Cross-language information retrieval system and device based on shared semantic model

Abstract

This invention provides a cross-language information retrieval system and device based on a shared semantic model. The system includes a monolingual semantic space construction module, a model construction module, and an application module. This invention can improve the semantic sharing performance between Chinese and Indo-European languages ​​rooted in English and French, which is beneficial for meeting the multilingual needs of domestic and international users and fully realizing the value of multilingual information, thus improving the experience of discovering and acquiring domain resources.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of information retrieval technology, and in particular to a cross-language information retrieval system and device based on a shared semantic model. Background Technology

[0002] Most existing cross-language information retrieval systems are based on machine translation, and translation theories that emphasize formal transformation over semantic transformation cannot reflect the essential characteristics of cross-language communication. With the rapid development of deep learning technology, distributed embeddings of natural language vocabulary can be accurately generated on deep models; these embeddings are called word embeddings. Word embeddings capture the linguistic regularities of the trained language, and these regularities transfer between different languages. Chinese, English, and French are the three languages ​​with the widest coverage and the most countries represented by the United Nations, playing a vital role in international communication. Existing methods are not effective in cross-language communication between languages ​​with heterogeneous feature spaces, such as Chinese and Indo-European languages ​​like English and French. The issue of semantic information sharing between Chinese, English, and French languages ​​urgently needs further research. Domain information faces challenges such as the difficulty in extracting domain semantic relations and the difficulty in easily obtaining domain-specific bilingual supervised corpora, posing difficulties for cross-language information retrieval. Especially in the field of public culture, cross-language information retrieval research is relatively weak in existing literature, and there is a lack of one-stop resource discovery platforms for multilingual users in practical applications: existing national-level public digital cultural service platforms do not support cross-language information retrieval, and resources cannot be accessed by multilingual users. This results in a very low adoption rate of relevant resources, failing to meet the current needs and requirements for cross-language information retrieval in the public cultural sphere. Therefore, developing a cross-language information retrieval system and device based on a shared semantic model to effectively overcome the shortcomings of the aforementioned related technologies has become an urgent technical problem to be solved by the industry. Summary of the Invention

[0003] To address the aforementioned problems in existing technologies, embodiments of the present invention provide a cross-language information retrieval system and device based on a shared semantic model.

[0004] Firstly, embodiments of the present invention provide a cross-language information retrieval system based on a shared semantic model, comprising: a monolingual semantic space construction module, used to construct a domain-oriented Chinese semantic space; before constructing a Chinese-English-French shared semantic model, a domain-oriented Chinese semantic space is constructed; through transfer training, semantic relationships between words in the public cultural domain are captured in the Chinese semantic space and projected into the shared semantic space; a model construction module, used to construct a Chinese-English-French cross-language shared semantic model; the model construction and verification part uses only a large monolingual corpus to construct a Chinese-English-French cross-language shared semantic embedding space that does not rely on parallel text; the core content is an extension and improvement of the MUSE project, proposing a shared semantic model; and an application module, used to apply the shared semantic model to realize cross-language information retrieval in the public cultural domain; with the final application of cross-language retrieval as the guide, a cross-language retrieval framework taking the public cultural domain as an example is constructed.

[0005] Based on the above system embodiments, the cross-language information retrieval system based on a shared semantic model provided in this embodiment of the invention uses transfer training corpus in the public culture domain of the monolingual semantic space construction module, which comes from the National Public Culture Cloud Platform. This includes introductions to cultural centers, libraries, museums, art galleries, and grassroots cultural centers, as well as details of folk customs and activities. The system completes the acquisition, preprocessing, and storage of relevant corpus data, and achieves the initial construction of a small corpus.

[0006] Based on the above system embodiments, the cross-language information retrieval system based on a shared semantic model provided in this embodiment of the invention generates a multilingual shared word embedding semantic space in the model building module. The cross-language transfer of word embeddings aims to establish semantic mapping relationships between words in different languages ​​by learning the transformation function of the corresponding word embedding space.

[0007] Based on the above system embodiments, the cross-language information retrieval system based on a shared semantic model provided in this embodiment of the invention decomposes the model construction module into three steps: cross-language adversarial learning, Proxtes solution fine-tuning, and pseudo-embedding and vector optimization, which correspond to the generation, adjustment, and optimization of the shared semantic space, respectively. The specific method for generating the shared semantic space includes: using a linear projection with a transformation matrix W, a vector space of one language can be transformed into a space of another language, let X = {x1,...,x...} n} and Y = {y1,...,y m Let} represent two sets of n and m word embeddings from the source language and the target language, respectively. Let WXi be a mapping of the source language word embeddings to share the semantic space with the target language. From the mapping WX = {W x1 ,...,W xnWe randomly sample elements from Y and Y, and train a model to identify and distinguish elements from different sources. This model is called a discriminator, and its parameters are θ. D The probability that the discriminator determines vector z to be a mapping of the source language word embedding is: The probability that vector z is a target language word embedding is: If the discriminator's cost function is determined, then it is the sum of the losses from misclassifications.

