CN113987155B - Conversational retrieval method integrating knowledge graph and large-scale user log - Google Patents
Conversational retrieval method integrating knowledge graph and large-scale user log Download PDFInfo
- Publication number
- CN113987155B CN113987155B CN202111412012.1A CN202111412012A CN113987155B CN 113987155 B CN113987155 B CN 113987155B CN 202111412012 A CN202111412012 A CN 202111412012A CN 113987155 B CN113987155 B CN 113987155B
- Authority
- CN
- China
- Prior art keywords
- entity
- representation
- query
- document
- representations
- 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.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 39
- 230000004927 fusion Effects 0.000 claims abstract description 24
- 238000004364 calculation method Methods 0.000 claims abstract description 16
- 238000013528 artificial neural network Methods 0.000 claims abstract description 12
- 230000000694 effects Effects 0.000 claims abstract description 12
- 230000008569 process Effects 0.000 claims abstract description 10
- 230000003993 interaction Effects 0.000 claims description 22
- 238000010586 diagram Methods 0.000 claims description 9
- 238000013461 design Methods 0.000 claims description 8
- 230000002452 interceptive effect Effects 0.000 claims description 8
- 238000012549 training Methods 0.000 claims description 7
- 238000011176 pooling Methods 0.000 claims description 6
- 239000011159 matrix material Substances 0.000 claims description 5
- 238000009826 distribution Methods 0.000 claims description 3
- 230000006870 function Effects 0.000 claims description 3
- 230000007246 mechanism Effects 0.000 claims description 3
- 238000012216 screening Methods 0.000 claims description 3
- 238000013507 mapping Methods 0.000 claims 1
- 238000013473 artificial intelligence Methods 0.000 abstract description 2
- 230000006399 behavior Effects 0.000 description 2
- 238000001914 filtration Methods 0.000 description 1
- 238000007634 remodeling Methods 0.000 description 1
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
- G06F16/332—Query formulation
- G06F16/3329—Natural language query formulation or dialogue systems
-
- 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
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Abstract
The invention realizes a conversational retrieval method integrating a knowledge graph and a large-scale user log by a method in the artificial intelligence field. First, by learning the representation of the entity in the current session on a fused entity graph, the context information corresponding to the entity is taken into account in the learning of the representation. The method then uses the learned entity representation and the original entity representation and word representation to effect a correlation calculation of the session context with the candidate document. And finally, sorting the candidate documents through the relevance scores. The method provided by the invention fuses the knowledge graph with the entity graph extracted from the query log, and calculates the fused entity representation through the graph attention neural network to fuse additional knowledge. The computed fusion entity representation is used for the matching process of the session context with the candidate document. Compared with the prior conversational search model, the method and the device try to introduce knowledge beyond the current conversation for the first time to improve the document search effect, and further improve the use experience of search engine users.
Description
Technical Field
The invention relates to the technical field of artificial intelligence, in particular to a conversational retrieval method integrating a knowledge graph and a large-scale user log.
Background
There are many session-based retrieval models. Some models learn about user intent by modeling remodelling relationships between multiple queries within a session. For example, the user has deleted certain terms compared to previous queries, indicating that these terms do not match the user's current query intent. Or the user has newly added words that are likely to more satisfy the user's current query intent. With the widespread use of depth models, many models began to attempt to obtain a representation of user intent through semantic modeling, and use this representation to measure the relevance of candidate documents to user intent. The latest session-based retrieval model orders two tasks through comprehensive query remodeling and documents, so that the two tasks are promoted to each other, and the effect of document retrieval is improved. In addition, the model [7] with the best effect at present improves the retrieval effect by modeling different behaviors of the user in the query process. These user actions include clicking and skipping certain documents.
The related effort of previous conversational searches uses only the information inside the conversation, i.e., the user's historical queries and historical click documents. However, such information is often limited, and it is sometimes insufficient to understand the intent of the user. For example, when a user mentions some ambiguous entities, the search engine is likely to misunderstand the user's query intent without the assistance of a knowledge-graph. In addition, co-occurrence relationships among many entities are implied in the large-scale user log, and are built by other users through their query behavior during the search process, which also plays an important role in understanding the user's intent of the current session, which was not considered in the past.
