Conversational AI multi-intention understanding and decision-making method for tourism scene

Abstract

The invention is suitable for the technical field of semantic understanding, and provides a dialogue type AI multi-intention understanding and decision-making method for a tourism scene, and the method comprises the steps: obtaining a tourism scene dialogue text which is input by a user and is expressed by a dialect; dialect expression is converted into standard Chinese statements through a dialect translation engine, and translation semantic deviation is corrected through a semantic correction model; performing multi-intention modeling on the converted standard Chinese statements, adopting a pre-training language model to combine with dynamic word segmentation processing, aligning special words of dialects to a standard Chinese word list through a dialect word mapping table, and adjusting semantic representation to adapt to dialect features; generating word list probability distribution, and generating a dialect compatible reply containing a multi-intention decision result in combination with an autoregression model; a dialogue state tracking model is trained according to geographic information, user historical interaction data and a cross-dialect corpus, intention switching and composite requests in multiple rounds of dialogues are analyzed, and dialect compatibility and intention recognition accuracy are effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of semantic understanding technology, and in particular to a conversational AI multi-intention understanding and decision-making method for tourism scenarios. Background Art

[0002] In recent years, conversational AI has been widely applied in tourism scenarios, such as intelligent customer service, itinerary planning, and personalized recommendations. Existing technologies primarily rely on pre-trained language models for intent recognition and text generation, achieving basic conversational interaction through end-to-end training. Common approaches for multi-intent understanding include rule-based intent classification, joint modeling within a multi-task learning framework, and contextual reasoning combined with knowledge graphs. Furthermore, some research has attempted to improve model robustness to diverse inputs through data augmentation techniques. However, traditional translation engines are prone to semantic bias when converting between different dialects, and there is a lack of dialect corpora tailored to tourism scenarios. Given the diverse regional cultural characteristics and long-tail demands involved in tourism, existing training data is insufficient, particularly the lack of high-quality corpora that align dialects with the standard language, which limits the model's generalization capabilities. This severely impacts the practicality and user experience of conversational AI in complex tourism scenarios.

[0003] In view of this, a conversational AI multi-intent understanding and decision-making method for tourism scenarios is needed. Summary of the Invention

[0004] This application provides a conversational AI multi-intention understanding and decision-making method for tourism scenarios, which is used to solve the problems of poor practicality and user experience of conversational AI in complex tourism scenarios.

[0005] This application provides a conversational AI multi-intention understanding and decision-making method for tourism scenarios, including:

[0006] Acquire a travel scene dialogue text input by a user, wherein the dialogue text includes regional dialect expressions;

[0007] Converting the dialect expressions in the dialogue text into standard Chinese sentences using a dialect translation engine, wherein the dialect translation engine is trained based on cross-dialect and standard Chinese training data generated by bidirectional back-translation, and correcting the translation semantic deviations in the converted standard Chinese sentences using a semantic correction model;

[0008] Multi-intent modeling is performed on the converted standard Chinese sentences. A pre-trained language model is combined with dynamic word segmentation processing. The dialect-specific vocabulary is aligned to the standard Chinese vocabulary through a dialect vocabulary mapping table, and the semantic representation is adjusted to adapt to the dialect characteristics.

[0009] Generate vocabulary probability distribution based on multi-intent representation, combine with autoregressive model to generate dialect-compatible responses containing multi-intent decision results, and optimize the alignment between generated results and user intent through cross-entropy loss;

[0010] A conversation state tracking model is trained based on geographic information, historical user interaction data, and cross-dialect data generated by aligning original dialect sentences, standard translations, and back-translated sentences to analyze intent switching and complex requests in multi-round conversations.

[0011] Furthermore, the dialect translation engine is trained based on cross-dialect and standard Chinese training data generated by bidirectional back-translation, including:

[0012] Obtaining original dialect sentences from the original dialect corpus;

[0013] Utilizing a multilingual translation engine to translate the dialect sentences into standard Chinese sentences;

[0014] Using the dialect translation engine to reversely translate the standard Chinese sentence to generate a back-translated dialect sentence containing dialect features;

[0015] The original dialect sentence, the standard Chinese sentence and the back-translated dialect sentence are aligned to generate cross-dialect and standard Chinese training sentence pairs.

[0016] Furthermore, the converting of the dialect expressions in the dialogue text into standard Chinese sentences by the dialect translation engine includes:

[0017] Perform word segmentation on the input dialect sentence and match it with the standard Chinese word list through the dialect vocabulary mapping table;

[0018] Use a context encoder to semantically encode the sequence after word segmentation to generate an intermediate semantic representation;

[0019] The attention-based decoder converts the intermediate semantic representation into standard Chinese sentences.

[0020] Furthermore, the semantic correction model is used to correct the translation semantic deviation in the converted standard Chinese sentence, including:

[0021] Extracting semantic features of the standard Chinese sentence and the back-translated dialect sentence, and calculating the semantic difference loss between the standard Chinese sentence and the back-translated dialect sentence;

[0022] Optimizing the dialect translation engine through adversarial training to minimize the semantic difference loss;

[0023] Semantic consistency constraints are introduced during the translation process to generate semantically consistent standard Chinese sentences and original dialect expressions.

[0024] Furthermore, the specific steps of performing multi-intent modeling on the converted standard Chinese sentence include:

[0025] The encoding layer of the pre-trained language model is used to extract contextual semantic features of the text, and dialect-adapted entity information is extracted from the text. The entity information includes the names of folk activities, regional service types, and dialect-specific expressions.

[0026] identifying user intent categories by hierarchically modeling the contextual semantic features, the user intent categories including service requests, time constraints, and geographic locations;

[0027] The user intent category is fused with the entity information to generate a multi-intent joint representation.

[0028] Furthermore, the dynamic word segmentation processing includes:

[0029] Constructing a dialect-standard vocabulary mapping table, which is a mapping relationship between dialect vocabulary and standard vocabulary generated by aligning the original dialect corpus with the standard Chinese corpus;

[0030] Pre-matching the dialect vocabulary in the mapping table when segmenting the dialogue text, and replacing the dialect vocabulary with standard vocabulary;

[0031] For unmatched words, the default word segmentation strategy of the pre-trained language model is used to preserve contextual coherence.

[0032] Furthermore, the method of aligning dialect-specific words to the standard Chinese vocabulary through the dialect vocabulary mapping table and adjusting the semantic representation to adapt to the dialect characteristics includes:

[0033] Add a dialect adaptation network after the word vector encoding layer of the pre-trained language model, and the adaptation network adjusts the word vector space through linear transformation;

[0034] Use contrastive learning loss function to constrain the consistency between the adapted semantic representation and the standard Chinese semantic representation;

[0035] Optimize the process parameters of intent classification, entity extraction and dialect adaptation networks based on multi-task joint training.

[0036] Furthermore, the method generates a vocabulary probability distribution based on multi-intent representation, combines an autoregressive model to generate a dialect-compatible response containing multi-intent decision results, and optimizes the alignment of the generated results with the user intent through cross-entropy loss, including:

[0037] The multi-intent joint representation is mapped to the vocabulary space through a preset mapping matrix to generate an initial probability distribution;

[0038] Introducing an intent weight distribution mechanism to adjust the vocabulary weights for different intents based on intent classification results;

[0039] The weighted probability distribution is smoothed by a normalization function to generate the target vocabulary probability distribution.

[0040] Furthermore, the method of generating a vocabulary probability distribution based on multi-intent representation, combining an autoregressive model to generate a dialect-compatible response containing multi-intent decision results, and optimizing the alignment of the generated results with the user intent through cross-entropy loss also includes:

[0041] A dynamic masking mechanism is used in the decoding stage to block candidate words that are not relevant to the current multi-intent decision;

[0042] generating a coherent response including dialect features and multi-intent results based on the output semantic representation of the dialect adaptation layer;

[0043] The semantic difference between the generated text and the true label is calculated through the cross entropy loss function, and the parameters of the mapping matrix and autoregressive model are optimized through backpropagation.

[0044] Furthermore, the conversation state tracking model is trained based on geographic information, historical user interaction data, and cross-dialect data to analyze intent switching and complex requests in multiple rounds of conversations. The cross-dialect data is generated by aligning original dialect sentences, standard translations, and back-translated sentences, including:

[0045] Construct a cross-dialect corpus by aligning original dialect sentences, standard Chinese translations, and back-translated dialect sentences to generate mixed training data containing dialect features and standard semantics;

[0046] Constructing a multi-round dialogue sequence using the collected geographic information data and historical user interaction data, where the geographic information data includes the location of scenic spots, transportation routes, and regional service distribution;

[0047] Determining the dependency of intent switching by performing context-aware encoding on the multi-turn dialogue sequence;

[0048] Based on the compound request parsing algorithm, the current conversation intent is associated with the historical conversation state, and the semantic context is updated to parse the compound requests in multi-round conversations.

[0049] It can be seen from the above technical solutions that the present invention has the following advantages:

[0050] The present invention converts user-entered regional dialect expressions into standard Chinese sentences. Based on cross-dialect and standard Chinese training data generated through bidirectional back-translation, the mapping accuracy of dialect-specific vocabulary is improved, reducing semantic loss. A semantic correction model dynamically corrects translation results to avoid misunderstandings of intent due to dialect polysemy. The contextual understanding capabilities of pre-trained language models and a dynamic word segmentation mechanism are combined to effectively distinguish complex intents, avoiding the limitations of single-intent decision-making. An autoregressive model generates responses based on the probability distribution of the vocabulary and outputs multi-intent decision results. Cross-entropy loss optimization is used to generate text that conforms to standard Chinese grammar and is compatible with dialect expression habits. Based on geographic information and user historical data, intent switching is tracked in real time to avoid logical disconnection in the conversation. A state tracking model trained on cross-dialect data associates multi-intent decisions with historical conversation states, enabling coherent processing of complex requests. Through precise dialect conversion, multi-intent joint modeling, dynamic semantic adaptation, and context-aware conversation management, the present invention achieves efficient multi-intent understanding and decision-making for conversational AI in regional tourism scenarios where dialects are frequently used, significantly improving dialect compatibility, intent recognition accuracy, and multi-round interaction coherence. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 This is a flow chart of an embodiment of a conversational AI multi-intent understanding and decision-making method for tourism scenarios in the present invention. DETAILED DESCRIPTION

[0052] The terms "first," "second," "third," "fourth," and the like (if any) in the specification and claims of the present invention and the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a particular order or sequential sequence. It should be understood that the numbers used in this manner are interchangeable where appropriate so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "corresponding to," and any variations thereof, are intended to cover non-exclusive inclusions, e.g., a process, method, system, product, or apparatus comprising a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to these processes, methods, products, or apparatus.

[0053] Example 1

[0054] The implementation method of the present invention can be implemented in the system, can be implemented in the server, and can also be implemented in the terminal, without specific limitation. The following will introduce the conversational AI multi-intention understanding and decision-making method for tourism scenarios in this application from the perspective of system implementation. Figure 1 , the method provided by the present invention comprises the following steps:

[0055] S11. Obtain the travel scenario dialogue text input by the user, where the dialogue text includes regional dialect expressions;

[0056] The system directly receives the text input by the user through the provided API interface or the front-end input box. An integrated Automatic Speech Recognition (ASR) module is used to convert speech into text. The pre-trained speech model adopted here supports dialect speech recognition, and the recognition accuracy is improved by loading travel scenario speech data. When a dialect is recognized, a text classification model based on Transformer, pre-trained on a multilingual corpus, is used to output the probability of the dialect category of the input text. For example, when the original input text is "酸嘢好食" (suan ye hao shi), the output dialect type label is "Guangxi Dialect". A dialect feature vocabulary is constructed, such as "酸嘢" (suan ye) → "pickled food", "落雨" (luo yu) → "rain", and dialect vocabulary is quickly detected through regular expressions or Trie trees.

[0057] Then, irrelevant symbols are removed, the encoding format is unified, and a Named Entity Recognition (NER) model is used to mark the dialect-specific vocabulary in the text. The original order of the dialogue and the multi-turn interaction information are retained and stored in a structured JSON format. Among them, the user identity information is desensitized, and only the necessary semantic content is retained. For example, when the user inputs "礼拜五帮订漓江边嘅房,再推荐啲酸嘢同早茶" (On Friday, help book a room by the Lijiang River, and then recommend some pickled food and morning tea), first the speech is converted into text, that is, the ASR module outputs "礼拜五帮订漓江边嘅房,再推荐啲酸嘢同早茶" (On Friday, help book a room by the Lijiang River, and then recommend some pickled food and morning tea). The model recognizes it as "Guangxi Dialect" with a probability of 0.92; mark "嘅" (ge) → "的" (de), "酸嘢" (suan ye) → "pickled food", "早茶" (morning tea) → "breakfast", and finally store the data.

[0058] This step uses a multi-modal input interface, a dialect classification model, and data preprocessing technology to obtain and standardize the travel scenario dialogue text containing dialects, laying a foundation for subsequent accurate translation and intention understanding.

[0059] S12. Convert the dialect expressions in the dialogue text into standard Chinese sentences through a dialect translation engine. The dialect translation engine is trained based on the cross-dialect and standard Chinese training data generated by bidirectional backtranslation, and a semantic correction model is used to correct the translation semantic deviation in the converted standard Chinese sentences;

[0060] In this embodiment, the dialect translation engine is trained based on the cross-dialect and standard Chinese training data generated by bidirectional backtranslation, including:

[0061] 1. Obtain the dialect sentences in the original dialect corpus;

[0062] Collect the original dialect corpus from the dialect dialogue records in the travel scenario, regional cultural materials, or user interaction logs. For example, "帮定房间" (help book a room) corresponds to "帮订间房" (help book a room) in Guangxi Cantonese. It is necessary to ensure the diversity and domain relevance of the corpus, covering common expressions in the travel scenario, including accommodation, transportation, scenic spot queries, etc.

[0063] 2. Use a multilingual translation engine to translate dialect sentences into standard Chinese sentences;

[0064] Use a general multilingual translation model to initially convert dialect sentences into standard Chinese, such as translating "帮订间房" to "预订一间房间". The translation engine needs to support low-resource language processing and improve the initial translation quality through a tourism terminology library.

[0065] 3. Use a dialect translation engine to back-translate the standard Chinese sentences, generating back-translated dialect sentences with dialect features;

[0066] Input the standard Chinese sentences generated in the previous step into the dialect translation engine and re-translate them into dialect expressions, such as back-translating "预订一间房间" to "订间房", retaining dialect features but possibly introducing grammar or vocabulary deviations. Optimize the dialect translation engine through iterative training to make the back-translation results conform to dialect habits and be close to the original semantics.

[0067] 4. Align the original dialect sentences, standard Chinese sentences, and back-translated dialect sentences to generate cross-dialect and standard Chinese training sentence pairs.

[0068] Align the original dialect sentences, standard Chinese translations, and back-translated dialect sentences semantically, such as the triple: "帮订间房" - "预订一间房间" - "订间房", and construct training data containing dialect diversity features. Use manual verification or automatic alignment algorithms to ensure semantic consistency and filter out low-quality data.

[0069] In this embodiment, the dialect expressions in the dialogue text are converted into standard Chinese sentences through a dialect translation engine, including:

[0070] 1. Perform word segmentation on the input dialect sentences and match the standard Chinese word list through a dialect vocabulary mapping table;

[0071] Use a rule-based or BERT-based tokenizer to perform fine-grained segmentation on dialect sentences, such as segmenting "酸嘢" into "酸" + "嘢"; construct a dialect-standard word list mapping table, such as "嘢" - "东西", "酸嘢" - "腌制食品", and replace dialect vocabulary with standard vocabulary. The mapping table needs to be updated dynamically to support flexible adaptation of regional cultural terms.

[0072] 2. Use a context encoder to perform semantic encoding on the segmented sequence to generate an intermediate semantic representation;

[0073] Use a Transformer encoding layer or BiLSTM to perform context encoding on the tokenized sequence, such as encoding "book a room" into a vector sequence; capture long-distance dependencies through the self-attention mechanism, such as the association between "Friday" and "hotel". The encoder needs to be pre-trained on mixed dialect-standard data to enhance the semantic understanding ability of dialect expressions.

[0074] 3. The decoder based on the attention mechanism converts the intermediate semantic representation into a standard Chinese sentence.

[0075] Adopt a Transformer decoder, combine the intermediate representation output by the encoder, and gradually generate standard Chinese tokens, such as "book", "a", "room". Dynamically focus on the key information output by the encoder during decoding, such as "Friday" corresponding to the "time" entity. Introduce a coverage mechanism to prevent repeated generation and optimize the output fluency through search.

[0076] The above steps generate high-quality training data through bidirectional backtranslation, solving the data bottleneck of dialect-standard Chinese alignment; through token mapping and attention decoding, accurate conversion from dialect to standard Chinese is achieved, avoiding semantic distortion.

[0077] S13. Perform multi-intent modeling on the converted standard Chinese sentence. Adopt a pre-trained language model combined with dynamic tokenization processing. Align dialect-specific vocabulary to the standard Chinese vocabulary through a dialect vocabulary mapping table, and adjust the semantic representation to adapt to dialect features;

[0078] In this embodiment, a semantic correction model is used to correct the translation semantic deviation in the converted standard Chinese sentence, including:

[0079] 1. Extract the semantic features of the standard Chinese sentence and the backtranslated dialect sentence, and calculate the semantic difference loss between the standard Chinese sentence and the backtranslated dialect sentence;

[0080] Use the pre-trained model BERT to extract semantic vectors (h std and h dial ∈R std , h dial ∈R d ) for the standard Chinese sentence S

[0081]

[0082] For example, the Guangxi dialect "luo yu" is backtranslated as "xia yu", but if the semantic correction is insufficient, it may be translated as "precipitation". By calculating the semantic difference loss between "xia yu" and "precipitation", the model output "xia yu" is optimized.

[0083] 2. Optimize the dialect translation engine through adversarial training to minimize the semantic difference loss;

[0084] Construct a dialect translation engine generator and a semantic consistency discriminator. The generator attempts to minimize the cosine similarity loss, and the discriminator distinguishes between real standard Chinese and generated sentences. The adversarial loss function is:

[0085] L adv = E[logD(S std )] + E[log1 - D(G(S std ))]

[0086] Where: G is the generator, and D is the discriminator.

[0087] 3. Introduce semantic consistency constraints during the translation process to generate semantically consistent standard Chinese sentences and original dialect expressions.

[0088] During the translation process, add a constraint term to force the generated sentence to be semantically aligned with the original dialect:

[0089] L total = L[[ID=​​​​​​​​​​​​​​​​​​​​​The intent classification layer performs multi-label classification on the semantic feature H and outputs an intent probability distribution:

[0096] P(intent|H) = Softmax(W c ·H + b c )

[0097] Where: W c ∈R d×k is the classification weight, and k is the number of intent categories.

[0098] 3. Integrate the user intent category with the entity information to generate a multi-intent joint representation.

[0099] Concatenate the intent probability distribution with the entity embedding vector and generate a joint representation through a fully connected layer:

[0100] h joint = ReLU(W f ·[h intent ; h entity + b f )

[0101] Where: is the semantic vector output for the intent respectively, is the semantic vector output for entity extraction, is the concatenation operation of the intent and entity vectors, is the weight matrix of the fully connected layer, used to map the concatenated vector to the joint representation space, b f ∈R d is the bias term of the fully connected layer, and h joint ∈R d is the final multi-intent joint representation, which integrates intent and entity information.

[0102] In this embodiment, the dynamic word segmentation process includes:

[0103] 1. Construct a mapping table between dialects and standard lexicons. The mapping table generates the mapping relationship between dialect vocabulary and standard vocabulary by aligning the original dialect corpus and the standard Chinese corpus;

[0104] Align the original dialect corpus and the standard corpus to generate a mapping table, such as "啱心水" - "符合喜好"; use the edit distance algorithm or semantic similarity calculation to automatically expand the mapping table. For example, the Guangxi dialect "早茶" is mapped to "早餐", and "行街" is mapped to "逛街".

[0105] 2. When segmenting the dialogue text, pre-match the dialect vocabulary in the mapping table and replace the dialect vocabulary with standard vocabulary;

[0106] 3. For unmatched words, the default word segmentation strategy of the pre-trained language model is used to preserve contextual coherence.

[0107] The tokenizer combines the mapping table and prioritizes replacing dialect words if a match occurs. Otherwise, the default strategy is used. For example, if the input is "I want to eat sour food", the tokenization is "I / want / eat / pickled food".

[0108] For unmatched words, such as the proper noun "Lijiang River," the pre-trained model's word segmentation strategy is used to ensure contextual coherence. For example, for the input "Lijiang River is there any good place to stay?", the word segmentation is "Lijiang River / side / what / good / accommodation."

[0109] In this embodiment, dialect-specific words are aligned to the standard Chinese vocabulary through a dialect vocabulary mapping table, and the semantic representation is adjusted to suit the dialect characteristics, including:

[0110] 1. Add a dialect adaptation network after the word vector encoding layer of the pre-trained language model. The adaptation network adjusts the word vector space through linear transformation;

[0111] 2. Use contrastive learning loss function to constrain the consistency of the adapted semantic representation with the standard Chinese semantic representation;

[0112] 3. Optimize the process parameters of intent classification, entity extraction and dialect adaptation networks based on multi-task joint training.

[0113] Add a linear transformation layer after the BERT word vector layer:

[0114] h adapted =W d ·h original +b d

[0115] Where: h adapted is the adapted word vector, h original is the original word vector, W d is the weight matrix of the dialect adaptation layer, b d is the bias term.

[0116] Construct positive sample pairs (standard Chinese and dialect sentences with the same semantics) and negative sample pairs (irrelevant sentences), and the loss function is:

[0117]

[0118] Where: h j is the semantic vector of each j-th negative sample, sin(h std ,h dial ) is h std and h dial The cosine similarity of , N is the number of negative samples.

[0119] For example, constrain the semantic vectors of "sour food" to be close to "pickled food" and away from "sour food".

[0120] The total loss function is the weighted sum of intent classification, entity extraction and contrast loss:

[0121] L total =αL intent +βL entity +γL cont

[0122] Where: L intent is the cross entropy loss for intent classification, L entity is the CRF loss of entity extraction, L cont For contrastive learning loss, α, β, and γ are the loss weights of each task respectively.

[0123] The following scenario is taken as an example where a user from Guangxi inputs "Please help me book a room by the Li River on Friday, and recommend some pickled foods and breakfast." First, "the room" is corrected to "the room", "the pickled foods" is translated into "pickled food", and "the breakfast" retains its original meaning; the word segmentation is "Please help me book a room by the Li River on Friday, and recommend some pickled foods and breakfast"; the intent is identified as "hotel booking" + "food recommendation", and the entities "by the Li River" (location) and "Friday" (time) are extracted; the response is generated as "I have booked a hotel by the Li River on Friday, and recommend a pickled food store and a nearby breakfast restaurant." Finally, the Guangxi dialect and cultural characteristics are accurately adapted to achieve efficient parsing and response to complex multi-intents.

[0124] S14. Generate a vocabulary probability distribution based on multi-intent representations, combine it with an autoregressive model to generate dialect-compatible responses containing multi-intent decision results, and optimize the alignment of the generated results with the user intent using cross-entropy loss.

[0125] In this embodiment, a vocabulary probability distribution is generated based on multi-intent representation, and an autoregressive model is used to generate dialect-compatible responses containing multi-intent decision results. The alignment of the generated results with the user intent is optimized using cross-entropy loss, including:

[0126] 1. Map the multi-intent joint representation to the vocabulary space through a preset mapping matrix to generate an initial probability distribution;

[0127] 2. Introduce an intent weight distribution mechanism to adjust the vocabulary weights for different intents based on intent classification results;

[0128] 3. Smooth the weighted probability distribution through the normalization function to generate the target vocabulary probability distribution.

[0129] Multiple intent joint representation h joint ∈R d Through the preset mapping matrix W0∈RV×d Projection to vocabulary space:

[0130] P base =W0·h joint

[0131] Where: V is the vocabulary size, W0 is the mapping matrix, P base ∈R V is the initial probability distribution.

[0132] Assign a weight vector based on the intent classification probability:

[0133]

[0134] Where: s i and s j is the confidence of intention i, j, α i is the normalized weight of intent i, and k is the total number of intents.

[0135] Weighted probability distribution:

[0136]

[0137] in: is the initial probability distribution subset corresponding to the i-th intention.

[0138] Introduce the temperature coefficient τ to adjust the distribution smoothness:

[0139]

[0140] Where: τ controls the smoothness of the distribution, τ>1 is more uniform, and τ<1 is more sharp.

[0141] Based on the multi-intent representation, the probability distribution of the vocabulary is generated. The autoregressive model is combined to generate dialect-compatible responses containing multi-intent decision results. The cross-entropy loss is used to optimize the alignment between the generated results and the user intent. The following also applies:

[0142] 1. A dynamic masking mechanism is used in the decoding stage to block candidate words that are not relevant to the current multi-intent decision;

[0143] 2. Generate a coherent response that includes dialect features and multi-intent results based on the output semantic representation of the dialect adaptation layer;

[0144] 3. The semantic difference between the generated text and the true label is calculated through the cross-entropy loss function, and the parameters of the mapping matrix and autoregressive model are optimized through backpropagation.

[0145] According to the current multi-intention decision results, such as "hotel reservation + attraction recommendation", the blocked word table M, such as "shopping" and "air ticket", is occasionally seen. The probability mask of the candidate words is memorized during decoding:

[0146]

[0147] Where: P final (w)∈[0,1], the original probability value of word w in the final vocabulary probability distribution, P masked (w) is the probability of the word after masking, and the probability of irrelevant words is forced to zero.

[0148] The autoregressive decoder combines the semantic vectors output by the dialect adaptation layer to gradually generate word sequences:

[0149] h t =Decoder(h adapted ,h t-1 )

[0150] Where: h adapted is the semantic vector output by the dialect adaptation layer, integrating dialect features, h t-1 is the hidden state generated by the decoder at the previous time step t-1, h t is the hidden state at the current time step t, used to predict the next word.

[0151] When generating, dialect-compatible words are given priority, such as replacing "suanye" with "pickled food".

[0152] Calculate and generate text Y=y1,y2,…,y T Loss with the true label:

[0153]

[0154] Where: y t To generate the tth word of the text, y <t is the first t-1 words generated, X is the semantic representation of the output text, P(y t │y <t ,X)∈[0,1] is the model prediction word y t Back-propagation optimizes the mapping matrix W0, the decoder and the dialect adaptation layer parameters.

[0155] For example, when a user inputs "Book a room by the Li River on Friday, and recommend some sour food and morning tea"; after the system maps multiple intent representations, the probability of words such as "room", "Friday", "sour food", and "morning tea" is higher; the weight of "hotel reservation" is set to 0.6, and the weight of "food recommendation" is set to 0.4, which jointly influence the word list distribution; irrelevant words such as "car rental" and "air ticket" are blocked; the decoder combines the output of the adaptation layer to generate "A hotel by the Li River has been booked on Friday, and a pickled food store and a nearby morning tea restaurant are recommended"; if the generated result omits "morning tea", the cross-entropy loss drives the model to adjust the parameters to complete the information.

[0156] S15. A conversation state tracking model is trained based on geographic information, historical user interaction data, and cross-dialect data to analyze intent switching and complex requests in multi-round conversations. The cross-dialect data is generated by aligning original dialect sentences, standard translations, and back-translated sentences.

[0157] 1. Build a cross-dialect corpus by aligning original dialect sentences, standard Chinese translations, and back-translated dialect sentences to generate mixed training data containing dialect features and standard semantics;

[0158] Dialectal dialogues are collected from Guangxi tourism scenarios, such as "Help me book a room by the Li River." These are then converted to standard Chinese using a dialect translation engine, such as "Book a room by the Li River." The standard Chinese is then back-translated into the dialect, such as "Book a room by the Li River." This preserves dialect characteristics but may introduce grammatical differences. Dynamic Time Warping (DTW) or manual verification is used to align the original dialect, standard translation, and back-translation to ensure semantic consistency. Triples are stored as structured training data, annotated with intent labels and entity information.

[0159] 2. Construct multi-round conversation sequences using collected geographic information data and historical user interaction data. Geographic information data includes the location of scenic spots, transportation routes, and regional service distribution.

[0160] Retrieve the locations of Guangxi attractions, such as "Lijiang Scenic Area" and "Yangshuo West Street," transportation routes, such as "Guilin to Yangshuo Bus," and service distribution, such as "Nearby Hotel Density," from a map API. Build a geographic knowledge graph, linking attractions, transportation, and service entities. Extract historical user interaction records, such as "User Preferred B&B" and "User Searched for Soup Shops," and store them as a time-series conversation sequence. Expand a single-turn conversation into a multi-turn interaction. For example, User 1 recommends a hotel near the Lijiang River; System 1 has already recommended Hotel A. Does System 1 need to check transportation? User 2 does, and also wants to know which sour shops are nearby.

[0161] 3. Determine the dependencies of intent switching by performing context-aware encoding on multi-turn dialogue sequences;

[0162] Use the Transformer encoder to encode the multi-round dialogue sequence and generate the context representation H∈R n×d :

[0163] H=TransformerEncoder([h1,h2,…,h n ])

[0164] Where: h i ,i∈(1,2,…n) is the semantic vector of the i-th round of dialogue.

[0165] The self-attention mechanism is used to analyze the relevance of intentions in adjacent rounds. For example, user 1’s intention is “hotel recommendation” → user 2’s intention is “transportation query”. The model captures the dependency between “hotel location” and “transportation route”.

[0166] Calculate the intention switching probability:

[0167] P(switch|H)=Sigmoid(W s H+b s )

[0168] Where: W s is the classification weight.

[0169] 4. Based on the compound request parsing algorithm, the current conversation intent is associated with the historical conversation state, and the semantic context is updated to parse compound requests in multiple rounds of conversation.

[0170] A dialogue state graph is constructed based on a graph neural network (GNN). Nodes represent intents, such as "hotel reservation" and "food recommendation," and edges represent the weight of intent associations. When a user enters a new request, the algorithm retrieves the historical state (such as the confirmed "check-in on Friday") and updates the current intent node: h current =GNN(h current ,h history ), generates a response based on the updated state diagram, for example: User: I want to change my check-in time to Saturday, and then recommend dinner; System: The time has been changed to Saturday, and recommends a restaurant by the Li River (associated with the historical "Li River" location).

[0171] The above-mentioned embodiments achieve efficient multi-intent understanding and decision-making of conversational AI in tourism scenarios in areas where dialects are frequently used through precise dialect conversion, multi-intent joint modeling, dynamic semantic adaptation and context-aware dialogue management, significantly improving dialect compatibility, intent recognition accuracy and multi-round interaction coherence.

[0172] Those skilled in the art will appreciate that the units of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition of each example has been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0173] In the embodiments provided herein, it should be understood that the division of units is merely a logical functional division. In actual implementation, other division methods may be employed, such as combining multiple units into one unit, splitting a unit into multiple units, or ignoring certain features. Furthermore, the functional units in the various embodiments of the present invention may be integrated into a single processing unit, each unit may exist physically as a separate unit, or two or more units may be integrated into a single unit. These integrated units may be implemented in either hardware or software functional units.

[0174] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-0nly Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk, etc., various media that can store program code.

[0175] It can be understood that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some or all of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the scope of the claims and description of the present invention.

Claims

1. A conversational AI multi-intention understanding and decision-making method for tourism scenarios, characterized by: include: Acquire a travel scene dialogue text input by a user, wherein the dialogue text includes regional dialect expressions; Converting the dialect expressions in the dialogue text into standard Chinese sentences using a dialect translation engine, wherein the dialect translation engine is trained based on cross-dialect and standard Chinese training data generated by bidirectional back-translation, and correcting the translation semantic deviations in the converted standard Chinese sentences using a semantic correction model; Multi-intent modeling is performed on the converted standard Chinese sentences. A pre-trained language model is combined with dynamic word segmentation processing. The dialect-specific vocabulary is aligned to the standard Chinese vocabulary through a dialect vocabulary mapping table, and the semantic representation is adjusted to adapt to the dialect characteristics. Generate vocabulary probability distribution based on multi-intent representation, combine with autoregressive model to generate dialect-compatible responses containing multi-intent decision results, and optimize the alignment between generated results and user intent through cross-entropy loss; A conversation state tracking model is trained based on geographic information, historical user interaction data, and cross-dialect data generated by aligning original dialect sentences, standard translations, and back-translated sentences to analyze intent switching and complex requests in multi-round conversations.

2. The conversational AI multi-intention understanding and decision-making method for tourism scenarios according to claim 1 is characterized in that: The dialect translation engine is trained based on cross-dialect and standard Chinese training data generated by bidirectional back-translation, including: Obtaining original dialect sentences from the original dialect corpus; translating the dialect sentences into standard Chinese sentences using a multilingual translation engine; Using the dialect translation engine to reversely translate the standard Chinese sentence to generate a back-translated dialect sentence containing dialect features; The original dialect sentence, the standard Chinese sentence and the back-translated dialect sentence are aligned to generate cross-dialect and standard Chinese training sentence pairs.

3. The conversational AI multi-intention understanding and decision-making method for tourism scenarios according to claim 2 is characterized in that: The converting the dialect expressions in the dialogue text into standard Chinese sentences by the dialect translation engine includes: Perform word segmentation on the input dialect sentence and match it with the standard Chinese word list through the dialect vocabulary mapping table; Use a context encoder to semantically encode the sequence after word segmentation to generate an intermediate semantic representation; The attention-based decoder converts the intermediate semantic representation into standard Chinese sentences.

4. The conversational AI multi-intention understanding and decision-making method for tourism scenarios according to claim 1 is characterized in that: The semantic correction model is used to correct the translation semantic deviation in the converted standard Chinese sentence, including: Extracting semantic features of the standard Chinese sentence and the back-translated dialect sentence, and calculating the semantic difference loss between the standard Chinese sentence and the back-translated dialect sentence; Optimizing the dialect translation engine through adversarial training to minimize the semantic difference loss; Semantic consistency constraints are introduced during the translation process to generate semantically consistent standard Chinese sentences and original dialect expressions.

5. The conversational AI multi-intention understanding and decision-making method for tourism scenarios according to claim 1 is characterized in that: The specific steps of performing multi-intent modeling on the converted standard Chinese sentence include: The encoding layer of the pre-trained language model is used to extract contextual semantic features of the text, and dialect-adapted entity information is extracted from the text. The entity information includes the names of folk activities, regional service types, and dialect-specific expressions. identifying user intent categories by hierarchically modeling the contextual semantic features, the user intent categories including service requests, time constraints, and geographic locations; The user intent category is fused with the entity information to generate a multi-intent joint representation.

6. The conversational AI multi-intention understanding and decision-making method for tourism scenarios according to claim 5 is characterized in that: The dynamic word segmentation processing includes: Constructing a dialect-standard vocabulary mapping table, which is a mapping relationship between dialect vocabulary and standard vocabulary generated by aligning the original dialect corpus with the standard Chinese corpus; Pre-matching the dialect vocabulary in the mapping table when segmenting the dialogue text, and replacing the dialect vocabulary with standard vocabulary; For unmatched words, the default word segmentation strategy of the pre-trained language model is used to preserve contextual coherence.

7. The conversational AI multi-intention understanding and decision-making method for tourism scenarios according to any one of claims 5 or 6, characterized in that: The method of aligning dialect-specific vocabulary to the standard Chinese vocabulary through the dialect vocabulary mapping table and adjusting the semantic representation to adapt to the dialect characteristics includes: Add a dialect adaptation network after the word vector encoding layer of the pre-trained language model, and the adaptation network adjusts the word vector space through linear transformation; Use contrastive learning loss function to constrain the consistency between the adapted semantic representation and the standard Chinese semantic representation; Optimize the process parameters of intent classification, entity extraction and dialect adaptation networks based on multi-task joint training.

8. The conversational AI multi-intention understanding and decision-making method for tourism scenarios according to claim 1 is characterized in that: The method generates a vocabulary probability distribution based on multi-intent representation, combines an autoregressive model to generate dialect-compatible responses containing multi-intent decision results, and optimizes the alignment of the generated results with the user intent through cross-entropy loss, including: The multi-intent joint representation is mapped to the vocabulary space through a preset mapping matrix to generate an initial probability distribution; Introducing an intent weight distribution mechanism to adjust the vocabulary weights for different intents based on intent classification results; The weighted probability distribution is smoothed by a normalization function to generate the target vocabulary probability distribution.

9. The conversational AI multi-intention understanding and decision-making method for tourism scenarios according to claim 8 is characterized in that: The method of generating a vocabulary probability distribution based on multi-intent representation, combining an autoregressive model to generate a dialect-compatible response containing multi-intent decision results, and optimizing the alignment of the generated results with the user intent through cross-entropy loss also includes: A dynamic masking mechanism is used in the decoding stage to block candidate words that are not relevant to the current multi-intent decision; generating a coherent response including dialect features and multi-intent results based on the output semantic representation of the dialect adaptation layer; The semantic difference between the generated text and the true label is calculated through the cross entropy loss function, and the parameters of the mapping matrix and autoregressive model are optimized through backpropagation.

10. The conversational AI multi-intention understanding and decision-making method for tourism scenarios according to claim 1 is characterized in that: The conversation state tracking model is trained based on geographic information, historical user interaction data, and cross-dialect data to analyze intent switching and complex requests in multiple rounds of conversations. The cross-dialect data is generated by aligning original dialect sentences, standard translations, and back-translated sentences, including: Construct a cross-dialect corpus by aligning original dialect sentences, standard Chinese translations, and back-translated dialect sentences to generate mixed training data containing dialect features and standard semantics; Constructing a multi-round dialogue sequence using the collected geographic information data and historical user interaction data, where the geographic information data includes the location of scenic spots, transportation routes, and regional service distribution; Determining the dependency of intent switching by performing context-aware encoding on the multi-turn dialogue sequence; Based on the compound request parsing algorithm, the current conversation intent is associated with the historical conversation state, and the semantic context is updated to parse the compound requests in multi-round conversations.

Images