AI translation system for realizing multilingual inter-translation

Abstract

The invention relates to the technical field of speech translation, in particular to an AI translation system for realizing multilingual inter-translation, which comprises a speech recognition and preliminary translation module, a translation optimization module and a translation quality evaluation module, the speech recognition and preliminary translation module is used for translating sound input by a user into characters, and the translation quality evaluation module is used for analyzing the industry of the characters to be translated; a similarity threshold value is set, translated characters output at the moment are distinguished as preliminary translated characters or optimized translated characters according to the similarity threshold value, the suitability of the translation quality and different use scenes is further guaranteed, and the similarity threshold value is reasonably set by analyzing the industry where the characters to be translated are located. Therefore, whether the preliminarily translated characters or the optimized translated characters meet the requirements of the corresponding scenes or not is flexibly distinguished, the defects in the prior art when diversified translation requirements are met are overcome, the requirement for natural and smooth communication in the daily dialogue scene can be met, and the quality and practicability of speech translation services in different scenes are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of speech translation, and more specifically, to an AI translation system for realizing multilingual translation. Background Art

[0002] Speech recognition technology is a technology that converts human speech into text. By analyzing, processing, and recognizing speech signals, human language is converted into text information that can be understood and processed by a computer. With the continuous progress of computer technology, signal processing technology, and artificial intelligence, speech recognition technology has been widely used;

[0003] However, at present, speech recognition technology has many limitations in the face of complex actual application scenarios. When users translate speech into text, because users need to translate in different usage scenarios, in daily conversation scenarios, even if language users in different regions express the same semantic content, there may be obvious differences in aspects such as their speech word order and word usage preferences. Most of the existing speech translation technologies only recognize speech and convert it into text according to a fixed pattern, and then simply perform literal translation, resulting in the translated text being rigid and not conforming to the natural expression habits of a specific region, causing obstacles for the recipient to understand. In the aspect of professional document translation, the requirements for accuracy, professionalism, and compliance with industry norms are extremely high. In view of this, we propose an AI translation system for realizing multilingual translation. Summary of the Invention

[0004] The purpose of the present invention is to solve the problem that text cannot be translated according to the user's translation scenario.

[0005] To achieve the above purpose, the present invention provides an AI translation system for realizing multilingual translation, including a speech recognition and preliminary translation module, a translation optimization module, and a translation quality evaluation module;

[0006] The speech recognition and preliminary translation module uses a web crawler to collect a large number of spectra and corresponding words corresponding to known different language word contents, constructs an acoustic model of different language word contents and corresponding spectra, collects the sound input by the user, analyzes the spectrum of the sound signal input by the user, inputs the spectrum into the acoustic model, recognizes the corresponding word content, and according to the order of the word content input by the user, translates the word content into the text specified by the user, and defines the translated text as the preliminary translated text;

[0007] The translation optimization module analyzes the hidden state sequence rules of the language specified by the user for translation, sorts the word content recognized by the speech recognition and preliminary translation module according to the hidden state sequence rules, and defines the text obtained by optimizing the translation of the word content in the speech recognition and preliminary translation module as the optimized translated text;

[0008] The translation quality assessment module calculates the similarity between the optimized translation text and the preliminary translation text, analyzes the industry to which the translation text belongs based on the preliminary translation text in the speech recognition and preliminary translation module, and sets a similarity threshold according to the industry standards.

[0009] If the similarity < similarity threshold, the optimized translation text in the translation optimization module is fed back to the user.

[0010] If the similarity > similarity threshold, the preliminary translation text in the speech recognition and preliminary translation module is fed back to the user, and the hidden state sequence rules in the translation optimization module are optimized according to the preliminary translation text in the speech recognition and preliminary translation module.

[0011] As a further improvement of this technical solution, the web crawler in the speech recognition and preliminary translation module accesses the required target according to preset rules and goals. After the crawler accesses the target web page, it uses web parsing technology to extract the text content. The known different language spectra and corresponding word contents in the speech recognition and preliminary translation module are obtained from official multilingual files through web crawler technology.

[0012] As a further improvement of this technical solution, the speech recognition and preliminary translation module converts the voice signal input by the user into a digital signal, divides the digital signal into the same frames, calculates the energy of each frame signal and the zero-crossing rate of each frame signal, and converts each frame of voice signal from the time domain to the frequency domain to obtain its spectrum.