[0008]

[0009] By training W to interfere with the discriminator and prevent it from making accurate predictions, the goal is to continuously make WX and Y more similar. The cost function of the mapping is:

[0010]

[0011] The discriminator aims to maximize its ability to identify word embeddings from WX or Y, while W aims to prevent the discriminator from effectively identifying WX and Y by making them as similar as possible.

[0012] Based on the above system embodiments, the cross-language information retrieval system based on a shared semantic model provided in this embodiment of the invention extracts a synthetic dictionary from the obtained shared embedding space using the model building module, and adjusts the initialized Indo-European shared semantic space using the Prokostian solution.

[0013] Based on the above system embodiments, the cross-language information retrieval system based on a shared semantic model provided in this embodiment of the invention includes the following steps in the model building module: constructing a cross-language shared semantic model for Chinese, English, and French: connecting two existing vectors, selecting a point on the connecting line that has the same angle, which is the point in the set of points with the minimum cosine distance that has the closest cosine similarity to the other two vectors, while maintaining the minimum sum of the pairwise Euclidean distances between the three vectors, where PE is:

[0014]

[0015] If the Chinese word lg cannot be explained in the English-French shared space due to a lexical gap, then the two nearest neighbors of lg that do not have a lexical gap problem are n. i and n j Let the terms lg and n in the CTL-PC spatial subgraph be... i ,n j The perpendicular intersection of the connecting lines is denoted as sec, and the distance from the root to the edge is distance(lg, sec). In the CLE-APE space, PE(lg) and PE(n) i ),PE(n jLet PE(sec) be the perpendicular intersection of the connecting lines. Then, the distance from PE(lg) to the bottom edge is distance(PE(lg),PE(sec)). Pseudo-embeddings of words with lexical gaps are generated using the vector triangle rule and similarity scaling.

[0016]

[0017]

[0018] PE(lg)=PE(sec)+distance(PE(lg),PE(sec))

[0019] The improved neighbor distance generation method normalizes the distance between two spatially identical word pairs through scaling, making... Multiply by the ratio of the two space sizes (scaleRate), locate the perpendicular intersection point to reduce the impact of different spatial dimensions, and connect n. i and n j Starting from lg, draw a perpendicular line to the base, intersecting at a point n. i and n j The distance ratio to the intersection point is l1:l2, where k is any Chinese word.

[0020] Based on the above system embodiments, the cross-language information retrieval system based on a shared semantic model provided in this embodiment of the invention, with the application module building a cross-language retrieval framework taking the public cultural field as an example, consists of four modules: inputting the search query, cross-language conversion, calculating scores through downstream retrieval algorithms, and outputting retrieval results. Cross-language conversion is the core functional module. According to language usage habits, search queries written in Chinese, English, and French are input respectively to obtain documents on topics of interest in the public cultural field. The input search and cross-language conversion modules are connected by two consecutive judgments: first, the language type of the search query is determined, and then the language of the current resource is determined. Based on the construction method of the shared semantic space, let English or French, which achieves cross-lingual embedding directly through unsupervised adversarial learning and Proxima Centauri fine-tuning, be one option, and Chinese, which achieves cross-lingual embedding through pseudo-embedding, be another option. Then, in the judgment stage, four retrieval-resource document matching situations will be generated. According to different matching situations, the cross-lingual retrieval framework, taking the public cultural domain as an example, adopts targeted strategies to expand the retrieval style, including K-nearest neighbor expansion and pseudo-related document expansion. After transformation, the downstream retrieval algorithm no longer contains any judgment steps, but calculates the retrieval score in the corresponding space along the original route.

[0021] Based on the above system embodiments, the cross-language information retrieval system based on a shared semantic model provided in this embodiment of the invention uses both a probabilistic statistical model to calculate and rank document relevance scores based on existing documents and queries expanded through word embeddings, and a method that transforms sentences from the bag-of-words approach used in traditional methods to a distributed approach to achieve semantic matching results. Let t be the embedding of terms obtained from the cross-language embedding space. It is N d The embedding of a document d in the shared space of a set of words is calculated as follows:

[0022]

[0023] Here, · is the semantic combination operator, which aggregates the term embeddings into a document embedding. Since the word embedding space exhibits linear linguistic regularity, and addition has demonstrated powerful performance in information retrieval tasks, vector addition is used as the combination. The query vector q is the sum of the word embeddings. The document uses a simple unweighted addition method: When using weighted addition, the embedding of each term is weighted by the inverse document frequency of the term, and the cosine similarity between the query and document embeddings in the shared cross-linguistic space is used as a score for the relevance of the search results.

[0024] In a second aspect, embodiments of the present invention provide an electronic device, comprising:

[0025] At least one processor; and

[0026] At least one memory communicatively connected to the processor, wherein:

[0027] The memory stores program instructions that are executed by the processor. The processor can execute the program instructions to perform a cross-language information retrieval system based on a shared semantic model provided by any of the various implementations of the first aspect.

