embodiment below in conjunction with the accompanying drawings.The present embodiment is with the technology of the present invention side
Implemented premised on case, the detailed implementation method and specific operation process are given, but protection scope of the present invention is unlimited
In following embodiments.
The present embodiment proposes a kind of entity search method that involvement entity description distribution indicates, mainly includes following
Two steps:
Effective term vector matrix training step, is trained according to existing training sample, obtains effective term vector square
Battle array；
Entity search related text is embedded into effective term vector matrix, obtains entity and search by entity search sequence step
The distributed nature of rope related text is expressed, and the correlation of entity search result is carried out according to the result of distributed nature expression
It calculates and sorts, obtain the output listing of entity search result.
Wherein, effective term vector matrix training step includes the following steps:
A1) all training samples are initialized, obtain all term vector matrixes；
A2 stochastical sampling) is carried out to all term vector matrixes that step A1) is obtained, and constructs and obtains multiple term vector squares
Battle array training sample；
A3 all term vector matrix training samples) are traversed, the loss of each term vector matrix training sample is calculated
Value, and all term vector matrixes are updated according to gradient descent algorithm method；
A4) judgment step A3) whether the sum of all penalty values for being calculated tend to restrain, if then entering step
A5), if otherwise return step A2)；
A5) current all term vector matrixes are exported as effective term vector matrix.
According to above-mentioned steps, the training process of effective term vector matrix is specifically as shown in Figure 2: first to all instructions
Practice sample to be initialized, the word for occurring entity search query text, entity text and entity description text
The term vector dimension of V random initializtion, each word v ∈ V is d, and then obtains all term vector matrixes, is being obtained entirely
After pronouns, general term for nouns, numerals and measure words vector matrix, by stochastical sampling, to construct term vector matrix training sample, each term vector matrix training sample
This specific format is (q, ent+, ent-, des+, des-): it contains entity search query text sample q, be positively correlated entity
Samples of text ent+, negatively correlated entity samples of text ent-, be positively correlated entity description samples of text des+It is retouched with negatively correlated entity
State samples of text des-；After construction has got well term vector matrix training sample, to all term vector matrix samples progress time
It goes through, and all term vector matrixes is updated by gradient descent algorithm method, while calculating all term vector matrix training
The sum of penalty values of sample continue to carry out structure to all term vector matrixes progress stochastical sampling if the sum of penalty values do not restrain
New term vector matrix training sample is made, to traverse again, if the sum of penalty values tend to restrain, stops traversal and iteration,
The corresponding all term vector matrixes of output, as effective term vector matrix.
In the above process, to the loss value calculating method of term vector matrix training sample specifically:
Loss=max { 0, m-cos (Rent ++Rdes +,Rq)+cos(Rent -+Rdes -,Rq)}
Wherein, loss is the penalty values of term vector matrix training sample, and m is the hyper parameter of model, and between (0,1)
Real number, for controlling the differentiation degree being positively correlated between entity and negatively correlated entity, i.e., it can guarantee to be positively correlated entity
Score is at least higher by m, R than negative sample entity after trainingent +The feature vector of the related entities that are positive samples of text, Rdes +
The related entities that are positive describe the feature vector of samples of text, RqFor the feature vector of entity search query text sample, Rent -For
The feature vector of negatively correlated entity samples of text, Rdes -The related entities that are negative describe the feature vector of samples of text.
It, then can be according to gradient descent algorithm to entirety after the penalty values that term vector matrix training sample has been calculated
Term vector matrix is updated, specifically:
Wherein, WnewFor updated all term vector matrixes, WoldFor all term vector matrixes before update, η is to fix
Learning rate, be usually arranged as 0.1, loss be term vector matrix training sample penalty values, ▽ loss be term vector matrix instruct
Practice the gradient of the penalty values of sample.
Once having obtained effective term vector matrix, then it can be saved to be used for entity search sequence step, passed through
A period of time has accumulated after enough new training samples again re -training to guarantee the order of accuarcy of effective term vector matrix i.e.
It can.
