An incomplete dialogue rewriting method and device and a storage medium
By generating an insertion and replacement operation matrix, the problem of semantic inconsistency in incomplete dialogue rewriting is solved, achieving complete dialogue supplementation and semantic consistency, and improving the understanding effect of multi-turn dialogue.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- 四川启睿克科技有限公司
- Filing Date
- 2022-11-14
- Publication Date
- 2026-05-22
AI Technical Summary
Existing methods for rewriting incomplete dialogues fail to effectively consider the high degree of overlap between incomplete and complete dialogues, resulting in semantically inconsistent generated dialogues or an inability to handle incomplete dialogues that contain both referents and omissions.
By acquiring and preprocessing historical dialogues, an insertion operation matrix and a replacement operation matrix are generated using a rewriting operation judgment model. Insertion and replacement operation information is extracted, and the original incomplete dialogue is rewritten based on this information.
It achieves a high degree of overlap between incomplete and complete dialogues, generates dialogues that are semantically consistent with historical dialogues, and can handle incomplete dialogues with referents and omissions, thus improving the dialogue comprehension effect.
Smart Images
Figure CN115795006B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of natural language processing technology, and in particular to a method, apparatus and storage medium for rewriting incomplete dialogues. Background Technology
[0002] In recent years, with the rapid development of natural language processing and big data technologies, human-computer dialogue systems have gradually become an important research area in the field of artificial intelligence. Various intelligent customer service systems and virtual assistants have been developed and widely applied in e-commerce, social media, and mobile phones. Based on the number of turns of interaction between the human and the dialogue system, dialogue systems can be divided into single-turn dialogue systems and multi-turn dialogue systems. Based on massive historical data and mature natural language processing technology, the response generation methods for single-turn dialogues are already quite mature. However, multi-turn dialogues are more prevalent in actual conversations. In multi-turn dialogues, for the sake of conciseness, fluency, and efficiency, users often input incomplete dialogue statements (references or omissions). For example: "What's the weather like in Chengdu?", "Cloudy.", "Why is it always like this?". The final complete dialogue should be: "Why is it always cloudy in Chengdu?", which shows both reference and omission. The prevalence of omissions and references in multi-turn dialogues greatly interferes with dialogue comprehension. A common solution is to rewrite and complete the incomplete dialogue, and then use single-turn dialogue methods for dialogue comprehension.
[0003] Currently, there are three main solutions for rewriting incomplete dialogues:
[0004] Option 1: Generative Approach. This method first uses an encoder to semantically encode historical dialogues and incomplete dialogues, and then uses a decoder to decode the semantic encoding to directly generate a complete dialogue. However, this method generates a complete dialogue starting from the first word, failing to consider the high degree of overlap between incomplete and complete dialogues, resulting in dialogues that do not semantically match the historical dialogue.
[0005] Option 2: Sequence labeling-based method. One solution using sequence labeling is to first label the start and end positions of key information, the completion position, the start and end positions of reference, and then insert the information fragment between the start and end positions of key information at the completion position, or replace the fragment between the start and end positions of reference with the information fragment between the start and end positions of key information. The problem is that it cannot rewrite incomplete dialogues that contain both references and omissions.
[0006] Option 3: A method combining generation and sequence labeling. This method first performs sequence labeling on the incomplete dialogue, obtaining a label sequence containing three types of tags: delete, retain, and replace. Then, for each label sequence, the sequence corresponding to the delete tag is deleted, and the sequence corresponding to the retain tag is retained. Finally, a generation method is used to generate the sequence corresponding to the replace tag. This type of method suffers from the same problem as generation-based methods: it does not consider the high degree of overlap between incomplete and complete dialogues, leading to the generation of dialogues that do not semantically match the historical dialogue information. Furthermore, errors in sequence labeling will cascade and propagate to the generation operation.
[0007] In summary, existing methods for rewriting incomplete dialogues have several drawbacks. They fail to consider the high degree of overlap between incomplete and complete dialogues, resulting in dialogues that do not semantically match the historical dialogue information, or they cannot rewrite incomplete dialogues that contain both referents and omissions. Summary of the Invention
[0008] This invention provides a method, apparatus, and storage medium for rewriting incomplete dialogues, in order to solve the problems in existing incomplete dialogue rewriting methods that fail to consider the high degree of overlap between incomplete dialogues and complete dialogues, resulting in the generation of dialogues that do not conform to the semantics of historical dialogue information, or that cannot rewrite incomplete dialogues that simultaneously contain referents and omissions.
[0009] The technical solution adopted in this invention is: to provide a method for rewriting incomplete dialogues, including:
[0010] Obtain historical dialogues and preprocess the historical dialogues and incomplete dialogues;
[0011] Input historical dialogues and incomplete dialogues into the rewrite operation judgment model to obtain the insertion operation matrix and the replacement operation matrix;
[0012] Extracting insertion operation information based on the insertion operation matrix;
[0013] Extract replacement operation information based on the replacement operation matrix;
[0014] The original incomplete dialogue is finally rewritten based on the insertion and replacement operation information.
[0015] Furthermore, the method for acquiring historical dialogues and preprocessing the historical dialogues and incomplete dialogues includes:
[0016] Retrieve one or more historical dialogue statements preceding the incomplete dialogue;
[0017] Preprocessing operations are performed on historical dialogues and incomplete dialogues. These preprocessing operations include removing special symbols, converting between half-width and full-width characters, and converting between simplified and traditional Chinese characters.
[0018] Furthermore, the method for inputting historical dialogue and incomplete dialogue content into the rewrite operation judgment model to obtain the insertion operation matrix and the replacement operation matrix includes:
[0019] Connecting historical dialogues and incomplete dialogues;
[0020] The rewrite operation is used to determine the semantic encoding of the concatenated historical dialogue and incomplete dialogue content by the encoder.
[0021] The rewrite operation judgment model decoder is used to decode the semantic encoded information to obtain the insertion operation matrix and the replacement operation matrix.
[0022] Furthermore, the method for extracting insertion operation information based on the insertion operation matrix includes:
[0023] Insertion operation information is extracted using an insertion operation information extraction algorithm, including:
[0024] Traverse each row of the insertion operation matrix corresponding to the incomplete dialogue. If there are two columns with a value of 1 in the traversed row and they correspond to the historical dialogue, extract the historical dialogue fragments corresponding to the two columns and mark the position of the row corresponding to the incomplete dialogue as the insertion position.
[0025] Save the historical dialogue fragment and its corresponding insertion position as a tuple in the insertion operation information.
[0026] Furthermore, the method for extracting replacement operation information based on the replacement operation matrix includes:
[0027] Replacement operation information is extracted using a replacement operation information extraction algorithm, including:
[0028] For each row of the replacement operation matrix corresponding to the incomplete dialogue, if there is a column with a value of 1 in the traversed row and it corresponds to the historical dialogue, record the row number and column number as h1 and c1 respectively.
[0029] Continue traversing each row of the replacement operation matrix from the row below h1. If there is a column with a value of 1 in the traversed row that corresponds to the historical dialogue, record the row number and column number as h2 and c2 respectively.
[0030] Extract fragments h1 to h2 from the incomplete dialogue as fragments to be replaced, and extract fragments c1 to c2 from the historical dialogue as the corresponding replacement fragments.
[0031] Save the segment to be replaced and the corresponding replacement segment as a tuple in the replacement operation information.
[0032] Furthermore, the method for finally rewriting the original incomplete dialogue based on insertion operation information and replacement operation information includes:
[0033] Iterate through each tuple in the insertion operation information and insert the historical dialogue fragment into the insertion position.
[0034] Iterate through each tuple in the replacement operation information and replace the fragment to be replaced in the incomplete dialogue with the replacement fragment.
[0035] The present invention also provides an incomplete dialogue rewriting device, comprising:
[0036] The data processing module is used to acquire historical dialogues and preprocess the historical dialogues and incomplete dialogues.
