A method for extracting and estimating time text in multi-round interactions of intelligent outbound calls

Abstract

The present invention discloses a method for extracting and calculating time text in multi-round interactions of intelligent outbound calls, comprising the following steps: S1, using an ASR algorithm to convert the real-time audio of a user speaking during the intelligent interaction process into text information for the multi-round interactions; S2, constructing a stop word library to remove irrelevant vocabulary content, screening valid text information through preprocessing, and performing text feature engineering processing and feature representation on the text information; S3, obtaining time feature text expression information in the interaction through a Bert-Attention intention recognition model, and dividing the text features in the intelligent interaction into time feature intentions. The present invention uses the Bert-Attention intention recognition model to divide text intentions based on whether they have time features, thereby narrowing the scope of text time node extraction; extracting time node text information through the HanLP framework and the Bert-BiLSTM-CRF named entity extraction fusion strategy, thereby improving the accuracy of time extraction; and realizing the ability of the intelligent outbound call system to schedule outbound calls based on the time mentioned by the user.

Description

Technical Field

[0001] The present invention relates to the field of communication payment, and in particular to a method for extracting and calculating time text in multi-round interactions of intelligent outbound calls. Background Art

[0002] Currently, in the financial loan field, intelligent outbound calls can remind users to repay their loans in a timely and effective manner. Usually, users will provide feedback on the expected repayment time during the interaction. This capability requires obtaining the specific effective time information mentioned by users in intelligent multi-round interactions. Traditional mainstream methods mostly use rule matching methods to match time field information from the text information described by users. However, when users' expressions in multi-round interactions are too colloquial and the time points are mentioned multiple times without clear meaning, they lack the ability to understand the time intentions of multi-round interactions and to calculate the effective time points. As a result, it is impossible to accurately obtain the effective time nodes expressed by users, resulting in large errors in the scheduled outbound call reminder function and the risk of harassing users.

[0003] Therefore, this patent proposes: 1. Through the intent recognition model, the text features in the intelligent interaction are divided into time feature intentions, and the time feature text expression information of multiple rounds of interaction is spliced ​​to effectively improve the ability to extract time information; 2. Multiple time features and event entities are extracted through the named entity model, and the word segmentation part-of-speech strategy is integrated to extract the user mentioned time extraction method, and the time rule inference logic is formulated to realize the ability to convert the text expression time into the standard time date. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to overcome the defects of the existing technology and provide a method for extracting and calculating time text in multi-round interactions of intelligent outbound calls. By extracting effective time node information during the intelligent outbound call interaction process and converting it into a standard digital time and date, a more accurate method for extracting the real time and date expressed by the user is achieved, thereby realizing the ability to schedule outbound call reminders based on the time and date mentioned by the user.

[0005] The present invention provides the following technical solutions:

[0006] The present invention provides a method for extracting and calculating time text in intelligent outbound multi-round interactions, comprising the following steps:

[0007] S1. Use ASR algorithm to convert the real-time audio of the user's speech during the intelligent interaction process into text information for multiple rounds of interaction;

[0008] S2. Build a stop word library to remove irrelevant vocabulary, filter valid text information through preprocessing, and perform text feature engineering and feature representation on the text information;

[0009] S3. Use the Bert-Attention intention recognition model to obtain the temporal feature textual representation information in the interaction, and divide the textual features in the intelligent interaction into temporal feature intentions, effectively improving the ability to extract temporal information and narrowing the scope of text matching at time nodes.

[0010] S4, the method for extracting effective time of user mentions in multi-round interactions is as follows Figure 2 As shown in the figure, the text information of the time intent category is divided into rule matching and named entity extraction to extract the time nodes. The specific steps are as follows:

[0011] 1) Build a delay time vocabulary to determine whether the user's text contains time delay logic. If it does, the IsDelay flag is set to 1; otherwise, the flag is set to 0. For example, "Wait for 5 minutes to call back," where "wait" is a keyword for delay time logic.