After having obtained available effective term vector matrix, entity search sort the step of specifically include:
B1 entity search query text, entity text and entity description text) are embedded into effective term vector matrix respectively
In, obtain the distributed nature expression of entity search related text:
B11) entity search query text is segmented, is inquired in effective term vector matrix according to word segmentation result pair
The vector answered, arrangement obtain the term vector matrix W of entity search query textq；
B12 corresponding vector) is inquired in effective term vector matrix using entity text as a word, arrangement obtains reality
The term vector W of body textent；
B13) entity description text is segmented, is inquired in effective term vector matrix according to word segmentation result corresponding
Vector, arrangement obtain the term vector matrix W of entity description textdes；
B2 feature selecting) is carried out to the distributed nature expression of the obtained entity search related text of step B1), respectively
Obtain the distributed nature vector of entity search query text, entity text and entity description text；
B3) according to the obtained distributed nature of entity search query text, entity text and entity description text to
Amount, carries out the correlation calculations of entity search result:
B31) to the distributed nature vector of the corresponding entity description text of the distributed nature vector sum of entity text into
Row combination, obtains candidate answers feature vector；
B32) the distributed nature vector sum candidate answers feature vector of computational entity query text carries out cosine phase
Like degree matching primitives, matching score is obtained；
B4) entity search result is ranked up according to correlativity calculation result, obtains the output column of entity search result
Table.
The process to be sorted according to the above-mentioned steps entity search is as follows, after having input entity search query text,
Multiple corresponding entity texts and entity description text can be obtained as entity search result, at this time respectively look into entity search
It is corresponding with effective term vector matrix that training obtains before to ask text, entity text and entity description text, and entity is searched
Rope query text and entity description text then first segment it since itself length is longer, obtain to participle
Each word inquires corresponding term vector in effective term vector matrix, and entity search can be then respectively obtained after integration and is looked into
The term vector matrix of text and the term vector matrix of entity description text are ask, and entity text is since this body length is little,
It is inquired in effective term vector matrix as a word, the corresponding term vector obtained after inquiry is i.e. as real
The term vector of body text.After obtaining above-mentioned term vector matrix and term vector, these vector sum matrixes are subjected to 1- most
Great Chiization can then respectively obtain the distributed nature vector of entity search query text, entity text and entity description text,
At this time by the distributed nature vector R of entity textentWith the distributed nature vector R of corresponding entity description textdesIt carries out
It combines, then the feature vector R of available candidate answerscand=Rent+Rdes.By the feature vector and entity of this candidate answers
The distributed nature vector of query text carries out similarity calculation, specific formula are as follows:
Wherein, RqFor the distributed nature vector of entity search query text, RcandFor candidate answers feature vector, obtain
Result according to sorting from large to small, the as relevancy ranking of search result outputs it and has obtained entity search result
Output listing, the result obtained in this way not only can utmostly avoid human intervention, but can make search result more subject to
Really, practical performance is greatly improved.
Claims (10)
1. a kind of entity search method for incorporating entity description distribution and indicating, which is characterized in that the method includes following steps
It is rapid:
Effective term vector matrix training step, is trained according to existing training sample, obtains effective term vector matrix；
Entity search related text is embedded into effective term vector matrix, obtains entity search phase by entity search sequence step
Close text distributed nature expression, according to distributed nature expression result carry out entity search result correlation calculations and
Sequence, obtains the output listing of entity search result.
2. the entity search method according to claim 1 for incorporating entity description distribution and indicating, which is characterized in that described
Effective term vector matrix training step includes the following steps:
A1) all training samples are initialized, obtain all term vector matrixes；
A2 stochastical sampling) is carried out to all term vector matrixes that step A1) is obtained, and constructs and obtains multiple term vector matrix training
Sample；
A3 all term vector matrix training samples) are traversed, the penalty values of each term vector matrix training sample, and root are calculated
All term vector matrixes are updated according to gradient descent algorithm method；
A4) judgment step A3) whether the sum of all penalty values for being calculated tend to restrain, if then entering step A5), if not
Then return step A2)；
A5) current all term vector matrixes are exported as effective term vector matrix.
3. the entity search method according to claim 2 for incorporating entity description distribution and indicating, which is characterized in that described
Term vector matrix training sample includes entity search query text sample, is positively correlated entity samples of text, negatively correlated entity text
Sample is positively correlated entity description samples of text and negatively correlated entity description samples of text.
4. the entity search method according to claim 2 for incorporating entity description distribution and indicating, which is characterized in that described
All term vector matrixes are updated according to gradient descent algorithm method specifically:
Wherein, WnewFor updated all term vector matrixes, WoldFor all term vector matrixes before update, η is fixed
Habit rate, loss are the penalty values of term vector matrix training sample,For the ladder of the penalty values of term vector matrix training sample
Degree.
5. the entity search method according to claim 2 for incorporating entity description distribution and indicating, which is characterized in that described
The penalty values of term vector matrix training sample specifically:
Loss=max { 0, m-cos (Rent ++Rdes +,Rq)+cos(Rent -+Rdes -,Rq)}
Wherein, loss is the penalty values of term vector matrix training sample, and m is the hyper parameter of model, Rent +The related entities that are positive text
The feature vector of sample, Rdes +The related entities that are positive describe the feature vector of samples of text, RqFor entity search query text sample
Feature vector, Rent -The feature vector of the related entities that are negative samples of text, Rdes -The related entities that are negative describe the spy of samples of text
Levy vector.
6. the entity search method for incorporating entity description distribution and indicating stated according to claim 1, which is characterized in that the reality
Body searching order step includes the following steps:
B1) entity search query text, entity text and entity description text are embedded into effective term vector matrix respectively, obtained
Distributed nature to entity search related text is expressed；
B2 feature selecting) is carried out to the distributed nature expression of the obtained entity search related text of step B1), respectively obtains reality
The distributed nature vector of body query text, entity text and entity description text；
B3 it) according to the distributed nature vector of obtained entity search query text, entity text and entity description text, carries out
The correlation calculations of entity search result；
B4) entity search result is ranked up according to correlativity calculation result, obtains the output listing of entity search result.
7. the entity search method according to claim 6 for incorporating entity description distribution and indicating, which is characterized in that described
Step B1) include the following steps:
B11) entity search query text is segmented, inquired in effective term vector matrix according to word segmentation result it is corresponding to
Amount, arrangement obtain the term vector matrix W of entity search query textq；
B12 corresponding vector) is inquired in effective term vector matrix using entity text as a word, arrangement obtains entity text
Term vector Went；
B13) entity description text is segmented, corresponding vector is inquired in effective term vector matrix according to word segmentation result,
Arrangement obtains the term vector matrix W of entity description textdes。
8. the entity search method according to claim 6 for incorporating entity description distribution and indicating, which is characterized in that described
Feature selecting specifically: all distributed nature vector is subjected to 1- maximum pond, respectively obtain entity search query text,
The distributed nature vector of entity text and entity description text.
9. the entity search method according to claim 6 for incorporating entity description distribution and indicating, which is characterized in that described
Step B3) include:
B31 group) is carried out to the distributed nature vector of the corresponding entity description text of the distributed nature vector sum of entity text
It closes, obtains candidate answers feature vector；
B32) the distributed nature vector sum candidate answers feature vector of computational entity query text carries out cosine similarity
Matching primitives obtain matching score.
10. the entity search method according to claim 9 for incorporating entity description distribution and indicating, which is characterized in that institute
State cosine similarity matching primitives specifically:
Wherein, RqFor the distributed nature vector of entity search query text, RcandFor candidate answers feature vector.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810187938.7A CN110309255A (en) | 2018-03-07 | 2018-03-07 | A kind of entity search method for incorporating entity description distribution and indicating |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810187938.7A CN110309255A (en) | 2018-03-07 | 2018-03-07 | A kind of entity search method for incorporating entity description distribution and indicating |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110309255A true CN110309255A (en) | 2019-10-08 |
Family
ID=68073350
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810187938.7A Pending CN110309255A (en) | 2018-03-07 | 2018-03-07 | A kind of entity search method for incorporating entity description distribution and indicating |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110309255A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112182154A (en) * | 2020-09-25 | 2021-01-05 | 中国人民大学 | Personalized search model for eliminating keyword ambiguity by utilizing personal word vector |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104598611A (en) * | 2015-01-29 | 2015-05-06 | 百度在线网络技术（北京）有限公司 | Method and system for sequencing search entries |