[0037] The rewrite operation matrix acquisition module inputs historical dialogues and incomplete dialogues into the rewrite operation judgment model to obtain the insertion operation matrix and the replacement operation matrix.
[0038] The insertion operation information extraction module is used to extract insertion operation information using an insertion operation information extraction algorithm.
[0039] The replacement operation information extraction module is used to extract replacement operation information using a replacement operation information extraction algorithm.
[0040] The dialogue modification module is used to rewrite the original incomplete dialogue based on the insertion and replacement operation information.
[0041] The present invention also provides a storage medium storing an incomplete dialogue rewriting program, wherein the incomplete dialogue rewriting program, when executed by a processor, implements the steps of the above-described incomplete dialogue rewriting method.
[0042] The beneficial effects of this invention are: it can achieve a high degree of overlap between incomplete dialogues and dialogues with complete information, generate dialogues that are semantically consistent with historical dialogue information, and rewrite incomplete dialogues that simultaneously contain one or more referents and omissions, thereby supplementing the information in complete dialogues and improving the effect of dialogue comprehension. Attached Figure Description
[0043] Figure 1 This is a flowchart illustrating the incomplete dialogue rewriting method disclosed in this invention.
[0044] Figure 2 This is a structural block diagram of the incomplete dialogue rewriting device disclosed in this invention. Detailed Implementation
[0045] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be described in further detail below with reference to the accompanying drawings, but the embodiments of the present invention are not limited thereto.
[0046] Example 1:
[0047] See Figure 1 This embodiment discloses a method for rewriting incomplete dialogues, including the following steps:
[0048] S1. Obtain historical dialogues and preprocess the historical dialogues and incomplete dialogues.
[0049] Specifically, first, one or more historical dialogue statements preceding the incomplete dialogue are obtained, and then preprocessing operations are performed on the historical dialogue and the incomplete dialogue, including but not limited to removing special symbols, converting between half-width and full-width characters, and converting between simplified and traditional Chinese characters.
[0050] It should be noted that the method of this invention does not require a separate step to determine whether the dialogue is complete. If the dialogue is complete, then the insertion operation matrix I and the replacement operation matrix R are matrices of all zeros. Thus, the insertion operation information and replacement operation information cannot be obtained according to S3 and S4. That is, S5 traverses an empty set and does not perform any operation.
[0051] The dialogue can be text generated from speech recognition or text entered by the user.
[0052] S2. Input the historical dialogue and incomplete dialogue content into the rewrite operation judgment model to obtain the insertion operation matrix and the replacement operation matrix.
[0053] Specifically, S2 includes:
[0054] S21. First, connect the historical dialogues and incomplete dialogues;
[0055] S22. Then, the rewrite operation is used to determine the semantic encoding of the concatenated historical dialogue and incomplete dialogue content by the model encoder.
[0056] S23. Finally, the rewrite operation judgment model decoder is used to decode the semantic encoded information to obtain the insertion operation matrix and the replacement operation matrix.
[0057] By concatenating historical dialogues, the semantic encoding of incomplete dialogues can incorporate historical dialogue information, thereby improving the decoding performance of the decoder in the rewriting operation judgment model. By decoding the semantically encoded information into insertion and substitution operation matrices, the problem of existing technologies being unable to rewrite incomplete dialogues containing both referential and ellipsis terms can be solved. Furthermore, the rewritten content is entirely derived from historical and incomplete dialogues, avoiding the semantic discrepancy between dialogues generated by generative methods and historical dialogue information.
[0058] Specifically, suppose the preprocessed incomplete dialogue is A: "Why is it always like this?", the previous two rounds of historical dialogue are B: "What's the weather like in Chengdu today?", "Cloudy.", and the rewritten complete dialogue is T: "Why is it always cloudy in Chengdu?". First, concatenate B and A to get I: "What's the weather like in Chengdu today? Cloudy. Why is it always like this?", then input T into the rewriting operation judgment model encoder to obtain the semantic encoding vector H = [h1,...,h 19 The rewrite operation judgment model encoder includes, but is not limited to, LSTM, GRU, CNN, Transformer, pre-trained language model BERT, and various variants of BERT. Finally, the rewrite operation judgment model decoder is used to decode the semantic encoded information to obtain the insertion operation matrix and the replacement operation matrix. First, the insertion operation decoding vector X1 = f is obtained. θ1 (H) and the replacement operation decoding vector X2 = f θ2 (H), where f θ1 and f θ2 Let X1 and X2 represent functions with parameters θ1 and θ2, respectively. Then, the transposes X1^T and X2^T corresponding to the dot products of X1 and X2 are used, respectively. Finally, it is determined whether the result of the dot product is greater than a set threshold to obtain the insertion operation matrix I and the replacement operation matrix R. Where I∈[0,1] 19×19 R∈[0,1] 19×19 .
[0059] S3. Extract insertion operation information based on the insertion operation matrix.
[0060] Specifically, insertion operation information is extracted using an insertion operation information extraction algorithm, including:
[0061] S31. Traverse each row of the insertion operation matrix corresponding to the incomplete dialogue. If there are two columns with a value of 1 in the traversed row and they correspond to the historical dialogue, extract the historical dialogue fragments corresponding to the two columns and mark the position of the row corresponding to the incomplete dialogue as the insertion position.
[0062] S32. Save the historical dialogue fragment and its corresponding insertion position as a tuple into the insertion operation information.
[0063] S33. By decoding historical dialogues and incomplete dialogues into insertion operation matrices, multiple omissions are allowed in incomplete dialogues, thus enabling the rewriting of multiple omissions in incomplete dialogues.
[0064] Specifically, for the input T in S2, traverse each element I[i] in the corresponding insertion operation matrix (i ∈ [13, 19]). If there are two elements I[i,j] and I[i,k] with a value of 1 in I[i], and j < k and k < 13, then save (T[j:k], i - 13) to the insertion operation information, where T[j:k] represents the conversation segment from the j-th to the k-th in T. For the input example, the final insertion operation information list is: [(Chengdu, 1)].
[0065] The above "value of 1" indicates a pointing relationship. For example, in the insertion operation matrix, the "wei" in "Why is it always like this" corresponds to I
[13] (the "wei" is in the 13th position in "What's the weather like in Chengdu today? Cloudy. Why is it always like this." And each value in I
[13] ∈ [0,1]19 corresponds one by one in order with "What's the weather like in Chengdu today? Cloudy. Why is it always like this."). When I
[13] [i] = 1, it means that "wei" points to the i-th position in I
[13] . In the example sentence, "Chengdu" is omitted before "wei", so except for the first (Cheng) and the second (du) elements in I
[13] with a value of 1, other elements have a value of 0 (if "Chengdu today" is omitted before "wei", then the first and the fourth have a value of 1, and others have a value of 0. That is, it points to the start and end of the omitted text).
[0066] S4. Extract replacement operation information based on the replacement operation matrix.
[0067] Specifically, use the replacement operation information extraction algorithm to extract replacement operation information, including:
[0068] S41. Traverse each row of the replacement operation matrix corresponding to the incomplete conversation. If there is 1 column with a value of 1 in the traversed row and it corresponds to the historical conversation, record the row number and column number as h1 and c1 respectively;
[0069] S42. Continue to traverse each row of the replacement operation matrix from the next row of h1. If there is 1 column with a value of 1 in the traversed row and it corresponds to the historical conversation, record the row number and column number as h2 and c2 respectively;
[0070] S43. Extract the segment from h1 to h2 in the incomplete conversation as the segment to be replaced, and extract the segment from c1 to c2 in the historical conversation as the corresponding replacement segment;
[0071] S44. Save the segment to be replaced and the corresponding replacement segment as a binary tuple to the replacement operation information.
[0072] By decoding the historical conversation and the incomplete conversation into the replacement operation matrix, multiple referential items in the incomplete conversation are allowed, so that multiple referential items in the incomplete conversation can be rewritten.
[0073] Specifically, for the input T in S2, traverse each element R[i] (i ∈ [13, 19]) in the corresponding insertion operation matrix. If there is an element R[i,j] with a value of 1 in R[i], and j < 13, record i and j; then continue traversing from the next row of R to find another element R[m,n] with a value of 1, where i < m < 19 and j < n < 13, and record m and n; finally, extract the text segment T[j:n] in T as the segment to be replaced, extract T[i:m] as the corresponding replacement segment, and save (T[i:m], T[j:n]) to the replacement operation information. For the input example in S2, the final operation information list is obtained: [(cloudy day, like this)].
[0074] S5. Perform the final rewriting of the original incomplete dialogue based on the insertion operation information and the replacement operation information.
[0075] Specifically, S5 includes:
[0076] S51. Traverse each binary tuple in the insertion operation information, and insert the historical dialogue segment therein after the insertion position;
[0077] S52. Traverse each binary tuple in the replacement operation information, and replace the segment to be replaced in the incomplete dialogue with the replacement segment.
[0078] Specifically, traverse the insertion operation information list in S3 to obtain the binary tuple (Chengdu, 1), and insert "Chengdu" before the first Chinese character in the incomplete dialogue to get "Why is it always like this in Chengdu"; then traverse the replacement operation information list in S4 to obtain the binary tuple (cloudy day, like this), and replace "like this" with "cloudy day" according to the binary tuple information to get "Why is it always cloudy in Chengdu".
[0079] According to the method provided by the embodiment of the present invention, a high degree of coincidence between the incomplete dialogue and the dialogue with complete information can be achieved, a dialogue that is semantically consistent with the historical dialogue information can be generated, and an incomplete dialogue with one or more references and ellipsis items can be rewritten simultaneously, so as to supplement the information in the complete dialogue and improve the effect of dialogue understanding.
[0080] Embodiment 2
[0081] Figure 2 Show the structural block diagram of an apparatus for rewriting an incomplete dialogue according to the present invention, including:
[0082] A data processing module 10, configured to obtain a historical dialogue and preprocess the historical dialogue and the incomplete dialogue.
[0083] Specifically, it retrieves one or more historical dialogue statements preceding the incomplete dialogue; and performs preprocessing operations on the historical dialogue and the incomplete dialogue, including but not limited to removing special symbols, converting between half-width and full-width characters, and converting between simplified and traditional Chinese characters.
[0084] The rewrite operation matrix acquisition module 20 is used to input historical dialogue and incomplete dialogue content into the rewrite operation judgment model to obtain the insertion operation matrix and the replacement operation matrix.
[0085] Specifically, the historical dialogue and the incomplete dialogue are concatenated; the rewrite operation judgment model encoder is used to semantically encode the content of the concatenated historical dialogue and the incomplete dialogue; and the rewrite operation judgment model decoder is used to decode the semantically encoded information to obtain the insertion operation matrix and the replacement operation matrix.
[0086] The insertion operation information extraction module 30 is used to extract insertion operation information using an insertion operation information extraction algorithm.
[0087] Specifically, an insertion operation information extraction algorithm is used to extract insertion operation information.
[0088] The replacement operation information extraction module 40 is used to extract replacement operation information using a replacement operation information extraction algorithm.
[0089] Specifically, a replacement operation information extraction algorithm is used to extract replacement operation information.
[0090] The dialogue modification module 50 is used to rewrite the original incomplete dialogue based on the insertion operation information and the replacement operation information.
[0091] Specifically, iterate through each tuple in the insertion operation information and insert the historical dialogue fragment into the insertion position; iterate through each tuple in the replacement operation information and replace the fragment to be replaced in the incomplete dialogue with the replacement fragment.
[0092] The incomplete dialogue rewriting device provided in Embodiment 2 of the present invention can achieve a high degree of overlap between incomplete dialogue and complete dialogue, generate dialogue that is semantically consistent with historical dialogue information, and rewrite incomplete dialogue with one or more referents and omissions, thereby supplementing the information in the complete dialogue and improving the effect of dialogue understanding.
[0093] Example 3
[0094] This embodiment discloses a storage medium storing an incomplete dialogue rewriting program. When the incomplete dialogue rewriting program is executed by a processor, it implements the steps of the incomplete dialogue rewriting method as described in Embodiment 1.
[0095] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for rewriting incomplete dialogues, characterized in that, include: Obtain historical dialogues and preprocess the historical dialogues and incomplete dialogues; Input historical dialogues and incomplete dialogues into the rewrite operation judgment model to obtain the insertion operation matrix and the replacement operation matrix; Extracting insertion operation information based on the insertion operation matrix; The method for extracting insertion operation information based on the insertion operation matrix includes: Traverse each row of the insertion operation matrix corresponding to the incomplete dialogue. If there are two columns with a value of 1 in the traversed row and they correspond to the historical dialogue, then extract the historical dialogue fragments corresponding to the two columns as the fragments to be inserted, and mark the index value of the row as the position to be inserted. The segment to be inserted and the position to be inserted are stored as a tuple in the insertion operation information; Extract replacement operation information based on the replacement operation matrix; The method for extracting replacement operation information based on the replacement operation matrix includes: Replacement operation information is extracted using a replacement operation information extraction algorithm, including: For each row of the replacement operation matrix corresponding to the incomplete dialogue, if there is a column with a value of 1 in the traversed row and it corresponds to the historical dialogue, record the row number and column number as h1 and c1 respectively. Continue traversing each row of the replacement operation matrix from the row below h1. If there is a column with a value of 1 in the traversed row that corresponds to the historical dialogue, record the row number and column number as h2 and c2 respectively. Extract fragments h1 to h2 from the incomplete dialogue as fragments to be replaced, and extract fragments c1 to c2 from the historical dialogue as the corresponding replacement fragments. Save the segment to be replaced and the corresponding replacement segment as a tuple in the replacement operation information; The original incomplete dialogue is finally rewritten based on the insertion and replacement operation information.
2. The incomplete dialogue rewriting method according to claim 1, characterized in that, The method for acquiring historical dialogues and preprocessing the historical dialogues and incomplete dialogues includes: Retrieve one or more historical dialogue statements preceding the incomplete dialogue; Preprocessing operations are performed on historical dialogues and incomplete dialogues. These preprocessing operations include removing special symbols, converting between half-width and full-width characters, and converting between simplified and traditional Chinese characters.
3. The incomplete dialogue rewriting method according to claim 1, characterized in that, The method for inputting historical dialogue and incomplete dialogue content into the rewrite operation judgment model to obtain the insertion operation matrix and the replacement operation matrix includes: Connecting historical dialogues and incomplete dialogues; The rewrite operation is used to determine the semantic encoding of the concatenated historical dialogue and incomplete dialogue content by the encoder. The rewrite operation judgment model decoder is used to decode the semantic encoded information to obtain the insertion operation matrix and the replacement operation matrix.
4. The incomplete dialogue rewriting method according to claim 1, characterized in that, The method for finally rewriting the original incomplete dialogue based on insertion operation information and replacement operation information includes: Iterate through each tuple in the insertion operation information and insert the historical dialogue fragment into the insertion position. Iterate through each tuple in the replacement operation information and replace the fragment to be replaced in the incomplete dialogue with the replacement fragment.
5. An incomplete dialogue rewriting apparatus, used to perform the incomplete dialogue rewriting method as described in any one of claims 1-4, characterized in that, include: The data processing module is used to acquire historical dialogues and preprocess the historical dialogues and incomplete dialogues. The rewrite operation matrix acquisition module inputs historical dialogues and incomplete dialogues into the rewrite operation judgment model to obtain the insertion operation matrix and the replacement operation matrix. The insertion operation information extraction module is used to extract insertion operation information using an insertion operation information extraction algorithm. The replacement operation information extraction module is used to extract replacement operation information using a replacement operation information extraction algorithm. The dialogue modification module is used to rewrite the original incomplete dialogue based on the insertion and replacement operation information.
6. A storage medium, characterized in that, The storage medium stores an incomplete dialogue rewriting program, which, when executed by a processor, implements the steps of the incomplete dialogue rewriting method as described in any one of claims 1-4.