[0028] Thirdly, embodiments of the present invention provide a non-transitory computer read storage medium that stores computer instructions, which enable a computer to implement a cross-language information retrieval system based on a shared semantic model provided by any of the various implementations of the first aspect.

[0029] The cross-language information retrieval system and device based on a shared semantic model provided in this invention can improve the semantic sharing performance between Chinese and Indo-European languages ​​rooted in English and French, which is conducive to meeting the multilingual needs of domestic and foreign users and fully leveraging the value of multilingual information, and improving the experience of discovering and acquiring domain resources. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is a schematic diagram of the cross-language information retrieval system based on a shared semantic model provided in an embodiment of the present invention;

[0032] Figure 2 A schematic diagram of the physical structure of an electronic device provided in an embodiment of the present invention;

[0033] Figure 3 This is a schematic diagram illustrating the operating principle of a cross-language information retrieval system based on a shared semantic model, as provided in an embodiment of the present invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. 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. The technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined to form a series of technical solutions. Such combinations are not constrained by the order of steps and / or structural composition patterns, but are based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0035] This invention provides a cross-language information retrieval system based on a shared semantic model. See also... Figure 1The system comprises: a monolingual semantic space construction module, used to construct a domain-oriented Chinese semantic space. Before building a shared semantic model in Chinese, English, and French, a domain-oriented Chinese semantic space is constructed. Through transfer training, semantic relationships between words in the public cultural domain are captured in the Chinese semantic space and projected into the shared semantic space; a model construction module, used to construct a cross-lingual shared semantic model in Chinese, English, and French. The model construction and verification part uses only a large monolingual corpus to construct a cross-lingual shared semantic embedding space in Chinese, English, and French that does not rely on parallel text. The core content is an extension and improvement of the MUSE project, proposing a shared semantic model; and an application module, used to apply the shared semantic model to achieve cross-lingual information retrieval in the public cultural domain. With the ultimate application of cross-lingual retrieval as the guide, a cross-lingual retrieval framework is built, taking the public cultural domain as an example. It should be noted that the shared semantic model is the M-APE shared semantic model, which uses and improves the method of the MUSE project to generate a shared semantic space without cross-language supervision. It introduces machine translation methods and focuses on combining domain adaptation and deep feature learning in one training process, and proposes the MUSE-Additional PseudoEmbedding (M-APE) model.

[0036] Based on the above system embodiments, as an alternative embodiment, the cross-language information retrieval system based on a shared semantic model provided in this embodiment of the invention uses transfer training corpus in the public culture domain of the monolingual semantic space construction module, which comes from the National Public Culture Cloud Platform. This includes introductions to cultural centers, libraries, museums, art galleries, and grassroots cultural centers, as well as details of folk customs and activities. The system completes the acquisition, preprocessing, and storage of relevant corpus data, and achieves the initial construction of a small corpus.

[0037] Based on the above system embodiments, as an optional embodiment, the cross-language information retrieval system based on a shared semantic model provided in this embodiment of the invention generates a multilingual shared word embedding semantic space in the model building module. The cross-language transfer of word embeddings aims to establish semantic mapping relationships between words in different languages ​​by learning the transformation function of the corresponding word embedding space.

[0038] Based on the above system embodiments, as an optional embodiment, the cross-language information retrieval system based on a shared semantic model provided in this embodiment of the invention decomposes the model construction module into three steps: cross-language adversarial learning, Proxtes solution fine-tuning, and pseudo-embedding and vector optimization, which correspond to the generation, adjustment, and optimization of the shared semantic space, respectively. The specific method for generating the shared semantic space includes: using a linear projection with a transformation matrix W, a vector space of one language can be transformed into a space of another language, let X = {x1,...,x...} n} and Y = {y1,...,ym Let} represent two sets of n and m word embeddings from the source language and the target language, respectively. Let WXi be a mapping of the source language word embeddings to share the semantic space with the target language. From the mapping WX = {W x1 ,...,W xn We randomly sample elements from Y and Y, and train a model to identify and distinguish elements from different sources. This model is called a discriminator, and its parameters are θ. D The probability that the discriminator determines vector z to be a mapping of the source language word embedding is: The probability that vector z is a target language word embedding is: If the discriminator's cost function is determined, then it is the sum of the losses from misclassifications.

[0039]

[0040] By training W to interfere with the discriminator and prevent it from making accurate predictions, the goal is to continuously make WX and Y more similar. The cost function of the mapping is:

[0041]

[0042] The discriminator aims to maximize its ability to identify word embeddings from WX or Y, while W aims to prevent the discriminator from effectively identifying WX and Y by making them as similar as possible.

[0043] Based on the above system embodiments, as an optional embodiment, the cross-language information retrieval system based on a shared semantic model provided in this embodiment of the invention extracts a synthetic dictionary from the obtained shared embedding space using the Prokostian solution, and adjusts the initialized Indo-European shared semantic space.

[0044] Based on the above system embodiments, as an optional embodiment, the cross-language information retrieval system based on a shared semantic model provided in this embodiment of the invention includes the following in the model building module: constructing a cross-language shared semantic model for Chinese, English, and French: connecting two existing vectors, selecting a point on the connecting line that has the same angle, this point being the point in the set of points with the minimum cosine distance that has the closest cosine similarity to the other two vectors, while maintaining the minimum sum of the pairwise Euclidean distances between the three vectors, where PE is:

[0045]

[0046] If the Chinese word lg cannot be explained in the English-French shared space due to a lexical gap, then the two nearest neighbors of lg that do not have a lexical gap problem are n. i and n j Let the terms lg and n in the CTL-PC spatial subgraph be... i ,n jThe perpendicular intersection of the connecting lines is denoted as sec, and the distance from the root to the edge is distance(lg, sec). In the CLE-APE space, PE(lg) and PE(n) i ),PE(n j Let PE(sec) be the perpendicular intersection of the connecting lines. Then, the distance from PE(lg) to the bottom edge is distance(PE(lg),PE(sec)). Pseudo-embeddings of words with lexical gaps are generated using the vector triangle rule and similarity scaling.

[0047]

[0048]

[0049] PE(lg)=PE(sec)+distance(PE(lg),PE(sec))

[0050] The improved neighbor distance generation method normalizes the distance between two spatially identical word pairs through scaling, making... Multiply by the ratio of the two space sizes (scaleRate), locate the perpendicular intersection point to reduce the impact of different spatial dimensions, and connect n. i and n j Starting from lg, draw a perpendicular line to the base, intersecting at a point n. i and n j The distance ratio to the intersection point is l1:l2, where k is any Chinese word. It should be noted that the cross-language embedding space output by the model is called the ClossLanguageEmbedding-AdditionalPseudoEmbedding (CLE-APE) space.

[0051] Based on the above system embodiments, as an alternative embodiment, the cross-language information retrieval system based on a shared semantic model provided in this embodiment of the invention, in its application module, establishes a cross-language retrieval framework taking the public cultural field as an example. This framework consists of four modules: inputting the search query, cross-language conversion, calculating scores through downstream retrieval algorithms, and outputting retrieval results. Cross-language conversion is the core functional module. According to language usage habits, search queries written in Chinese, English, and French are input respectively to obtain documents on topics of interest in the public cultural field. The input search and cross-language conversion modules are connected by two consecutive judgments: first, determining the language type of the search query, and then determining the current... Based on the shared semantic space construction method, let English or French, which achieves cross-language embedding directly through unsupervised adversarial learning and Proxima Centauri fine-tuning, be one option, and Chinese, which achieves cross-language embedding through pseudo-embedding, be another option. Then, in the judgment stage, four retrieval-resource document matching scenarios will be generated. According to different matching scenarios, the cross-language retrieval framework, taking the public culture domain as an example, adopts targeted strategies to expand the retrieval scenarios, including K-nearest neighbor expansion and pseudo-related document expansion. After transformation, the downstream retrieval algorithm no longer includes any judgment steps in the score calculation section, but instead calculates the retrieval score in the corresponding space along the original route.

[0052] Based on the above system embodiments, as an alternative embodiment, the cross-language information retrieval system based on a shared semantic model provided in this embodiment of the invention uses a module that directly employs a probabilistic statistical model to calculate and rank document relevance scores based on existing documents and queries expanded through word embeddings. It also utilizes a distributed approach to transform sentences from the bag-of-words method used in traditional methods to achieve semantic matching results. Let t be the embedding of terms obtained from the cross-language embedding space. Let d be a document in a set of Nd words. The embedding of document d in the shared space is calculated as follows:

[0053]

[0054] Here, · is the semantic combination operator, which aggregates the term embeddings into a document embedding. Since the word embedding space exhibits linear linguistic regularity, and addition has demonstrated powerful performance in information retrieval tasks, vector addition is used as the combination. The query vector q is the sum of the word embeddings. The document uses a simple unweighted addition method: When using weighted addition, the embedding of each term is weighted by the inverse document frequency of the term, and the cosine similarity between the query and document embeddings in the shared cross-linguistic space is used as a score for the relevance of the search results.

[0055] The cross-language information retrieval system and device based on a shared semantic model provided in this invention can improve the semantic sharing performance between Chinese and Indo-European languages ​​rooted in English and French, which is conducive to meeting the multilingual needs of domestic and foreign users and fully leveraging the value of multilingual information, and improving the experience of discovering and acquiring domain resources.

[0056] In another embodiment, the cross-language information retrieval system based on a shared semantic model includes:

[0057] Step S101: The shared semantic model initializes and generates a shared semantic space for Indo-European languages ​​by performing adversarial learning on existing monolingual semantic spaces with different distributions.

[0058] Step S102: The shared semantic model extracts a synthetic dictionary from the obtained shared embedding space and adjusts the initialized Indo-European shared semantic space using the Procosts solution.

[0059] Step S103: The pseudo-embedding step proposed in this invention will linearly combine the English and French vectors in the Indo-European shared semantic space generated in the first two steps to optimize the cross-language embedding of Chinese and generate the final version of the cross-language shared semantic space.

[0060] Step S104: Based on the shared semantic model, cross-language conversion between query and resource text is achieved, enabling different languages ​​to be used in the same space. This facilitates subsequent retrieval algorithms in calculating the similarity score between the query and the resource. The algorithm here... Figure 3 The meeting will provide a detailed description.

[0061] Step S105: Based on the shared semantic model, after the user inputs a query, the search results are calculated for similarity and reordered. The algorithm here is... Figure 3 It will be described in detail.

[0062] Figure 3 An embodiment of cross-language information retrieval research based on a shared semantic model according to the present invention is described in detail, including:

[0063] Step S201: This invention uses the FastText method to load embedding vectors from a pre-computed embedding space and performs transfer learning through a domain-based corpus. The resulting Chinese semantic space will help to better represent domain semantic relationships. Before developing the pre-trained language model, the data is filtered and cleaned, including but not limited to using regular expressions to remove numbers, emojis, punctuation marks, URLs, HTML tags, and symbolic symbols from the text, in order to reduce noise and memory consumption in the text content and improve training speed. In subsequent steps, this paper further projects the domain semantic relationships captured in Chinese onto a shared semantic space of Chinese, English, and French.

[0064] In step S202, the present invention uses the three monolingual spaces corresponding to the Chinese monolingual semantic space learned after transfer training, the English monolingual semantic space pre-trained by FastText, and the French monolingual semantic space pre-trained by FastText as input to the shared semantic space.

[0065] Step S203: The shared semantic model initially generates a shared semantic space. A mapping W is learned to make the source language space closer to the target language space; a discriminator is learned so that, even with interference from the mapping W, it can distinguish whether the sampled word embeddings come from the source language space or the target language space. The stronger the discriminator's recognition ability, the more the mapping W can confuse the discriminator, making the two spaces increasingly similar. At this point, it is equivalent to generating a shared, unified space where different spaces tend to overlap.

[0066] In step S204, the shared semantic model is fine-tuned using Proxima C++ to change the spatial metric via CSLS, generating a new bidirectional dictionary. Subsequently, adversarial training is applied to update W using this generated dictionary. The training results improve the performance of low-frequency words and increase similarity related to isolated word embeddings.

[0067] In step S205, after generating the Indo-European shared semantic space, the shared semantic model obtains Chinese word embeddings for cross-linguistic retrieval through linear calculations of English and French word embeddings. By considering both spatial distance and perspective, the newly generated vector is made to be most similar to the two existing vectors, with minimal difference. For words with lexical gaps, the semantic relationships captured in the Chinese space are projected into the Indo-European shared semantic space using a neighbor distance generation method, generating a cross-linguistic shared semantic space.

[0068] Step S206: After three steps of generation, adjustment, and optimization, the shared semantic model outputs a shared semantic space, which is used for subsequent cross-language information retrieval.

[0069] In step S207, the user enters search queries written in Chinese, English, and French, respectively, according to their language usage habits.

[0070] Step S208 connects the input retrieval and cross-language conversion sections with two consecutive judgments: first, the language type of the retrieval expression is determined, and then the language type of the current resource is determined. Based on the shared semantic space construction method, let English, French, and Indo-European languages, which achieve cross-language embedding directly through unsupervised adversarial learning and Proxima Centauri fine-tuning, be one option; let Chinese, which achieves cross-language embedding through pseudo-embedding, be another option. Then, the judgment stage will generate the above four retrieval expression-resource document matching scenarios. Based on different matching scenarios, the framework adopts targeted strategies to expand the retrieval expression, including the K-nearest neighbor expansion and pseudo-related document expansion mentioned in Section 2.3.1. It is worth noting that, to fully utilize information from the public cultural domain, when both the retrieval expression and the resource language are Chinese, the retrieval expression expansion directly uses Chinese word embeddings from the Chinese semantic space oriented towards the public cultural domain, after transfer training. For the other three cases, the retrieval expression expansion will use Chinese-English-French word embeddings from the shared semantic space. After conversion, the "Calculate Score via Downstream Retrieval Algorithm" section no longer contains any judgment steps; instead, it calculates the retrieval score in the corresponding space along the original route.

[0071] Step S209: The score calculation directly uses a probabilistic statistical model, based on existing documents and queries expanded through word embeddings, to calculate and rank document relevance scores under probabilistic statistics; it also uses a method that transforms sentences or documents from the bag-of-words approach used in traditional methods to the distributed approach used in deep learning methods to achieve better semantic matching results.

[0072] Step S210: Sort and return the search results.

[0073] The cross-language information retrieval technology based on a shared semantic model proposed in this invention comprises several algorithms and computational processes. The first is the transfer training computation process, where vector dimension, number of epochs, learning rate, and sub-word length are four important basic hyperparameters tuned in FastText technology. The length of a word's vector size is called the vector dimension; larger vectors carry more data but are more difficult to train, time-consuming, and require more data. Epochs are the number of times the model trains on a batch of data; the larger the corpus, the fewer iterations are possible. The learning rate controls the magnitude of weight updates, as well as the training speed and accuracy. Sub-word length specifies the length of the substrings to be considered, used for different processing tasks, such as parsing words from a vocabulary. This invention uses FastText's default parameters. Through transfer training, a domain-adapted Chinese corpus with a dimension of 300 is obtained after training. French and English vectors use unsupervised word embeddings with a dimension of 300 trained using FastText.

[0074] Secondly, during the training of the adversarial generative network, the mapping W is 300×300 in size. The mapping W is initialized with an identity matrix, where the values ​​on the main diagonal are 1 and the rest are 0. For the discriminator, this invention uses a multilayer perceptron with two hidden layers of size 2048 and a Leaky-ReLU activation function. The discriminator's input is corrupted with dropout noise of 0.1, and the discriminator's output is smoothed with a smoothing coefficient of 0.2. The cross-linguistic domain adversarial learning model employs stochastic gradient descent with a batch size of 32, a learning rate of 0.1, and a gradually decreasing step size; the learning rate for both the discriminator and W decays to 0.95. In particular, when the supervised validation criterion is reduced, the learning rate is halved to 0.5 by dividing it by 2, thus reducing the oscillation of gradient descent. The learning process is continuously repeated on different mini-batches, covering the entire training set, until the parameters converge. In each training step, the word embeddings for the discriminator are uniformly sampled.

[0075] This invention proposes to obtain pseudo-embedding vectors using geometric relationships in vector space when implementing pseudo-embedding and neighbor distance generation methods, rather than directly solving the formulas. Specifically, taking the neighbor distance generation method as an example, when searching for the intersection point sec in the CTL-PC space, a point on the line connecting ni and nj is continuously located by iteratively moving in small steps until the cosine similarity between the line connecting this point and lg and the line connecting ni and nj approaches 0 and is below a set threshold. This point is then considered the perpendicular intersection point sec we are looking for. In this process, the idea of ​​effective substitution is employed. Similarly, when using the ordinary pseudo-embedding generation method, the final word embedding with the same angle is also found step by step through iterative traversal.

[0076] Step S301: The shared semantic model generates a shared semantic space in Chinese, English, and French. Blue dots represent English, green dots represent Chinese, and red dots represent French. Visualization reveals the patterns of the space occupied by these words. On one hand, we achieve similarity between words with the same meaning in Chinese, English, and French within the cross-linguistic shared semantic space. The same meaning, regardless of the language used, is located close together, indicating high similarity. On the other hand, we achieve a comprehensive understanding of the overall meaning regardless of the language. For example, the words "father"-"père" and "daughter"-"fille" distributed in the lower left corner represent personal identities; "language"-"langue" and "ethnolinguistics"-"ethnolinguistique" in the upper right corner represent linguistic synonyms; and the semantic clusters from top to bottom in the middle represent building groups, fruits, and animals. Words with the same meaning have similar contexts and are closer together. Conversely, words with different meanings are farther apart because they have different contexts.

[0077] In step S302, when performing a foreign language search on the public cultural cloud platform using "opera," the returned search results only include one piece of information whose article title contains the keyword "opera": "Weekend Series: Weekend Music Gathering—MiniOpera 'Jiang Jie'," as shown in the left side panel. Clearly, this does not meet the needs of multilingual users for retrieving public cultural digital resources. The cross-language information retrieval proposed in this invention effectively solves this problem, also using "opera" as the search term.

[0078] Step S303 uses three high-frequency terms in the public culture field—"library-bibliothèque," "museum-musée," and "concert-concert"—as search terms. Regardless of the language used for the search, at least two relevant documents written in each of the first 10 search results will be included, and documents written in three languages ​​will be included in the first 20 search results. Unlike machine translation technology that searches and ranks based on the query terms themselves, when performing cross-language information retrieval based on a shared semantic model, the first documents retrieved in Chinese, English, and French may not be the same, but they are all correct and relevant documents that meet the requirements. For example, when conducting cross-language information retrieval of articles related to museums in the field of public culture, the first articles retrieved in English and French both had titles such as "Why was the earliest museum in China born in Shanghai?" and "Pourquoi lepremier musée de Chine est-il néàShanghai?". However, when searching for articles related to concerts in the field of public culture, the first articles retrieved in English, Chinese, and French were all different from each other.

[0079] Step S304: Improve the cross-language lexical gap problem by using a shared semantic model for cross-language information retrieval. For words with lexical gaps, different semantic dimensions are extracted from candidate words to explain the words with lexical gaps. The method of extracting semantic dimensions is as follows: among candidates with the same semantic meaning, only the top-ranked candidate is retained as the semantic dimension, and the rest of the candidates are eliminated. This process is repeated, using different semantic dimensions to obtain five semantic dimensions for English and French, with some overlap between the English and French semantic dimensions. For example, "lion dance" is a folk culture, and through the extraction of semantic dimensions from the shared semantic space, it is explained by a series of semantics such as "celebration, acrobatics, celebration activities, parade, dance, revelry, and ceremony". Again, taking "lion dance" as an example, if a user whose native language is not Chinese wants to find Chinese articles related to lion dance on the public culture cloud platform by using a combination of query terms in their native language, the user's query term combination attempts to describe the activity from different perspectives. These different perspectives of words themselves constitute semantic information, and word embedding will express their semantic information in numerical form. For example, when searching using combinations of 'celebration', 'acrobatics', and 'dance', the cross-language retrieval based on a shared semantic model embeds the query terms into a 300-dimensional shared semantic space, creating computational vectors. This shared semantic space contains vectors of all words in Chinese, English, and French. Query term expansion and document similarity calculations are then performed within this space. In the search results, 3 out of the top 5 documents contain related information about lion dance, and 5 out of the top 10 documents contain related information about lion dance.

[0080] The system in this embodiment of the invention is implemented using an electronic device; therefore, it is necessary to introduce the relevant electronic device. For this purpose, this embodiment of the invention provides an electronic device, such as... Figure 2 As shown, the electronic device includes at least one processor, a communication interface, at least one memory, and a communication bus, wherein the at least one processor, the communication interface, and the at least one memory communicate with each other via the communication bus. The at least one processor invokes logical instructions stored in the at least one memory to implement various systems provided in the system embodiments.

[0081] Furthermore, when the logical instructions in at least one of the aforementioned memories are implemented as software functional units and sold or used as independent products, they are stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, is embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (a personal computer, server, or network device) to execute all or part of the system described in the various embodiments of the present invention. The aforementioned storage medium includes: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks—various media for storing program code.

[0082] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, located in one place, or distributed across multiple network units. The purpose of this embodiment is achieved by selecting some or all of the modules according to actual needs. Those skilled in the art will understand and implement this without any inventive effort.

[0083] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment is implemented by means of software plus necessary general-purpose hardware platforms, and of course, also by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, are embodied in the form of a software product. This computer software product is stored in a computer read-only storage medium, such as ROM / RAM, magnetic disk, optical disk, and includes several instructions to cause a computer device (a personal computer, server, or network device) to implement the methods or systems described in or in part of the various embodiments.

[0084] The flowcharts and block diagrams in the accompanying drawings illustrate the achievable architecture, functions, and operations of systems, methods, and computer program products according to various embodiments of the present invention. Based on this understanding, each block in a flowchart or block diagram represents a module, segment, or portion of code containing one or more execution instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks occur in a different order than those shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, or sometimes in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, is implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0085] It should be noted that the terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising..." does not exclude the presence of the same element in the process, method, article, or apparatus that includes said element.

[0086] 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 some technical features can be replaced; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A cross-language information retrieval system based on a shared semantic model, characterized in that, include: The monolingual semantic space construction module is used to construct a domain-oriented Chinese semantic space. Before building a shared semantic model in Chinese, English and French, a domain-oriented Chinese semantic space is constructed. Through transfer training, semantic relationships between words in the public cultural domain are captured in the Chinese semantic space and projected into the shared semantic space. The model building module is used to construct a cross-lingual shared semantic model in Chinese, English, and French. The model building and validation part uses only a large monolingual corpus to construct a cross-lingual shared semantic embedding space in Chinese, English, and French that does not rely on parallel text. The core content is an extension and improvement of the MUSE project, proposing a shared semantic model. The application module is used to apply the shared semantic model to realize cross-lingual information retrieval in the public cultural field. With the ultimate application of cross-lingual retrieval as the goal, a cross-lingual retrieval framework is built with the public cultural field as an example. The model building module is decomposed into three steps: cross-lingual adversarial learning, Proxtes solution fine-tuning, and pseudo-embedding and vector optimization, corresponding to the generation, adjustment, and optimization of the shared semantic space, respectively. The specific method for generating the shared semantic space includes: using a linear projection with a transformation matrix W, it is possible to transform the vector space of one language into the space of another language, such that... and Given two sets of n and m word embeddings from the source and target languages ​​respectively, WXi is a mapping of the source language word embeddings to share the semantic space with the target language. We randomly sample elements from Y and train a model to identify and distinguish elements from different sources. This model is called a discriminator. Let the parameters of the discriminator be... The probability that the discriminator determines vector z to be a mapping of the source language word embedding is: The probability that vector z is a word embedding of the target language is: If the discriminator's cost function is determined, then it is the sum of the losses from misclassifications. By training W to interfere with the discriminator and prevent it from making accurate predictions, the goal is to continuously make WX and Y more similar. The cost function of the mapping is: The discriminator aims to maximize its ability to identify word embeddings from WX or Y, while W aims to prevent the discriminator from effectively identifying WX and Y by making them as similar as possible.

2. The cross-language information retrieval system based on a shared semantic model according to claim 1, characterized in that, The transfer training corpus for the public culture domain in the monolingual semantic space construction module comes from the National Public Culture Cloud Platform, including introductions to cultural centers, libraries, museums, art galleries, and grassroots cultural centers, as well as details of folk customs and activities. The acquisition, preprocessing, and storage of the relevant corpus are completed, and a small corpus is initially built.

3. The cross-language information retrieval system based on a shared semantic model according to claim 2, characterized in that, The model building module generates a multilingual shared word embedding semantic space. The cross-language transfer of word embeddings aims to establish semantic mapping relationships between words in different languages ​​by learning the transformation function of the corresponding word embedding space.

4. The cross-language information retrieval system based on a shared semantic model according to claim 1, characterized in that, The model building module extracts a synthetic dictionary from the obtained shared embedding space and adjusts the initialized Indo-European shared semantic space using the Procosts solution.

5. The cross-language information retrieval system based on a shared semantic model according to claim 4, characterized in that, The model building module's construction of a cross-lingual shared semantic model for Chinese, English, and French includes: connecting two existing vectors, selecting a point on the connecting line that has the same angle; this point is the one in the set of points with the minimum cosine distance that has the closest cosine similarity to the other two vectors, while minimizing the sum of the pairwise Euclidean distances among the three vectors. The PE is: If the Chinese word "lg" cannot be explained in the English-French shared space due to a lexical gap, then the two nearest neighbors of "lg" that do not have a lexical gap problem are... and Let the word lg in the CTL-PC spatial subgraph be equal to... , The perpendicular intersection of the connecting lines is denoted as sec, and the distance from the term lg to the bottom edge is distance(lg, sec). In the CLE-APE space, PE(lg) and , even Let PE(sec) be the perpendicular intersection of the lines. Then, the distance from PE(lg) to the bottom edge is distance(PE(lg),PE(sec)). Pseudo-embeddings of words with lexical gaps are generated using the vector triangle rule and similarity scaling. The improved neighbor distance generation method normalizes the distance between two spatially identical word pairs through scaling, making... Multiply by the scaleRate of the two spaces, locate the perpendicular intersection point, reduce the impact of different spatial dimensions, and connect. and Draw a perpendicular line from lg to the base, intersecting it at a single point. and The ratio of the distances to the intersection point is : , k is any Chinese word.

6. The cross-language information retrieval system based on a shared semantic model according to claim 5, characterized in that, The application module establishes a cross-language retrieval framework, using the public culture domain as an example. It consists of four sections: inputting the search query, cross-language conversion, calculating scores via downstream retrieval algorithms, and outputting search results. Cross-language conversion is the core functional section. Based on language usage habits, search queries written in Chinese, English, and French are input to obtain documents on topics of interest in the public culture domain. The input search and cross-language conversion sections are connected by two consecutive judgments: first, determining the language of the search query, and then determining the language of the current resource. Based on the shared semantic space construction method, English or French, which achieves cross-language embedding directly through unsupervised adversarial learning and Proxima Centauri fine-tuning, is one option; Chinese, which achieves cross-language embedding through pseudo-embedding, is another option. This results in four possible matching scenarios between the search query and the resource document. Based on these different matching scenarios, the cross-language retrieval framework, using the public culture domain as an example, adopts targeted strategies to expand the search query, including K-nearest neighbor expansion and pseudo-relevant document expansion. After conversion, the downstream retrieval algorithm score calculation section no longer includes any judgment steps; instead, it calculates the search score in the corresponding space along the original route.

7. The cross-language information retrieval system based on a shared semantic model according to claim 6, characterized in that, The application module directly uses probabilistic statistical models to calculate and rank document relevance scores based on existing documents and queries expanded through word embeddings. It also transforms sentences from the traditional bag-of-words approach to a distributed model to achieve semantic matching results. Let t be the embedding of terms obtained from the cross-lingual embedding space. It is by The embedding of a document d in the shared space of a set of words is calculated as follows: in, This is a semantic composition operator that aggregates the term embeddings into a document embedding. Since the word embedding space exhibits linear linguistic regularity, and addition has demonstrated powerful performance in information retrieval tasks, vector addition is used as the composition method. The query vector q is the sum of the word embeddings. The document uses a simple unweighted addition method: When using weighted addition, the embedding of each term is weighted by the inverse document frequency of the term, and the cosine similarity between the query and document embeddings in the shared cross-linguistic space is used as a score for the relevance of the search results.

8. An electronic device, characterized in that, include: At least one processor, at least one memory, a communication interface, and a bus; wherein, The processor, memory, and communication interface communicate with each other through the bus; The memory stores program instructions that are executed by the processor, which invokes the program instructions to implement the system as claimed in any one of claims 1 to 7.

9. A non-transitory computer read storage medium, characterized in that, The non-transitory computer read storage medium stores computer instructions, which cause the computer to implement the system as claimed in any one of claims 1 to 7.

Images