Disclosure of Invention
Therefore, the invention firstly provides a conversational retrieval method integrating a knowledge graph and a large-scale user log, firstly, the representation of the entity in the current conversation is learned on a fused entity graph, the context information corresponding to the entity is considered in the process of learning the representation of the entity in the current conversation,
specifically: for a given session context S, it contains a historical query and corresponding historical click documentWherein q is i Representing the i-th query in the session, +.>User click document set corresponding to the ith query>Wherein->Indicating the 1 st clicked document corresponding to the i-th query, n indicating the number of clicked documents, and the current query being q t The candidate document is d t First, word representations in each document and query are obtained, original entity representations and fusion entity representations, for all words [ w ] in a certain query 1 ,…,w |q| ]Obtaining the embedding of each word through Fasttext: /> Using these embeddings as input, contextual semantic information of words is fused into the representation of words by bi-directional LSTM, specifically let +.>And +.>Hidden representations of the i-th word obtained for forward and backward LSTM respectively, and concatenating them together results in a representation of each word:
the learned entity representation is then used with the original entity representation and the word representation to effect a correlation calculation of the session context with the candidate document,
specifically: for the original entity representation, acquiring an initial embedded representation of each entity through a TransE, and embedding the entity into the entity through a fully connected layer:
acquiring new entity representation by means of a fused entity representation moduleThe fusion entity representation module obtains the representation of the entity, and calculates a ranking feature for the s-th historical query and the click document corresponding to the s-th historical query through a single-round interaction module>The ranking features of each historical query are then integrated together into an overall contextual ranking feature>Wherein->Representing the importance of the s-th query, and calculating by designing an attention mechanism capable of perceiving the semantics of the query:
wherein the method comprises the steps ofAnd->Respectively represent the current query q t And history query q s And this representation is spliced by three representations: />Wherein->And->The word representation, the original entity representation and the fused entity representation are summed up, respectively:
the resulting overall context ranking feature f context Obtaining a context matching score S through a full connectivity layer c ＝MLP(f context ) In addition, for the ranking features corresponding to the current queryLikewise, a current query matching score is obtained via a full connectivity layer>Taking the average tf-idf values of words and entities into consideration, the number of public words in candidate documents and all queries, embedding the calculated similarity values between the candidate documents and the words and entities of historical queries as statistical features, and obtaining another matching score S through a full connection layer by using the statistical features f The three scores are spliced together to obtain the final matching score R (d) of the candidate document through a full connection layer t |q t ,S)；
Finally, ranking the candidate documents by taking the matching score output by the model as a basis;
the model training process adopts a model way, and the design of the loss function is as follows:
wherein R (d|q) is shorthand for R (d|q, S). S is S T Including all of the sessions in the training set,containing all clicked and un-clicked documents corresponding to query q, where d + Representing clicked document d - Representing an untclicked document. All model parameters are updated by a back propagation algorithm.
The realization mode of the fusion entity representation module is as follows: firstly, constructing a fusion entity diagram comprising a knowledge graph, entities in a query log and relations among the entities, and specifically, nodes in the fusion entity diagram are formed by two parts: querying entities in the log and entities in the knowledge graph;
the entities in the query log refer to the entities contained in the query and click documents in each session in the query log, and on the basis of the entities in the query log, triples are screened in the knowledge graph, wherein the screening condition is that the head entity or the tail entity of the triples appears in the entity set of the query log, and the edges between the entities are divided into three categories: two entities appear in the same triplet in the knowledge graph, two entities appear in two queries before and after the query log, and two entities respectively appear in one query in the query day and the click document corresponding to the two entities;
the representation of the node of a given entity, in particular of a given entity e, is then calculated by means of a graph attention neural network i The representation of the corresponding node is calculated by a multi-layer graph attention neural network and the output of the last layer of network is used as the final entity representation, each layer of representation being calculated in the following manner:
wherein the method comprises the steps ofRepresenting a layer k neural network calculationOutput of N (e) i ) Representing entity e i Neighbor node set,/->Representing entity e i The weight of the j-th neighbor of (2), which weight needs to be based on this neighbor and e i The context association degree is calculated by the following calculation modes:
wherein R is e Is entity e i A representation of the query or document in which it resides, which is obtained by summing the original representations of all entities within the query or document.
The single-round interactive module designs three kinds of representations: word representation, original entity representation, new fusion entity representation, and nine interaction matrices obtained by calculation:
[M ww ,M we ,M wi ,M ew ,M ee ,M ei ,M iw ,M ie ,M ii ]
the elements of these matrices correspond to interaction scores of two representations, where w represents a word, e represents an entity, and i is a new fused entity representation. The nine interaction matrices are calculated as follows:
wherein W is M Is used to map the fused entity representation with the word or original entity representation into the same semantic space, and then, at each matchObtaining k-dimensional matching features on the matching matrix by using a kernel pooling method in a KNRM model
Wherein mu k Sum sigma k Is a parameter of kernel pooling for measuring different similarity distribution, and then obtaining an interaction feature by weighting and summing the matching features obtained by 9 matrix calculation
The single-round interaction module is designed with a matching method based on representation, namely, for single documents in a click document set corresponding to each query, word representations are used, three document representations are calculated by an original entity representation and a fusion entity representation, the calculation mode is that the representations of each word or entity are added together, and then the representations of each document in the click document set are averaged to obtain the representations of the three click document sets:three representations corresponding to the candidate document are obtained by adding the representations of each word or entity in the candidate document: />For each representation, matching features of the set of click documents and the candidate documents are calculated by:
can be obtained in the same wayAnd->The three matching features and the interaction feature obtained before +.>Spliced together, a single round interactive ordering feature is obtained>
The invention has the technical effects that:
additional knowledge is fused by fusing the knowledge graph with the entity graph extracted from the query log and computing a fused entity representation through the graph attention neural network. The computed fusion entity representation is used for the matching process of the session context with the candidate document. Compared with the prior conversational search model, the method and the device try to introduce knowledge beyond the current conversation for the first time to improve the document search effect, and further improve the use experience of search engine users.
Drawings
FIG. 1 is a diagram of the overall structure of the model;
FIG. 2 is a schematic diagram of a single-wheel interactive module structure
Detailed Description
The following is a preferred embodiment of the present invention and a technical solution of the present invention is further described with reference to the accompanying drawings, but the present invention is not limited to this embodiment.
The method for improving the conversational search effect by introducing the external knowledge is tried for the first time, and the applied external knowledge comprises a knowledge graph based on the real relationship and a large-scale user log recorded by a search engine. . Firstly, constructing an entity diagram by utilizing a knowledge graph and a user log, and learning the representation of the entity in the current session on the entity diagram, wherein the context information corresponding to the entity is considered in the process of learning the representation. The method then uses the learned entity representation and the original entity representation and word representation to effect a correlation calculation of the session context with the candidate document. And finally, sorting the candidate documents through the relevance scores.
In the conversational search problem, given the context S of a conversation, a historical query and a corresponding historical click document are includedWherein q is i Representing the i-th query in the session, +.>User click document set corresponding to the ith query>Wherein->The 1 st click document corresponding to the i-th query is represented, and n represents the number of the clicked documents. In addition, given the current query q t And candidate document d t We calculate the relevance score R (d t |q t ,S)。
The overall structure of the model is shown in fig. 1. First we obtain word representations, original entity representations, and fused entity representations in each document and query. For all words in a query [ w ] 1 ,…,w |q| ]First, the embedding of each word is obtained through Fasttext:using these embeddings as input, the contextual semantic information of the word is fused into the representation of the word through the bi-directional LSTM. Specifically, let->And +.>Hidden representations of the i-th word obtained by forward and backward LSTM, respectively. Stitching them together results in a representation of each word: />
For the original entity representation, an initial embedded representation of each entity is first obtained by a TransE. Since the entities in each query or document do not form a continuous sequence, the use of LS TM to obtain semantic representations of the entities is not considered. However, in order to obtain an entity representation in the same semantic space as the word representation, it is necessary to embed the entity additionally through a fully connected layer:
in addition, in order to introduce additional information in the knowledge graph and the query log, the invention obtains a new entity representation through a fusion entity representation module (IER)The IER module obtains representations of entities by using the graph ideographic neural network on a map of fused entities, as will be described in detail in the following section.
After the word representation, the original entity representation and the fused entity representation are obtained, the candidate documents may be computed as relevance scores for each of the historical queries and their corresponding click documents, respectively. In particular, for the s-th historical query and its corresponding click document, a ranking feature may be computed by a Single-turn interaction module (Single-turn interaction)The ranking features of each historical query are then integrated together into an overall contextual ranking featureWherein->Importance of representing the s-th queryThe degree is calculated by a query-level attention mechanism:
wherein the method comprises the steps ofAnd->Respectively represent the current query q t And history query q s And this representation is spliced by three representations: />Wherein->And->The word representation, the original entity representation and the fused entity representation are summed up, respectively:
the resulting overall context ranking feature f context Obtaining a context matching score S through a full connectivity layer c ＝MLP(f context ). In addition, for the ranking features corresponding to the current queryLikewise, a current query matching score is obtained via a full connectivity layer>In addition, consider some statistical features, including the average tf-idf values of words and entities, candidate documentsThe number of public words in the files and all queries, and the calculated similarity value is embedded in the words and entities of the candidate documents and the historical queries. These features are passed through a full connection layer to obtain another matching score S f The three scores are spliced together to obtain the final matching score R (d) of the candidate document through a full connection layer t |q t ,S)。
Fusion entity representation module (IER):
in order to introduce additional information contained in the knowledge graph and the query log into the model, the invention designs a fusion entity representation module. The module first needs to build a fusion entity graph. The fused entity graph includes relationships between knowledge maps and entities in the query log. Specifically, the nodes in the fusion entity graph are composed of two parts: the entities in the query log and the entities in the knowledge graph. First, the entities in the query log refer to the queries within each session in the query log and the entities contained within the click document. And screening the triples in the knowledge graph on the basis of the entities in the query log. The condition for filtering is that the head entity or the tail entity of the triplet appears in the entity set of the query log. After all entities in the graph have been determined, edges between the entities also need to be determined. Edges between entities fall into three categories: 1) Two entities appear in the same triplet in the knowledge graph; 2) Two entities appear in both the front and back queries of the query log; 3) The two entities appear in one query and its corresponding click document on the query day, respectively. These three classes of edges describe objective associations between entities and two different co-occurrence associations, respectively. It should be noted that there are no direct edges between entities that appear in the same query or document, but they all have common neighbors, thus equating to indirectly existing associations. This design is to avoid redundancy of edges in the figure.
After the fusion graph is constructed, the present invention calculates a representation of a given entity node through the graph attention neural network. Specifically, given an entity e i Computing a representation of a corresponding node through a multi-layer graph attention neural network and using the output of the last layer network as the finalIs an entity representation of (c). Each layer represents the way in which the computation is as follows:
wherein the method comprises the steps ofRepresenting the output of the calculation of the layer k neural network, N (e i ) Representing entity e i Is described herein). />Representing entity e i The weight of the j-th neighbor of (2), which weight needs to be based on this neighbor and e i The context association degree is calculated by the following calculation modes:
wherein R is e Is entity e i A representation of the query or document in which it resides, which is obtained by summing the original representations of all entities within the query or document.
Single-wheel interaction module (Single-turn interaction)
The block diagram of the single-wheel interactive module is shown in fig. 2. The structure is inspired by the EDRM model, which uses four interaction matrices: query word-document word, query entity-document entity, query entity-document word, and query word-document entity. Three representations are designed in the present invention: word representation, original entity representation, new fusion entity representation, thus nine interaction matrices can be calculated:
[M ww ,M we ,M wi ,M ew ,M ee ,M ei ,M iw ,M ie ,M ii ]
the elements of these matrices correspond to interaction scores of the two representations, M we And M wi The following are examples:
wherein W is M Is used to map the fused entity representation into the same semantic space as the word or original entity representation. Subsequently, a kernel pooling method as mentioned in the KNRM model is used on each matching matrix to obtain k-dimensional matching features
Wherein mu k Sum sigma k Is a parameter of kernel pooling for measuring different similarity distributions. The matching features obtained by the calculation of the 9 matrixes are weighted and summed to obtain an interaction feature
In addition, the module designs a representation-based matching method in order to consider the role of clicking on documents corresponding to each historical query. For a single document in the set of click documents corresponding to each query, three document representations are computed using the word representations, the original entity representation and the fused entity representation, in such a way that the representations of each word or entity are added together. The representation of each document within the set of click documents is then averaged to obtain three representations of the set of click documents:similarly, three representations corresponding to the candidate document are obtained by summing the representations of each word or entity in the candidate document: />For each representation, matching features of the set of click documents and the candidate documents are calculated by:
can be obtained in the same wayAnd->The three matching features and the interaction feature obtained before +.>Spliced together, a single round interactive ordering feature is obtained>
Model training
The model training process adopts a model way, and the design of the loss function is as follows:
wherein R (d|q) is shorthand for R (d|q, S). S is S T Including all of the sessions in the training set,containing all clicked and un-clicked documents corresponding to query q, where d + Representing clicked document d - Representing an untclicked document. All model parameters are updated by a back propagation algorithm.
Claims (3)
1. A conversational search method integrating knowledge graph and large-scale user log, which improves conversational search effect by introducing external knowledge, wherein the external knowledge comprises knowledge graph based on reality relationship and large-scale user log recorded by search engine, and is characterized in that: firstly, constructing an entity graph by utilizing a knowledge graph and a user log, and learning the representation of the entity in the current session on the entity graph, so that the context information corresponding to the entity can be considered in the process of learning the representation of the entity,
specifically: for a given session context S, it contains a historical query and corresponding historical click documentWherein q is i Representing the i-th query in the session, +.>User click document set corresponding to the ith query>Wherein->Indicating the 1 st clicked document corresponding to the i-th query, n indicating the number of clicked documents, and the current query being q t The candidate document is d t First, word representations in each document and query are obtained, original entity representations and fusion entity representations, for all words [ w ] in a certain query 1 ,…,w |q| ]Obtaining the embedding of each word through Fasttext: /> Using these embeddings as input, contextual semantic information of words is fused into the representation of words by bi-directional LSTM, specifically let +.>And +.>Hidden representations of the i-th word obtained for forward and backward LSTM respectively, and concatenating them together results in a representation of each word:
the learned entity representation is then used with the original entity representation and the word representation to effect a correlation calculation of the session context with the candidate document,
specifically: for the original entity representation, acquiring an initial embedded representation of each entity through a TransE, and embedding the entity into the entity through a fully connected layer:
acquiring new entity representation by means of a fused entity representation moduleThe fusion entity representation module obtains the representation of the entity, and calculates a ranking feature for the s-th historical query and the click document corresponding to the s-th historical query through a single-round interaction module>The ranking features of each historical query are then integrated together into an overall contextual ranking feature>Wherein->Representing the importance of the s-th query, and calculating by designing an attention mechanism capable of perceiving the semantics of the query:
wherein the method comprises the steps ofAnd->Respectively represent the current query q t And history query q s And this representation is spliced by three representations: />Wherein->And->The word representation, the original entity representation and the fused entity representation are summed up, respectively:
the resulting overall context ranking feature f context Obtaining a context matching score S through a full connectivity layer c ＝MLP(f context ) In addition, for the ranking features corresponding to the current queryLikewise, a current query matching score is obtained via a full connectivity layer>Considering the average tf-idf values of words and entities, candidate documents and the number of common words in all queries, and candidate documents and historical queriesEmbedding the counted similarity value between the words and the entities as statistical features, and obtaining another matching score S by using the statistical features through a full connection layer f The three scores are spliced together to obtain the final matching score R (d) of the candidate document through a full connection layer t |q t ,S)；
Finally, ranking the candidate documents by taking the matching score output by the model as a basis;
the model training process adopts a model way, and the design of the loss function is as follows:
wherein R (d|q) is shorthand for R (d|q, S) T Including all of the sessions in the training set,containing all clicked and un-clicked documents corresponding to query q, where d + Representing clicked document d - Representing an uncapped document, all model parameters are updated by a back propagation algorithm.
2. The conversational search method of claim 1, wherein the knowledge graph and the large-scale user logs are integrated, wherein the conversational search method is characterized by: the realization mode of the fusion entity representation module is as follows: firstly, constructing a fusion entity diagram comprising a knowledge graph, entities in a query log and relations among the entities, and specifically, nodes in the fusion entity diagram are formed by two parts: querying entities in the log and entities in the knowledge graph;
the entities in the query log refer to the entities contained in the query and click documents in each session in the query log, and on the basis of the entities in the query log, triples are screened in the knowledge graph, wherein the screening condition is that the head entity or the tail entity of the triples appears in the entity set of the query log, and the edges between the entities are divided into three categories: two entities appear in the same triplet in the knowledge graph, two entities appear in two queries before and after the query log, and two entities respectively appear in one query in the query day and the click document corresponding to the two entities;
the representation of the node of a given entity, in particular of a given entity e, is then calculated by means of a graph attention neural network i The representation of the corresponding node is calculated by a multi-layer graph attention neural network and the output of the last layer of network is used as the final entity representation, each layer of representation being calculated in the following manner:
wherein the method comprises the steps ofRepresenting the output of the calculation of the layer k neural network, N (e i ) Representing entity e i Neighbor node set,/->Representing entity e i The weight of the j-th neighbor of (2), which weight needs to be based on this neighbor and e i The context association degree is calculated by the following calculation modes:
wherein R is e Is entity e i A representation of the query or document in which it resides, which is obtained by summing the original representations of all entities within the query or document.
3. The conversational search method of claim 2, wherein the knowledge graph and the large-scale user logs are integrated, wherein the conversational search method is characterized by: the single-round interactive module designs three kinds of representations: word representation, original entity representation, new fusion entity representation, and nine interaction matrices obtained by calculation:
[M ww ,M we ,M wi ,M ew ,M ee ,M ei ,M iw ,M ie ,M ii ]
the elements of these matrices correspond to interaction scores of two representations, where w represents a word, e represents an entity, i is a new fused entity representation, and the nine interaction matrices are calculated as follows:
wherein W is M Is used for mapping the fusion entity representation and the word or the original entity representation into the same semantic space, and then k-dimensional matching features are obtained on each matching matrix by using a kernel pooling method in a KNRM model
Wherein mu k Sum sigma k Is a parameter of kernel pooling for measuring different similarity distribution, and then obtaining an interaction feature by weighting and summing the matching features obtained by 9 matrix calculation
The single-wheel interactive module is designed with a baseIn the matching method of the representations, namely, for single documents in the click document set corresponding to each query, three document representations are calculated by using word representations, original entity representations and fusion entity representations, wherein the calculation mode is to add the representations of each word or entity together, and then the representations of each document in the click document set are averaged to obtain the representations of the three click document sets:three representations corresponding to the candidate document are obtained by adding the representations of each word or entity in the candidate document: />For each representation, matching features of the set of click documents and the candidate documents are calculated by:
can be obtained in the same wayAnd->The three matching features and the interaction feature obtained before +.>Spliced together, a single round interactive ordering feature is obtained>
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111412012.1A CN113987155B (en) | 2021-11-25 | 2021-11-25 | Conversational retrieval method integrating knowledge graph and large-scale user log |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111412012.1A CN113987155B (en) | 2021-11-25 | 2021-11-25 | Conversational retrieval method integrating knowledge graph and large-scale user log |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113987155A CN113987155A (en) | 2022-01-28 |
| CN113987155B true CN113987155B (en) | 2024-03-26 |
Family
ID=79731813
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111412012.1A Active CN113987155B (en) | 2021-11-25 | 2021-11-25 | Conversational retrieval method integrating knowledge graph and large-scale user log |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113987155B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114741627B (en) * | 2022-04-12 | 2023-03-24 | 中国人民解放军32802部队 | Internet-oriented auxiliary information searching method |
| CN116304111B (en) * | 2023-04-10 | 2024-02-20 | 深圳市兴海物联科技有限公司 | AI call optimization processing method and server based on visual service data |
| CN116775848B (en) * | 2023-08-23 | 2023-11-07 | 宁波吉利汽车研究开发有限公司 | Control method, device, computing equipment and storage medium for generating dialogue information |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2017041372A1 (en) * | 2015-09-07 | 2017-03-16 | 百度在线网络技术（北京）有限公司 | Man-machine interaction method and system based on artificial intelligence |
| CN111177357A (en) * | 2019-12-31 | 2020-05-19 | 中国人民大学 | Memory neural network-based conversational information retrieval method |
| CN112069399A (en) * | 2020-08-25 | 2020-12-11 | 中国人民大学 | Personalized search system based on interactive matching |
| CN112732944A (en) * | 2021-01-30 | 2021-04-30 | 吉林大学 | New method for text retrieval |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US10803055B2 (en) * | 2017-12-15 | 2020-10-13 | Accenture Global Solutions Limited | Cognitive searches based on deep-learning neural networks |
-
2021
- 2021-11-25 CN CN202111412012.1A patent/CN113987155B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2017041372A1 (en) * | 2015-09-07 | 2017-03-16 | 百度在线网络技术（北京）有限公司 | Man-machine interaction method and system based on artificial intelligence |
| CN111177357A (en) * | 2019-12-31 | 2020-05-19 | 中国人民大学 | Memory neural network-based conversational information retrieval method |
| CN112069399A (en) * | 2020-08-25 | 2020-12-11 | 中国人民大学 | Personalized search system based on interactive matching |
| CN112732944A (en) * | 2021-01-30 | 2021-04-30 | 吉林大学 | New method for text retrieval |
Non-Patent Citations (2)
| Title |
|---|
| 一种面向多源知识图谱的样例查询方法;汤楠;申德荣;寇月;聂铁铮;;计算机研究与发展;20151015(S1);全文 * |
| 马飞翔 ; 廖祥文 ; 於志勇 ; 吴运兵 ; 陈国龙 ; .基于知识图谱的文本观点检索方法.山东大学学报(理学版).(11),全文. * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113987155A (en) | 2022-01-28 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113987155B (en) | Conversational retrieval method integrating knowledge graph and large-scale user log | |
| CN112784092B (en) | Cross-modal image text retrieval method of hybrid fusion model | |
| CN110826336B (en) | Emotion classification method, system, storage medium and equipment | |
| CN111597314B (en) | Reasoning question-answering method, device and equipment | |
| CN110674850A (en) | Image description generation method based on attention mechanism | |
| CN110347932B (en) | Cross-network user alignment method based on deep learning | |
| CN106021364A (en) | Method and device for establishing picture search correlation prediction model, and picture search method and device | |
| CN112084331A (en) | Text processing method, text processing device, model training method, model training device, computer equipment and storage medium | |
| US11768869B2 (en) | Knowledge-derived search suggestion | |
| CN112966091B (en) | Knowledge map recommendation system fusing entity information and heat | |
| CN114358657B (en) | Post recommendation method and device based on model fusion | |
| CN112988917A (en) | Entity alignment method based on multiple entity contexts | |
| CN116601626A (en) | Personal knowledge graph construction method and device and related equipment | |
| CN114780748A (en) | Priori weight enhancement-based completion method of knowledge graph | |
| CN114528898A (en) | Scene graph modification based on natural language commands | |
| CN113779996A (en) | Standard entity text determination method and device based on BilSTM model and storage medium | |
| CN115203550A (en) | Social recommendation method and system for enhancing neighbor relation | |
| CN113380360B (en) | Similar medical record retrieval method and system based on multi-mode medical record map | |
| CN114840747A (en) | News recommendation method based on comparative learning | |
| CN114706989A (en) | Intelligent recommendation method based on technical innovation assets as knowledge base | |
| CN114332519A (en) | Image description generation method based on external triple and abstract relation | |
| CN114996490A (en) | Movie recommendation method, system, storage medium and device | |
| CN115344794A (en) | Scenic spot recommendation method based on knowledge map semantic embedding | |
| CN113821610A (en) | Information matching method, device, equipment and storage medium | |
| CN113076490A (en) | Case-related microblog object-level emotion classification method based on mixed node graph |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |