Method for directional entity information extraction based on large model

Abstract

The invention discloses a method for extracting directional entity information based on a large model, which is used for realizing the extraction of directional information in a certain range by fusing the capability of the large model and a retrieval result, and comprises the following steps of: S1, building a word segmentation service; s2, filtering a word segmentation result, and performing directional recall based on a filtering result; s3, recalling related entity contents of the statements based on Co < l > Bert; s4, performing result fusion on the Co < l > Bert recall and the word segmentation result recall; s5, sorting the combined contents, and extracting a TopN result; s6, the TopN is input and provided for the large model, and entity extraction is carried out based on the large model; and S7, searching and matching the identification result of the large model, and verifying the validity of the result. According to the method, the entity identification problem of different industries and scenes can be solved with lower cost, higher expansibility can be achieved, business differentiation expression can be supported in a model layer, and the complexity and difficulty of overall implementation are reduced.

Description

Technical Field

[0001] The invention belongs to the technical field of data processing, and in particular is a method for extracting directional entity information based on a large model. Background Art

[0002] Entity recognition is an important task in the field of NLP, and its effect directly affects the specific performance of related tasks. In the business of some data-based enterprises, the recognition of addresses (provinces, cities, districts, counties, streets, communities), models, series, brands, etc. is a particularly important task, and the accuracy of the recognition of related content directly affects the subsequent business processes.

[0003] In data-based enterprises whose businesses are mainly based on conversations, the effect of entity recognition directly affects the effect of conversations. The relevant information generated in the conversation directly affects the subsequent business processes. For example, in conversations in the home improvement industry, relevant orders will be dispatched based on the user's address. If there is a problem with address recognition, it will cause problems with dispatching orders. In traditional entity recognition tasks, specialized entity recognition models are often used for processing, which often have problems with high startup costs and insufficient generalization. With the emergence of large models, especially open source large models, new technical routes have been brought to entity recognition. Based on open source models, better generalization effects can be achieved, but there are still many usability issues in actual implementation, such as insufficient accuracy compared to vertical models and easy hallucinations.

[0004] Disadvantages of existing technology:

[0005] 1. The existing traditional entity recognition model has a high startup cost. That is, to train an entity recognition model, a large amount of labeled data is required for training. The cost of constructing data is relatively high, and a long startup cost is required.

[0006] 2. The generalization is weak. Traditional models are not effective in identifying long-tail problems and need to continuously enrich the training corpus to continuously improve.

[0007] 3. Poor scalability. For new entity types, it requires extremely high costs, which often means re-labeling historical data;

[0008] 4. Unable to support differentiation, that is, the same type can only have a unique answer, and business differentiation cannot be achieved. For example, in the scenario of customer A, the "BYD Qin New Energy" is expected to represent the car model "Qin New Energy", but in the scenario of customer B, the car series "BYD Qin" is expected to be extracted. Such business differentiation cannot be achieved at the model level. Summary of the invention

[0009] In view of the shortcomings of the prior art, the present invention provides a method for extracting targeted entity information based on a large model. As a new entity recognition solution, the method integrates a retrieval method and performs targeted content enhancement based on the retrieval. By performing word segmentation on the content and recalling relevant information based on the word segmentation, more relevant special content can be recalled for reference by the large model when it is used. The method integrates a sorting method to sort the retrieved and recalled content and adopts content with higher relevance. The method supports business differentiated retrieval, so that in the same context, more differentiated results can be supported, entity recognition problems in different industries and scenarios can be solved at a lower cost, higher scalability can be achieved, and business differentiated performance can be supported at the model layer, reducing the complexity of the overall implementation.

[0010] A method for extracting directional entity information based on a large model, which combines the capabilities of the large model with the search results to extract directional information within a certain range, the steps are as follows:

[0011] S1 builds word segmentation services, builds its own word segmentation module based on the word segmentation tool, and performs fast word segmentation based on industry dictionaries and business dictionaries;

[0012] S2 filters the word segmentation results and performs targeted recall based on the filtered results, that is, uses the word segmentation results to recall relevant information;

[0013] S3 recalls relevant entity content of sentences based on Col Bert;

[0014] S4 combines the results of Col Bert recall and word segmentation recall;

[0015] S5 sorts the merged content and extracts the TopN results;

[0016] S6 provides TopN as input to the big model and extracts entities based on the big model;

[0017] S7 searches and matches the recognition results of the large model to verify the validity of the results.

[0018] Preferably, Ji eba word segmentation is adopted in step S1.

[0019] Preferably, the word segmentation result in step S2 is mainly the vocabulary of the target entity category, and the semantic recall of the target entity is performed based on the word segmentation result in a vector recall manner.

[0020] Preferably, Milvus is used to perform vector search in step S2.

[0021] Preferably, in step S2, SimCSE is used to generate sentence vectors.

[0022] Preferably, the word segmentation in step S1 may use any other word segmentation tool that can customize the dictionary.

[0023] Preferably, the vector representation of the content in step S2 may adopt other embedding methods, such as the embedding method provided by the large model.

[0024] Preferably, any other engine may be used for vector recall in step S2.

[0025] Preferably, in the sentence recall in step S3, Bert can be used to replace Col Bert.

[0026] Preferably, the result sorting in step S5 can be replaced by other relevance sorting methods.

[0027] The present invention has the following beneficial effects and advantages:

[0028] 1. Low startup cost, no need for a lot of annotations, can be directly sorted based on the relevant public information, getting rid of the original method that requires a lot of special annotations;

[0029] 2. Good generalization. For long-tail problems, large models have strong generalization, which can better solve long-tail problems and improve the comprehensive capabilities of the model;

[0030] 3. Providing differentiation capabilities on the model side solves the problem of unchanged output results of traditional entity recognition models. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0032] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the invention, so the present invention is not limited by the specific implementation disclosed below.

[0033] Example 1

[0034] A method for extracting directional entity information based on a large model, such as Figure 1 As shown, the ability of the large model and the retrieval results are integrated to extract directional information within a certain range. The method includes the steps of:

[0035] 1. Build a word segmentation service and build your own word segmentation module based on common word segmentation tools. Here, Ji eba word segmentation is used to perform fast word segmentation based on common industry dictionaries and business dictionaries;

[0036] 2. Filter the word segmentation results and perform targeted recall based on the filtered results, that is, use the word segmentation results to recall relevant information. The word segmentation results here are mainly the vocabulary of the target entity category. The semantic recall of the target entity is performed based on the vector recall method based on the word segmentation results. Here, Milvus is used for vector search and SimCSE is used for sentence vector generation;

[0037] 3. Recall relevant entity content of sentences based on Col Bert;

[0038] 4. Fusion of Col Bert recall and word segmentation recall results;

[0039] 5. Sort the merged content and extract the TopN results;

[0040] 6. Provide TopN as input to the big model and perform entity extraction based on the big model;

[0041] 7. Search and match the recognition results of the large model to verify the validity of the results.

[0042] Example 2

[0043] Any other word segmentation tool that can customize the dictionary can be used for word segmentation, such as HanLP; the vector representation of the content can use other embeddings, such as the embedding provided by the big model; vector recall can use any other engine, such as facets or other vector engines; sentence recall can use Bert to replace Col Bert; result sorting can be replaced by other relevance sorting methods.

[0044] Example 3

[0045] The large model can adopt third-party large models, such as Alibaba's Qwen, Volcano Engine's Doubao, etc., or it can adopt open source models for simple fine-tuning.

[0046] It should be noted that the embodiments mentioned above illustrate rather than limit the present invention, and those skilled in the art will be able to design alternative embodiments without departing from the scope of the claims. Although the present invention is disclosed as above, the present invention is not limited thereto. Any person skilled in the art may make various changes and modifications without departing from the spirit and scope of the present invention, so the scope of protection of the present invention shall be subject to the scope defined by the claims.

Claims

1. A method for extracting directional entity information based on a large model, characterized in that: It combines the capabilities of the large model with the search results to extract directional information within a certain range. The steps are as follows: S1 builds word segmentation services, builds its own word segmentation module based on the word segmentation tool, and performs fast word segmentation based on industry dictionaries and business dictionaries; S2 filters the word segmentation results and performs targeted recall based on the filtered results, that is, uses the word segmentation results to recall relevant information; S3 recalls relevant entity content of statements based on ColBert; S4 performs result fusion on ColBert recall and word segmentation result recall; S5 sorts the merged content and extracts the TopN results; S6 provides TopN as input to the big model and extracts entities based on the big model; S7 searches and matches the recognition results of the large model to verify the validity of the results.

2. The method for extracting directional entity information based on a large model according to claim 1, characterized in that: Jieba word segmentation is adopted in step S1.

3. The method for extracting directional entity information based on a large model according to claim 1, characterized in that: The word segmentation result in step S2 is mainly the vocabulary of the target entity category, and the semantic recall of the target entity is performed based on the vector recall method for the word segmentation result.

4. The method for extracting directional entity information based on a large model according to claim 1, characterized in that: In step S2, Milvus is used to perform vector search.

5. The method for extracting directional entity information based on a large model according to claim 2, characterized in that: In the step S2, SimCSE is used to generate the sentence vector.

6. The method for extracting directional entity information based on a large model according to claim 1, characterized in that: The word segmentation in step S1 may use any other word segmentation tool that can customize the dictionary.

7. The method for extracting directional entity information based on a large model according to claim 3 is characterized in that: The vector representation of the content in step S2 may adopt other Embeddings, such as Embeddings provided by a large model.

8. The method for extracting directional entity information based on a large model according to claim 3, characterized in that: In step S2, the vector recall may use any other engine.

9. The method for extracting directional entity information based on a large model according to claim 1, characterized in that: In the step S3, the sentence recall may use Bert to replace ColBert.

10. The method for extracting directional entity information based on a large model according to claim 1, characterized in that: The result sorting in step S5 may be performed by other relevance sorting methods.

Images