| CN104615767A (en) * | 2015-02-15 | 2015-05-13 | 百度在线网络技术（北京）有限公司 | Searching-ranking model training method and device and search processing method |
| CN105809473A (en) * | 2016-02-29 | 2016-07-27 | 北京百度网讯科技有限公司 | Training method, service recommending method for coupling model parameters and associated apparatus |
| CN105912625A (en) * | 2016-04-07 | 2016-08-31 | 北京大学 | Linked data oriented entity classification method and system |
| CN105930413A (en) * | 2016-04-18 | 2016-09-07 | 北京百度网讯科技有限公司 | Training method for similarity model parameters, search processing method and corresponding apparatuses |
| CN106547887A (en) * | 2016-10-27 | 2017-03-29 | 北京百度网讯科技有限公司 | Method and apparatus is recommended in search based on artificial intelligence |
| CN106815252A (en) * | 2015-12-01 | 2017-06-09 | 阿里巴巴集团控股有限公司 | A kind of searching method and equipment |
-
2018
- 2018-03-07 CN CN201810187938.7A patent/CN110309255A/en active Pending
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104598611A (en) * | 2015-01-29 | 2015-05-06 | 百度在线网络技术（北京）有限公司 | Method and system for sequencing search entries |
| CN104615767A (en) * | 2015-02-15 | 2015-05-13 | 百度在线网络技术（北京）有限公司 | Searching-ranking model training method and device and search processing method |
| CN106815252A (en) * | 2015-12-01 | 2017-06-09 | 阿里巴巴集团控股有限公司 | A kind of searching method and equipment |
| CN105809473A (en) * | 2016-02-29 | 2016-07-27 | 北京百度网讯科技有限公司 | Training method, service recommending method for coupling model parameters and associated apparatus |
| CN105912625A (en) * | 2016-04-07 | 2016-08-31 | 北京大学 | Linked data oriented entity classification method and system |
| CN105930413A (en) * | 2016-04-18 | 2016-09-07 | 北京百度网讯科技有限公司 | Training method for similarity model parameters, search processing method and corresponding apparatuses |
| CN106547887A (en) * | 2016-10-27 | 2017-03-29 | 北京百度网讯科技有限公司 | Method and apparatus is recommended in search based on artificial intelligence |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112182154A (en) * | 2020-09-25 | 2021-01-05 | 中国人民大学 | Personalized search model for eliminating keyword ambiguity by utilizing personal word vector |
| CN112182154B (en) * | 2020-09-25 | 2023-10-10 | 中国人民大学 | Personalized search model for eliminating keyword ambiguity by using personal word vector |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108804641B (en) | Text similarity calculation method, device, equipment and storage medium | |
| CN111444320B (en) | Text retrieval method and device, computer equipment and storage medium | |
| CN104915340B (en) | Natural language question-answering method and device | |
| CN108829801A (en) | A kind of event trigger word abstracting method based on documentation level attention mechanism | |
| JP5216063B2 (en) | Method and apparatus for determining categories of unregistered words | |
| CN110737758A (en) | Method and apparatus for generating a model | |
| US20090089047A1 (en) | Natural Language Hypernym Weighting For Word Sense Disambiguation | |
| CN110674252A (en) | High-precision semantic search system for judicial domain | |
| JP5710581B2 (en) | Question answering apparatus, method, and program | |
| CN110704621A (en) | Text processing method and device, storage medium and electronic equipment | |
| CN112163425A (en) | Text entity relation extraction method based on multi-feature information enhancement | |
| CN111695349A (en) | Text matching method and text matching system | |
| CN112861524A (en) | Deep learning-based multilevel Chinese fine-grained emotion analysis method | |
| CN112287656B (en) | Text comparison method, device, equipment and storage medium | |
| CN110929498A (en) | Short text similarity calculation method and device and readable storage medium | |
| CN109710921A (en) | Calculation method, device, computer equipment and the storage medium of Words similarity | |
| CN110659392B (en) | Retrieval method and device, and storage medium | |
| CN108595413B (en) | Answer extraction method based on semantic dependency tree | |
| Pan et al. | A content-based neural reordering model for statistical machine translation | |
| CN110309255A (en) | A kind of entity search method for incorporating entity description distribution and indicating | |
| Khassanov et al. | Enriching rare word representations in neural language models by embedding matrix augmentation | |
| KR102609616B1 (en) | Method and apparatus for image processing, electronic device and computer readable storage medium | |
| Yang et al. | Court similar case recommendation model based on word embedding and word frequency | |
| KR20200073524A (en) | Apparatus and method for extracting key-phrase from patent documents | |
| CN110413735B (en) | Question and answer retrieval method and system, computer equipment and readable storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |