A dialogue recommendation system based on a prompt learning method

Abstract

The application discloses a dialogue recommendation system based on a prompt learning method, and comprises the following steps: S1: through self-supervised learning based on entity prediction, semantic information of dialogue text and a knowledge graph is fused as a task-shared prompt, and knowledge information required by a dialogue recommendation system is supplemented for a pre-trained language model; S2: through prompt design specific to a dialogue task, a pre-trained language model is driven to generate a template sentence with an item slot as an intermediate result of a reply; and S3: through prompt specific to a recommendation task, a pre-trained language model is driven to generate an item of interest of a user. With the aid of a knowledge graph enhanced prompt learning technology, in the case of a fixed large-scale pre-trained language model, task sharing and task-specific prompts are added, so that one model can complete the dialogue and recommendation tasks with high quality, and generate a reply sentence and a recommended item that are consistent in results.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence natural language processing and recommendation systems, and in particular to a dialogue recommendation system based on a prompting learning method. Background Technology

[0002] In the era of big data, information is exploding. To alleviate information overload and extract commercial value, recommendation systems have emerged. They aim to recommend information or items of interest to users based on their preferences and needs. Recommendation systems have been widely adopted due to their value, as seen in e-commerce platforms like Taobao and JD.com. Traditional recommendation systems provide services through a one-way, one-time interaction. Since users cannot inform the system of their current preferences and needs, the system often estimates preferences by analyzing historical user behavior, such as click history, shopping records, and product reviews, before providing recommendations. This design has two problems: first, users' historical behavior is often limited, making it difficult to effectively estimate their preferences; second, users' needs change over time, and the system cannot perceive their latest needs through a one-way interaction. These problems lead to poor user experience and low relevance of recommended items, hindering the realization of their commercial value.

[0003] With the rise of intelligent assistants, more and more people are interacting with machines through dialogue. Machines use voice and semantic technology to perceive user needs and help them complete corresponding tasks. Dialogue, as a medium, can effectively break down the information asymmetry between users and systems, better help systems perceive user preferences, and make adjustments based on user feedback, thereby providing users with a better user experience.

[0004] Therefore, in order to improve the service quality of recommendation systems, this invention proposes a dialogue recommendation system based on a cue learning method. This system can drive a large-scale pre-trained model to infer users' preferences based on their historical dialogues with them through knowledge-enhanced cues, thereby generating high-quality recommended items and response statements.

[0005] Large-scale pre-trained models are widely used in natural language processing for various tasks involving language understanding and generation. In this invention, pre-trained models are used to model contextual information such as dialogue text and predict recommended items and generate responses to users.

[0006] Cue learning techniques can be used for lightweight fine-tuning of pre-trained models. By adding a small number of learnable parameters to a fixed pre-trained model, it can transfer the model to downstream tasks, significantly reducing the training and storage costs of deploying pre-trained models across multiple business scenarios. In this invention, cue learning can model the semantic information of dialogue and recommendation tasks, thus enabling the solution of both tasks using only a single large pre-trained model.

[0007] Knowledge graphs have been widely used in recommender systems to enrich the semantic information of items. In this invention, knowledge graphs can help the system better understand entities mentioned in dialogues and model user preferences for these entities, thereby improving performance on dialogue and recommendation tasks.

[0008] Currently, there are two main types of dialogue recommendation systems that utilize the following technologies:

[0009] Technology 1 (Model such as) Figure 1 As shown: This paper describes a method that integrates word-oriented and entity-oriented knowledge graphs based on mutual information maximization technology. A knowledge-enhanced recommendation module and a dialogue module are designed based on this method to generate items to recommend to the user and responses to the user, respectively. First, the representations of the two knowledge graphs are pre-trained using a loss function based on mutual information maximization. Then, a gating mechanism is used to fuse word-level and entity-level information as user representations. Candidate items are ranked by matching them with item representations. Finally, a Transformer-based encoder-decoder model is used to generate responses to the user based on historical dialogues and combined with word-level and entity-level information.

[0010] Technique 1 uses different models for the dialogue module and the recommendation module. When generating responses, the dialogue module's model cannot perceive the items predicted by the recommendation module, resulting in inconsistencies between the generated responses and the recommended items, which significantly impacts the user experience.

[0011] Technology 2 (Model such as) Figure 2 (As shown): This method uses item reviews as external knowledge to enrich the semantic information of the dialogue. Specifically, first, a Transformer-based model is used to retrieve review texts that are consistent with the sentiment implied in the dialogue history. Then, entities are extracted from these reviews, and a self-attention mechanism is used to fuse the entities contained in the dialogue history as user representations. These user representations are then matched with item representations to rank candidate items. Finally, a Transformer-based encoder-decoder model is used to generate a reply to the user based on the dialogue history, combined with entity information and review texts.

[0012] Technique 2 uses comment text as external knowledge, which requires maintaining a database of item comments, resulting in additional overhead. In addition, due to exposure bias, comment data for some items is very limited, and for new items, there is no usable comment data at all. These problems limit the scalability and maintainability of Technique 2.

[0013] The information disclosed in this background section is intended only to enhance the understanding of the general background of the invention and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0014] The purpose of this invention is to provide a dialogue recommendation system based on a cue learning method. The system is based on a large-scale pre-trained model and drives the model to generate high-quality response statements and recommended items that are consistent with the results by adding task-shared and task-specific cue.

[0015] This invention first integrates semantic information of entities and text in a dialogue through an entity-oriented pre-training task, using it as a shared prompt for both the dialogue and recommendation tasks. Then, it adds specific continuous prompts to the dialogue task, driving the model to generate a template with item slots as a response. Finally, it uses this template and the recommendation task-specific continuous prompts together to drive the model to generate a list of recommended items. Based on the number of slots, it selects the item with the highest recommendation probability to fill the template, thus obtaining the final response statement.

[0016] To achieve the above objectives, the present invention adopts the following technical solution:

[0017] This invention provides a dialogue recommendation system based on a prompting learning method, characterized by the following steps:

[0018] S1: By fusing semantic information from dialogue text and knowledge graphs through self-supervised learning as prompts for task sharing, the pre-trained model is supplemented with the knowledge needed for dialogue recommendation systems.

[0019] S2: Through dialogue task-specific prompt design, drive the pre-trained model to generate template statements with item slots as intermediate results for responses;

[0020] S3: By recommending task-specific prompts, the pre-trained model is driven to generate items that the user is interested in.

[0021] As a further technical solution, step S1 specifically includes:

[0022] We extract two basic semantic units, words and entities, from dialogue text to represent text-level and item-level knowledge, respectively. We also design a cross-fusion mechanism and a pre-training task based on self-supervised learning to fuse these two types of knowledge.

[0023] First, semantic units, namely words and entities, are encoded. For words, a pre-trained RoBERTa model based on bidirectional Transformer is used as the encoder. To reduce training and storage overhead, the weights of the RoBERTa model are kept fixed, and the encodings of each word are concatenated to obtain a text-level representation. ( It represents the word i, n W (This refers to the number of words); for entities, entity linking technology is used to link the dialogue text to the knowledge graph DBpedia, and the representation of each entity is obtained by encoding it through the relational graph neural network R-GCN, and then concatenated to obtain the item-level representation. ( It is the representation of entity i, n E (This refers to the number of entities);

[0024] After obtaining the above two levels of representation, in order to bridge the semantic gap between them, a cross-fusion mechanism based on bilinear transformation is used to associate these two semantic units:

[0025]

[0026]

[0027]

[0028] Here, A is the similarity matrix between the two representations, which is obtained by introducing the transformation matrix W1. and These correspond to the semantically fused word representation and entity representation, respectively;

[0029] To learn the parameters introduced by the above process, a self-supervised task is designed, which allows the model to predict the removed parts of linked entities based on the dialogue text. The calculation method is as follows:

[0030]

[0031] Where e represents the entity to be predicted, It is a dialog context with added entity representations. h is the context representation obtained by encoding the model. e It is the representation of entity e obtained after semantic fusion;

[0032] The cross-entropy loss function is used to optimize the parameter Θ introduced by semantic fusion. fuse Θ plm Keep it fixed; no optimization required.

[0033] As a further technical solution, step S2 specifically includes:

[0034] The dialogue task aims to generate informative statements to understand user preferences or recommend items of interest. By enhancing word-level semantic information in prompts, the responses become more relevant to the context. The specific design is as follows:

[0035]

[0036] in, It is a text representation that has undergone semantic fusion, P gen C is a dialogue task-specific continuous prompt, and C is the dialogue history text;

[0037] In the above prompt design, optimize P gen Parameter Θ gen Cross-entropy is used as the loss function, and the specific calculation process is as follows:

[0038]

[0039] Where N is the number of training samples, l i W is the length of the response statement in the i-th sample. ＜j It refers to all words preceding the j-th position in the reply.

[0040] As a further technical solution, step S2 also includes: sharing intermediate results; specifically:

[0041] The special word "[ITEM]" is added to the vocabulary of the pre-trained model as an item slot, and all items in the responses of the training samples are replaced with "[ITEM]". At each time step, the pre-trained model will generate "[ITEM]" or other words from the vocabulary; the items to be recommended generated by the recommendation module will be used to fill these slots.

[0042] As a further technical solution, step S3 specifically includes:

[0043] By enhancing entity-level semantic information in the prompts, the recommended items better meet user needs. The specific design is as follows:

[0044]

[0045] in, It is an entity representation that has undergone semantic fusion, P recIt is a continuous prompt specific to the recommended task, C is the dialogue history text, and S is the recommended template with item slots generated by the dialogue module;

[0046] In the above prompt design, optimize P rec Parameter Θ rec Cross-entropy is used as the loss function, and the specific calculation process is as follows:

[0047]

[0048] Where N is the number of training samples, M is the number of items, and y j,i Let 1 represent that item i is a recommended item in the j-th training sample, and Pr j (i) is calculated as follows:

[0049] Pr j (i) = Softmax(h) u ·h i )

[0050] in, Through the The representation of the dialogue context obtained through pre-trained model encoding and pooling operations, h i It is the representation of item i obtained through semantic fusion.

[0051] By adopting the above technical solution, the present invention has the following beneficial effects:

[0052] 1) By using prompting learning technology, the pre-trained model can be transferred to two downstream tasks with a fixed model. This not only utilizes the powerful language understanding and generation capabilities of the pre-trained model, but also allows for the sharing of model parameters, reducing the fine-tuning and storage costs of large models. 2) The response templates generated in the dialogue task are used as part of the prompts to supplement the contextual information of the recommendation task, thereby better helping the model to complete the task using its own knowledge. Attached Figure Description

[0053] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0054] Figure 1 This is a model diagram of existing technology one;

[0055] Figure 2 This is a model diagram of existing technology two;

[0056] Figure 3 This is a model diagram of the dialogue recommendation system of the present invention;

[0057] Figure 4 This is a schematic diagram of the core device of the dialogue recommendation system of the present invention;

[0058] Figure 5 The following is a sample diagram of a real dialogue from this invention. Detailed Implementation

[0059] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0061] Combination Figure 3 As shown in the figure, this embodiment provides a dialogue recommendation system based on a prompting learning method, and the specific process is as follows:

[0062] First, semantic information from dialogue text and knowledge graphs is fused through self-supervised learning as prompts for task sharing, supplementing the pre-trained model with the knowledge needed for the dialogue recommendation system. The specific process is as follows:

[0063] This invention extracts two basic semantic units, words and entities, from dialogue text to represent text-level and item-level knowledge, respectively. It also designs a cross-fusion mechanism and a pre-training task based on self-supervised learning to fuse these two types of knowledge.

[0064] Specifically, this invention first encodes two semantic units: words and entities. For words, a pre-trained RoBERTa model based on bidirectional Transformer is used as the encoder. To reduce training and storage overhead, the weights of the RoBERTa model are kept fixed, and the encodings of each word are concatenated to obtain a text-level representation. For entities, we use entity linking technology to link the dialogue text to the knowledge graph DBpedia, and encode each entity using the relational graph neural network R-GCN to obtain a representation, which is then concatenated to obtain an item-level representation.

[0065] After obtaining the above two levels of representation, in order to eliminate the semantic gap between them, this invention designs a cross-fusion mechanism based on bilinear transformation to associate these two semantic units:

[0066]

[0067]

[0068]

[0069] Here, A is the similarity matrix between the two representations, which is obtained by introducing the transformation matrix W1. and These correspond to the word representation and entity representation after semantic fusion, respectively.

[0070] To learn the parameters introduced in the above process, we designed a self-supervised task that allows the model to predict the removed parts of linked entities based on the dialogue text. The calculation method is as follows:

[0071]

[0072] Where e represents the entity to be predicted, It is a dialog context with added entity representations. h is the context representation obtained by encoding the model. e This is the representation of entity e obtained after semantic fusion. This invention uses the cross-entropy loss function to optimize the parameter Θ introduced by semantic fusion. fuse Θ plm Keep it fixed; no optimization required.

[0073] Secondly, through dialogue task-specific prompt design, the pre-trained model is driven to generate template statements with item slots as intermediate results for responses. The specific process is as follows:

[0074] The dialogue task aims to generate informative statements to obtain user preferences or recommend items of interest to users. To this end, this invention enhances word-level semantic information in prompts, making the response statements more relevant to the context. The specific design is as follows:

[0075]

[0076] in, It is a text representation that has undergone semantic fusion, P gen C is a dialogue task-specific continuous prompt, and C is the dialogue history text.

[0077] In the above prompt design, only P needs to be optimized. gen Parameter Θ gen We use cross-entropy as the loss function, and the specific calculation process is as follows:

[0078]

[0079] Where N is the number of training samples, l i W is the length of the response statement in the i-th sample. ＜j It refers to all words preceding the j-th position in the reply.

[0080] To address the inconsistency between predictions from the dialogue and recommendation modules in existing technologies, this technology shares intermediate results in addition to sharing the pre-trained model. Specifically, the dialogue task no longer generates complete sentences, but rather templates with item slots. These templates are used as part of the recommendation task's prompts, supplementing the context needed for the recommendation task and better assisting the pre-trained model in generating context-relevant items. In implementation, this invention adds the special word "[ITEM]" as an item slot to the vocabulary of the pre-trained model and replaces all items in the training sample responses with "[ITEM]". At each time step, the pre-trained model generates either "[ITEM]" or other words from the vocabulary. The items to be recommended generated by the recommendation module are then used to fill these slots.

[0081] Finally, by providing task-specific suggestions, the pre-trained model is driven to generate items that the user is interested in. The specific process is as follows:

[0082] The recommendation task aims to predict items that a user may be interested in, and user preferences are often reflected through entities. Therefore, this invention enhances the semantic information at the entity level in the prompts, so that the recommended items better meet the user's needs. The specific design is as follows:

[0083]

[0084] in, It is an entity representation that has undergone semantic fusion, P rec It is a continuous prompt specific to the recommended task, C is the dialogue history text, and S is the recommended template with item slots generated by the dialogue module.

[0085] Similarly, in the above prompt design, only P needs to be optimized. rec Parameter Θ rec We use cross-entropy as the loss function, and the specific calculation process is as follows:

[0086]

[0087] Where N is the number of training samples, M is the number of items, and y j,i Let 1 represent that item i is a recommended item in the j-th training sample, and Pr j (i) is calculated as follows:

[0088] Pr j(i) = Softmax(h) u ·h i )

[0089] in, Through the The representation of the dialogue context obtained through pre-trained model encoding and pooling operations, h i It is the representation of item i obtained through semantic fusion.

[0090] This invention employs a pipeline approach in its application: First, an entity linking tool is used to link the dialogue text to the knowledge graph DBpedia; then, dialogue task-specific hints are added to the pre-trained model, driving the model to generate response templates with item slots; next, recommendation task-specific hints are added to the pre-trained model, driving the model to generate recommended items; finally, based on the number of slots in the template, the items with the highest probability are selected from the list of recommended items and filled into the slots to obtain the final response statement.

[0091] Combination Figure 4 As shown, the core device of the present invention includes: an entity linker, a cue encoder, and a pre-trained model.

[0092] The entity linker takes dialogue text as input and extracts entities contained in the text that exist in the knowledge graph (DBpedia) through the entity linking algorithm, which are used as item-level semantic units in the prompt.

[0093] The main components of the prompt encoder are a semantic fusion module based on RoBERTa and R-GCN and a task-specific prompt encoder. The input is dialogue text and the entities in it. After the semantic fusion module, the entity-level and text-level representations are obtained and used as prompts shared by the two tasks. For dialogue tasks, it is concatenated with task-specific continuous prompts and dialogue history. For recommendation tasks, it is concatenated with task-specific continuous prompts, dialogue history and response templates with item slots.

[0094] The main component of the pre-trained model is a large-scale pre-trained language model, DialoGPT. The input is the prompt encoder based on the corresponding input provided by the task. For dialogue tasks, the output is a template with item slots. For recommendation tasks, the output is the item to be recommended. Finally, the item is filled into the slot in the template to obtain the final response statement.

[0095] Combination Figure 5As shown, a user dislikes animated films but prefers movies starring Julia Roberts. The actual response recommends another movie starring Julia Roberts. This invention first uses entity linking technology to identify entities in the dialogue text and links them to the knowledge graph DBpedia. Then, through semantic fusion technology, it captures the dialogue semantics and the key entity Julia Roberts as a prompt for the dialogue task, generating a response template containing this entity. This template is then input as part of the prompt into a pre-trained model for recommendation, generating another movie starring Julia Roberts, My Best Friend's Wedding (1997).

[0096] Compared to the existing technology, this invention introduces prompting learning technology, which unifies the modeling of dialogue and recommendation tasks through a pre-trained model. This allows for modeling the semantic information contained in the dialogue at both the text and item levels, and also enables a single model to complete both downstream tasks, bridging the semantic gap caused by the use of two separate models in the first technology. Furthermore, this invention alleviates the inconsistency in generated results caused by the two models by sharing a response template as an intermediate result, thus ensuring a better user experience.

[0097] Compared to the prior art, this invention does not require maintaining a database of item reviews. Furthermore, the prompt learning technique can fix the pre-trained model, and only needs to learn the corresponding prompt parameters for different application scenarios to complete model transfer. All of these factors significantly enhance the transferability and scalability of this invention.

[0098] In summary, this invention introduces prompting learning technology and solves the two tasks of dialogue and recommendation through a large-scale pre-trained model, which allows model parameters to be shared as much as possible. On this basis, the intermediate result of response templates is further shared. These innovations make the model's prediction results more consistent in the two tasks and effectively improve the user experience of the dialogue recommendation system.

[0099] Furthermore, this invention integrates semantic information from both dialogue text and item levels as part of the prompts using self-supervised learning technology, enabling the model to better understand user preferences in the dialogue, thereby improving the effectiveness of response generation and item recommendation, and effectively enhancing the service quality of the dialogue recommendation system.

[0100] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A dialogue recommendation system based on a prompting learning method, characterized in that, Includes the following steps: S1: By fusing semantic information from dialogue text and knowledge graphs through self-supervised learning as prompts for task sharing, the pre-trained model is supplemented with the knowledge needed for dialogue recommendation systems. S2: Through dialogue task-specific prompt design, drive the pre-trained model to generate template statements with item slots as intermediate results for responses; S3: Drive the pre-trained model to generate items that the user is interested in by recommending task-specific prompts; Step S2 is as follows: The dialogue task aims to generate informative statements to understand user preferences or recommend items of interest. By enhancing word-level semantic information in prompts, the responses become more relevant to the context. The specific design is as follows: ； in, It is a text representation that has undergone semantic fusion. It is a continuous prompt specific to the dialogue task. It is a dialogue history text; In the above prompt design, optimization parameters Cross-entropy is used as the loss function, and the specific calculation process is as follows: ； in, It is the number of training samples. It is the first The length of the response statements in each sample It is the first reply All words preceding that position.

2. The dialogue recommendation system based on the prompting learning method according to claim 1, characterized in that, Step S1 is as follows: We extract two basic semantic units, words and entities, from dialogue text to represent text-level and item-level knowledge, respectively. We also design a cross-fusion mechanism and a pre-training task based on self-supervised learning to fuse these two types of knowledge. First, semantic units, namely words and entities, are encoded. For words, a pre-trained RoBERTa model based on bidirectional Transformer is used as the encoder. To reduce training and storage overhead, the weights of the RoBERTa model are kept fixed, and the encodings of each word are concatenated to obtain a text-level representation. ;in, It is a word The expression, This refers to the number of words; for entities, entity linking technology is used to link the dialogue text to the knowledge graph DBpedia, and the representation of each entity is obtained by encoding it through the relational graph neural network R-GCN, and then concatenated to obtain the item-level representation. It is a physical entity The expression, It is the number of entities; After obtaining the above two levels of representation, in order to bridge the semantic gap between them, a cross-fusion mechanism based on bilinear transformation is used to associate these two semantic units: ； ； ； in, It is the similarity matrix between these two representations, which is obtained by introducing a transformation matrix. get, and These correspond to the semantically fused word representation and entity representation, respectively; To learn the parameters introduced by the above process, a self-supervised task is designed, which allows the model to predict the removed parts of linked entities based on the dialogue text. The calculation method is as follows: ； in, Indicates the entity to be predicted. It is a dialog context with added entity representations. It is the context representation obtained by encoding the model. It is a physical entity The representation obtained after semantic fusion; The cross-entropy loss function is used to optimize the parameters introduced by semantic fusion. , Keep it fixed; no optimization required.

3. The dialogue recommendation system based on the prompting learning method according to claim 1, characterized in that, Step S2 also includes: sharing intermediate results; specifically: Add to the vocabulary of the pre-trained model This special word is used as an item slot, and all items in the training sample responses are replaced with it. The pre-trained model will generate at each time step Or other words from the vocabulary; the items to be recommended generated by the recommendation module will be used to fill these slots.

4. The dialogue recommendation system based on the prompting learning method according to claim 1, characterized in that, Step S3 is as follows: By enhancing entity-level semantic information in the prompts, the recommended items better meet user needs. The specific design is as follows: ； in, It is an entity representation that has undergone semantic fusion. It is a series of prompts specific to the recommended task. It is a dialogue history text. It is a recommended template with item slots generated by the dialogue module; In the above prompt design, optimization parameters Cross-entropy is used as the loss function, and the specific calculation process is as follows: ； in, It is the number of training samples. It refers to the number of items. Representative items It is the first Items recommended from the training samples. The calculation method is as follows: ； in, Through the The representation of the dialogue context obtained through pre-trained model encoding and pooling operations. It is an item The representation obtained through semantic fusion; , Parameters introduced for semantic fusion.

Images