[0013] As a further improvement of this technical solution, the translation optimization module analyzes the hidden state sequence rules through the analysis of a large corpus of annotated grammatical components. The files in the corpus of annotated grammatical components are obtained through the web crawler in the speech recognition and preliminary translation module.

[0014] As a further improvement of this technical solution, the steps for the translation optimization module to analyze the hidden state sequence rules are as follows:

[0015] Step 1: Statistically count the number of transfer frequency times between different grammatical components in the corpus.

[0016] Step 2: For each grammatical state, statistically count the frequency of occurrence of different parts of speech.

[0017] Step 3: Construct an emission probability matrix based on the statistical results of the transfer frequency between different grammatical components, where the rows represent the grammatical states, the columns represent the parts of speech, and the elements in the matrix represent the probability of emitting a specific part of speech in a certain grammatical state. Based on the obtained state transition probability matrix and emission probability matrix, the hidden state sequence rules are obtained.

[0018] As a further improvement of this technical solution, the working principle of the translation optimization module for calculating the similarity between the optimized translation text and the preliminary translation text is as follows:

[0019] Semantic role annotation is performed on each word in the translation text according to the hidden state sequence rule. The hidden state sequence rule contains the internal logic and structural characteristics of the translated word when expressing semantics, and plays a key guiding role in semantic role annotation

[0020] When processing the translation text, based on the analysis and annotation of the hidden state sequence rule, attach the corresponding semantic role label to each word in the translation text, construct the semantic role set of each translation text, and calculate the similarity by comparing the semantic role sets

[0021] As a further improvement of this technical solution, the translation quality assessment module analyzes the industry where the translation text is located according to the preliminary translation text in the speech recognition and preliminary translation module, obtains the professional vocabulary library of each industry through the web crawler in the speech recognition and preliminary translation module, and matches the vocabulary in the preliminary translation text with the vocabulary in the professional vocabulary library;

[0022] After all the vocabulary matching is completed, count the number of matching times of each professional vocabulary library. The industry corresponding to the professional vocabulary library with the most matching times is the industry where the preliminary translation text is located

[0023] As a further improvement of this technical solution, when the similarity in the translation quality assessment module > the similarity threshold, the translation optimization module in the industry where the translation text is located is called to collect the corpus. The corpus of the industry analyzes the hidden state sequence rule in the translation optimization module again. During the analysis process, the files in the corpus that are not in the industry where the translation text is located are removed, and the hidden state sequence rule is analyzed again by calling the corpus corresponding to the industry

[0024] Compared with the prior art, the beneficial effects of the present invention are:

[0025] In this AI translation system for realizing multi - language mutual translation, the voice recognition and preliminary translation module collects the voice input by the user and translates the voice input by the user into the translation text specified by the user. In this process, the text translation order of the voice input by the user is the same as the order of the voice input by the user, which is defined as the preliminary translation text. It respects and retains the personalized expression habits of users in different regions in the daily conversation scenario, restores its original semantic and word order characteristics, ensures that the translation text can fit the natural and casual style in the daily communication scenario, and is convenient for both communication parties to understand and communicate;

[0026] However, considering the different language rules when users input voices, the translation optimization module analyzes the hidden state sequence rules of the translation language, reorders the word content in the voices input by users, and defines it as optimized translated text. By deeply analyzing the hidden state sequence rules of the translation language, especially in combination with the industry characteristics of each professional field, precise reordering of word content, term proofreading, and sentence structure optimization are performed on the preliminary translated text, so that the translated text strictly follows the accuracy, logic, and standardization required by professional documents.

[0027] Then, the translation quality assessment module analyzes the industry to which the text to be translated belongs, sets a similarity threshold, and distinguishes whether the output translated text is preliminary translated text or optimized translated text according to the similarity threshold, further ensuring the adaptability of translation quality to different usage scenarios. By analyzing the industry to which the text to be translated belongs and reasonably setting the similarity threshold, it can flexibly distinguish whether the preliminary translated text or the optimized translated text meets the requirements of the corresponding scenario, making up for the deficiencies of the existing technology in dealing with diverse translation needs, so that the present invention can not only meet the needs of natural and smooth communication in daily conversation scenarios, but also improve the quality and practicality of voice translation services in different scenarios. Brief Description of the Drawings

[0028] Figure 1 It is the overall module schematic diagram of the present invention.

[0029] The meanings of the various reference numerals in the figure are as follows:

[0030] 100, voice recognition and preliminary translation module; 200, translation optimization module; 300, translation quality assessment module. Detailed Embodiments

[0031] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0032] An AI translation system for realizing multilingual translation includes a voice recognition and preliminary translation module 100, a translation optimization module 200, and a translation quality assessment module 300.

[0033] The speech recognition and preliminary translation module 100 uses a web crawler to collect a large number of spectra corresponding to known word contents in different languages and the corresponding words, constructs an acoustic model of word contents in different languages and the corresponding spectra, collects the sound input by the user, analyzes the spectrum of the sound signal input by the user, inputs the spectrum into the acoustic model to identify the corresponding word content, and according to the order of the word content input by the user, translates the word content into the text specified by the user, and defines the translated text as the preliminary translated text;

[0034] Specifically, in the speech recognition and preliminary translation module 100, the web crawler accesses the required target according to the preset rules and objectives. After the crawler accesses the target web page, it will use web parsing technologies such as parsing methods based on markup languages such as HTML and XML to extract text content. In the speech recognition and preliminary translation module 100, the known spectra in different languages and the corresponding word contents are obtained through web crawler technology from official multilingual documents such as United Nations documents, multilingual regulations of various governments, translation versions of literary works such as different language translations of world famous works, translations of documents in professional fields such as scientific papers and medical literature, and multilingual web page content on the Internet, etc.;

[0035] The spectra in different languages and the corresponding word contents are as follows: For example, the text corresponding to a piece of speech is "hello", and the spectrum corresponding to each speech frame during the word pronunciation process is known.

[0036] Furthermore, the speech recognition and preliminary translation module 100 converts the sound signal input by the user into a digital signal, divides the digital signal into the same frames, calculates the energy of each frame signal and the zero-crossing rate of each frame signal, and converts each frame of speech signal from the time domain to the frequency domain to obtain its spectrum;

[0037] The speech recognition and preliminary translation module 100 collects the sound input by the user through an audio input device such as a microphone. The audio input device such as a microphone converts the input sound signal into an electrical signal;

[0038] The electrical signal is converted into a digital signal through analog-to-digital conversion and defined as an audio signal. The audio signal is divided into frames of 20 - 30 ms, and there is a 50% overlap between frames. On the premise of ensuring that the speech information is not lost, the speech information is converted into the corresponding spectrum;

[0039] Calculating the energy of each frame signal reflects the strength of the speech signal at a certain moment: where x(n,m) represents the m-th sampling point of the n-th frame, and N is the frame length;

[0040] Calculating the zero-crossing rate of each frame signal, that is, the number of times the signal crosses the zero value from positive to negative or from negative to positive, reflects the frequency change of the speech signal: where sgn is the sign function;

[0041] Each frame of speech signal is transformed from the time domain to the frequency domain through the FFT algorithm to obtain its spectrum. The FFT algorithm quickly calculates the discrete Fourier transform, converts the signal in the time domain into a representation in the frequency domain, analyzes the energy distribution of the speech signal at different frequencies, and reflects the energy magnitude of the speech signal at each frequency component.

[0042] The translation optimization module 200 analyzes the hidden state sequence rules of the user-specified translation language, sorts the word content recognized by the speech recognition and preliminary translation module 100 according to the hidden state sequence rules, and defines the text obtained by optimizing the translation of the word content in the speech recognition and preliminary translation module 100 as the optimized translation text.

[0043] The translation optimization module 200 analyzes the hidden state sequence rules through the analysis of a large number of corpora with annotated grammatical components. The files in the corpus with annotated grammatical components are obtained through the web crawler in the speech recognition and preliminary translation module 100. The files in the corpus include a large number of official multilingual documents such as United Nations documents, multilingual regulations of various governments, etc., translated versions of literary works such as different language translations of world famous works, translations of documents in professional fields such as scientific papers and medical literature, and multilingual web page content on the Internet.

[0044] To avoid incorrect order of the user input sound in the speech recognition and preliminary translation module 100, the steps for the translation optimization module 200 to analyze the hidden state sequence rules are as follows:

[0045] Step 1: Statistically calculate the transition frequencies between different grammatical components in the corpus, calculate the number of times of transferring from the subject state to the predicate state, the number of times of transferring from the predicate state to the object state, etc. The calculation example is as follows:

[0046] The number of times of transferring from the subject state to the predicate state is where n sv is the number of times of transferring from the subject state to the predicate state, and n s is the total number of times the subject appears. In this way, all possible grammatical component transfer situations are statistically calculated to obtain a state transition probability matrix.

[0047] Step 2: For each grammatical state, statistically calculate the frequencies of different parts of speech. In the subject state, statistically calculate the number of times of parts of speech such as nouns, pronouns, and noun phrases appear. The calculation example is as follows:

[0048] The probability of emitting a noun part of speech in the subject state where n s-N is the number of times a noun appears in the subject state, and n svThe total number of occurrences of all words in the subject state. Similarly, the probabilities of verbs in the predicate state, nouns or pronouns in the object state, etc. are statistically analyzed;

[0049] Step 3: Construct an emission probability matrix based on the results of statistically analyzing the transition frequencies between different grammatical components. Here, the rows represent grammatical states, the columns represent part-of-speech tags, and the elements in the matrix represent the probabilities of emitting specific part-of-speech tags in a certain grammatical state. Based on the statistically obtained state transition probability matrix and emission probability matrix, the hidden state sequence rules are derived.

[0050] The purpose of optimizing the translated text and the preliminary translated text is to convert the voice input by the user in the speech recognition and preliminary translation module 100 into the core meaning that the user wants to convey. When calculating the similarity between the optimized translated text and the preliminary translated text, it is judged as a whole whether they are consistent in conveying the core meaning. Even if there may be some differences in vocabulary, grammatical structure, etc., as long as they can equivalently express the key content of the original text, it means they are similar to a certain extent;

[0051] The working principle of the further translation optimization module 200 for calculating the similarity between the optimized translated text and the preliminary translated text is as follows:

[0052] Perform semantic role annotation for each word in the translated text according to the hidden state sequence rules. The content of semantic role annotation includes information such as subject, predicate, and object;

[0053] The hidden state sequence rules contain the internal logic and structural characteristics of translated words when expressing semantics, and play a key guiding role in semantic role annotation. In some languages, specific orders of sentence structures such as verbs in the second position and subjects in the first position, as well as word form changes, are closely related to semantic roles. These rules are abstracted as hidden state sequence rules;

[0054] When processing the translated text, based on the analysis and annotation of the hidden state sequence rules, ultimately, each word in the translated text will be attached with a corresponding semantic role label, forming a structured annotation result to determine its corresponding semantic role

[0055] After semantic role annotation, construct the semantic role sets of the translated texts respectively, and calculate the similarity by comparing the semantic role sets; The core idea is that if the distribution of semantic roles assumed by words in two texts is relatively similar, even if the word order is different, the core meanings they express are likely to be similar. The calculation formula is as follows:

[0056] The semantic role set obtained after semantic role annotation of the optimized translated text is R = {r1, r2, …, r n}, the semantic role set obtained after semantic role annotation of the preliminary translation text is S = {s1, s2, …, s n};

[0057] Calculate the intersection I = R ∩ S and union U = R ∪ S of the two semantic role sets in sequence, and measure the similarity of the semantic role sets by calculating the Jaccard similarity:

[0058] similarity SRL The value range is [0, 1]. The closer the value is to 1, the more similar the optimized translation text and the preliminary translation text are in terms of semantic role distribution.

[0059] The translation quality assessment module 300 calculates the similarity between the optimized translation text and the preliminary translation text, analyzes the industry to which the translation text belongs based on the preliminary translation text in the speech recognition and preliminary translation module 100, and sets a similarity threshold according to the industry standard;

[0060] Furthermore, the translation quality assessment module 300 analyzes the industry to which the translation text belongs based on the preliminary translation text in the speech recognition and preliminary translation module 100, obtains the professional vocabulary libraries of each industry through the web crawler in the speech recognition and preliminary translation module 100, and matches the vocabulary in the preliminary translation text with the vocabulary in the professional vocabulary libraries;

[0061] The vocabulary set of the preliminary translation text is W = {w1, w2, …, w n}, and m professional vocabulary libraries of different industries are obtained through the web crawler. Each vocabulary library V i = {i = 1, 2, …, m}, which are respectively denoted as k i indicating the number of vocabulary in the i-th vocabulary library;

[0062] Define a matching function f(w, V) to determine whether the vocabulary w matches the vocabulary in the vocabulary library V. If it matches, return 1; otherwise, return 0;

[0063] For each professional vocabulary library V i , its matching count C i The calculation formula is:

[0064] After completing the matching of all vocabulary, count the matching times of each professional vocabulary library. The industry corresponding to the professional vocabulary library with the most matching times is the industry to which the preliminary translation text belongs;

[0065] Determine that the industry to which the user's translation text belongs is I, compare the matching counts C1, C2, …, C m , and find the vocabulary library index argnax corresponding to the maximum matching counti (C i ), then I = the industry name corresponds to the index argnax i (C i 's professional vocabulary library);

[0066] Some industries have high requirements for information integrity. Therefore, it is necessary to set a similarity threshold according to the industry standards. In the scientific research field, the translation of academic papers must completely and accurately convey key contents such as research purposes, methods, results, and conclusions;

[0067] By setting the similarity threshold, it is ensured that the optimized translation text and the preliminary translation text are consistent in these key information points. For the translation of some news industries, due to the timeliness of news and the prominent characteristics of important information, the threshold setting will focus on the similarity of key elements such as the subject, time, place, and event content of the news event to ensure that news information is not lost or distorted;

[0068] The following are some examples of setting similarity thresholds according to their standards in different industries: Medical translation has extremely high requirements for accuracy. Whether it is medical terms, disease diagnosis descriptions, treatment plans, or drug information, etc., there must be no errors, because this is directly related to the health of patients and medical decisions. At the same time, the translation of medical literature, etc., also needs to follow strict logic and standard formats;

[0069] Set the similarity weight of the medical industry to 0.9. Given the strict requirements of the medical industry for the accuracy of terms, the term similarity weight is set to 0.6, which accounts for a relatively large proportion of the total weight of 0.9. Medical terms play a key role in accurately conveying professional knowledge. A wrong translation of a term may lead to serious consequences; the sentence structure weight is set to 0.3. Medical texts often rely on strict and standardized sentence structures to clearly elaborate logical relationships such as disease manifestations, pathological mechanisms, and treatment processes.

[0070] If the similarity in the translation quality assessment module 300 is < the similarity threshold, then the optimized translation text in the translation optimization module 200 is fed back to the user, indicating that the order of the user's input speech is incorrect;

[0071] If the similarity > the similarity threshold, then the preliminary translation text in the speech recognition and preliminary translation module 100 is fed back to the user. Optimize the translation optimization module 200 according to the preliminary translation text in the speech recognition and preliminary translation module 100, which means that the preliminary translation text of the speech recognition and preliminary translation module 100 has a certain quality basis for the order of the user's input speech. Information helpful for improving the rules can be mined from it through the translation quality assessment module 300 to optimize the hidden state sequence rules, so as to further improve the accuracy and quality of subsequent translations;

[0072] Specifically, when the similarity in the translation quality evaluation module 300 is greater than the similarity threshold, the corpus collected by the translation optimization module 200 for the industry where the translated text is located is retrieved. The hidden state sequence rules in the translation optimization module 200 are analyzed again using the corpus of the corresponding industry. During the analysis process, files in the corpus that do not belong to the industry where the translated text is located are excluded. By retrieving the corpus of the corresponding industry for re-analysis of the hidden state sequence rules, the unique rules and habits of language expression in this industry can be more accurately mined, further optimizing the hidden state sequence rules in the translation optimization module 200, improving the translation accuracy and professionalism of the translation system for texts in this industry. Excluding files that do not belong to the industry where the translated text is located can avoid interference from language patterns of other industries and make the analysis results more in line with the language characteristics of the target industry.

[0073] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art of this industry should understand that the present invention is not limited by the above embodiments. The above embodiments and the descriptions in the specification are only preferred examples of the present invention and are not used to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and all these changes and improvements fall within the scope of the present invention claimed. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. An AI translation system for multi-language translation, characterized in that: It includes a speech recognition and preliminary translation module (100), a translation optimization module (200) and a translation quality assessment module (300); The speech recognition and preliminary translation module (100) uses a web crawler to collect a large number of frequency spectra and words corresponding to known word contents in different languages, constructs an acoustic model of word contents in different languages ​​and corresponding frequency spectra, collects the sound input by the user, analyzes the frequency spectra of the sound signal input by the user, inputs the frequency spectra into the acoustic model, recognizes the corresponding word content, and translates the word content into text specified by the user according to the order of the word content input by the user, and defines the translated text as preliminary translated text; The translation optimization module (200) analyzes the hidden state sequence rule of the user-specified translation language, sorts the word contents recognized by the speech recognition and preliminary translation module (100) according to the hidden state sequence rule, and the sorted result is a text that optimizes the translation of the word contents in the speech recognition and preliminary translation module (100), which is defined as the optimized translation text; The translation quality assessment module (300) calculates the similarity between the optimized translation text and the preliminary translation text, analyzes the industry to which the translation text belongs based on the preliminary translation text in the speech recognition and preliminary translation module (100), and sets a similarity threshold based on the industry standard; If the similarity is less than the similarity threshold, the optimized translation text in the translation optimization module (200) is fed back to the user; If the similarity is greater than the similarity threshold, the preliminary translation text in the speech recognition and preliminary translation module (100) is fed back to the user, and the hidden state sequence rule in the translation optimization module (200) is optimized according to the preliminary translation text in the speech recognition and preliminary translation module (100).

2. The AI ​​translation system for realizing multi-language translation according to claim 1, characterized in that: The web crawler in the speech recognition and preliminary translation module (100) accesses the desired target according to preset rules and targets. After accessing the target web page, the crawler uses web page parsing technology to extract text content. The speech recognition and preliminary translation module (100) has known different language spectra and corresponding word contents, which are obtained from official multi-language documents through web crawler technology.

3. The AI ​​translation system for realizing multi-language translation according to claim 2, characterized in that: The speech recognition and preliminary translation module (100) converts the sound signal input by the user into a digital signal, divides the digital signal into the same frames, calculates the energy of each frame signal and the zero-crossing rate of each frame signal, and converts each frame of speech signal from the time domain to the frequency domain to obtain its spectrum.

4. The AI ​​translation system for realizing multi-language translation according to claim 3, characterized in that: The translation optimization module (200) analyzes the hidden state sequence rules by collecting a large number of corpora with annotated grammatical components, and the files in the corpora with annotated grammatical components are obtained by the web crawler in the speech recognition and preliminary translation module (100).

5. The AI ​​translation system for realizing multi-language translation according to claim 4, characterized in that: The steps of analyzing the hidden state sequence rule by the translation optimization module (200) are as follows: Step 1: Count the frequency of transitions between different grammatical components in the corpus; Step 2: For each grammatical state, count the frequency of occurrence of different parts of speech; Step 3: Construct a transmission probability matrix based on the statistical results of the transition frequencies between different grammatical components, where the rows represent grammatical states, the columns represent parts of speech, and the elements in the matrix represent the probability of emitting a specific part of speech under a certain grammatical state. Based on the statistically obtained state transition probability matrix and transmission probability matrix, the hidden state sequence rules are derived.

6. The AI ​​translation system for realizing multi-language translation according to claim 5, characterized in that: The working principle of the translation optimization module (200) for calculating the similarity between the optimized translation text and the preliminary translation text is as follows: Semantic role labeling is performed for each word in the translated text according to the hidden state sequence rule. The hidden state sequence rule contains the inherent logic and structural characteristics of the translated word when expressing semantics, and plays a key guiding role in semantic role labeling. When processing the translated text, based on the analysis and annotation of the hidden state sequence rules, a corresponding semantic role label is attached to each word in the translated text, and a semantic role set for each translated text is constructed. The similarity is calculated by comparing the semantic role sets.

7. The AI ​​translation system for realizing multi-language translation according to claim 3, characterized in that: The translation quality assessment module (300) analyzes the industry to which the translated text belongs based on the preliminary translated text in the speech recognition and preliminary translation module (100), obtains professional vocabulary libraries of various industries through the web crawler in the speech recognition and preliminary translation module (100), and matches the vocabulary in the preliminary translated text with the vocabulary in the professional vocabulary library; After all the words are matched, the matching times of each professional vocabulary database are counted, and the industry corresponding to the professional vocabulary database with the largest number of matches is the industry of the preliminary translation text.

8. The AI ​​translation system for realizing multi-language translation according to claim 7, characterized in that: When the similarity in the translation quality assessment module (300) is greater than the similarity threshold, the translation optimization module (200) is called out to collect the corpus of the industry in which the translated text is located, and the corpus of the industry in which the translated text is located is analyzed again with the hidden state sequence rules in the translation optimization module (200). During the analysis, files in the corpus that are not in the industry in which the translated text is located are removed, and the hidden state sequence rules are analyzed again by calling out the corpus of the corresponding industry.

Images