[0012] 2) Build a Bert-BiLSTM-CRF named entity extraction model, input text feature information, and output the time entities present in the current text, such as "5:00 PM: time entity", thereby clarifying the set of time named entities in the text;

[0013] 3) Use the HanLP framework to segment the text and mark the part of speech, match the quantifiers, number words, and time words in the text, and splice the part of speech time node text collection that appears in the text;

[0014] 4) Take the intersection of the time nodes shared by the time named entity and the part-of-speech time node text set, remove the repeated inclusion relationship time text, and obtain the single-round interaction time node text set;

[0015] S5. Construct a time text rule calculation method. The calculation process is as follows: Figure 3 As shown, it realizes the conversion of text time to digital time and date; the specific steps are as follows:

[0016] 1) Determine whether there is an afternoon word, such as "evening". If so, record the IsPM flag as 1; otherwise, record the flag as 0;

[0017] 2) Determine whether a time mapping relationship exists, such as "tomorrow: +1day", and store the elapsed time according to the time mapping flag PassTime;

[0018] 3) Formulate regular extraction rules for time units to match the year, month, day, hour, minute, and second information that appears in the text, and set the time unit to be T2 (Y, m, d, H, M, S);

[0019] 4) Calculate the time and date based on the time information extracted from the text. If IsDelay is 0, the time and date are calculated as follows:

[0020] T=today().replace(T2)+IsDelay*T2+IsPM*H 12 +PassTime

[0021] If IsDelay is 1, the time and date are calculated as follows:

[0022] T=today()+IsDelay*T2+IsPM*H 12 +PassTime

[0023] In the above formula, T is the final date and time, H 12 Indicates 12 hours. The today function obtains the current date and time T1. The calculation method is to determine whether the shift logic IsDelay uses the replace function to replace the time unit of T1 with the time unit extracted to T2, and finally add the time offset IsPM*H 12 , PassTime, IsDelay*T2 to get the specific time and date;

[0024] S6. Establish effective time verification rules to determine whether the converted time and date meet the effective time. If yes, the extraction is successful and the outbound call continues with the original process. Otherwise, multiple rounds of inquiries are required to clarify the specific time mentioned by the user, and the time and date are extracted and converted again.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] 1. The Bert-Attention intent recognition model is used to classify text intent based on whether it has temporal features, narrowing the scope of text time node extraction.

[0027] 2. Extract time node text information through the HanLP framework and the Bert-BiLSTM-CRF named entity extraction fusion strategy to improve the accuracy of time extraction;

[0028] 3. Use the time calculation formula to convert text time information into a calculable time and date, and then convert the time and date into a scheduled outbound call task, enabling the intelligent outbound call system to schedule outbound calls based on the time mentioned by the user. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0030] Figure 1 This is a flow chart of the text time calculation method for intelligent outbound multi-round interaction of the present invention;

[0031] Figure 2 This is a flow chart of the intelligent interactive user mention effective time extraction method of the present invention;

[0032] Figure 3 This is a logic flow chart for calculating the time rule of intelligent interactive user mentions of the present invention. DETAILED DESCRIPTION

[0033] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention and are not intended to limit the present invention. In the accompanying drawings, the same reference numerals all refer to the same components.

[0034] Example 1

[0035] like Figure 1-3 The present invention provides a method for extracting and calculating time text in intelligent outbound multi-round interactions. The detailed embodiment includes the following process steps:

[0036] S1. The intelligent outbound call timed repayment reminder business scenario usually uses multiple outbound calls to remind users of their repayment ability. During the intelligent interaction process, the user-side audio track is stripped out, and the real-time audio stream is converted into text information by using the ASR algorithm to convert the user's speech content.

[0037] S2. Build a stop word library, remove irrelevant vocabulary, filter valid text information in interactive content, and perform text feature engineering and feature representation:

[0038] 1) Text feature engineering: First, remove single-word spoken vocabulary and stop words from the ASR-parsed text. Then, remove punctuation and special characters from the text. Next, use HanLP to segment the text into a set of segmented words. Finally, remove duplicates from the set of segmented words and concatenate them according to the original text order.

[0039] 2) Text feature representation: Both the intent recognition and named entity recognition models use the Bert pre-trained model as the base model. The text information is encoded using the BertTokenizer tagger and the BertEncoder encoder to achieve text feature representation.

[0040] S3. Build a Bert-Attention intent recognition model, input the text semantic features of the interaction process, mine the temporal feature information through the Attention mechanism layer, and finally output the text intent classification through the Softmax classifier to determine whether the current text belongs to the temporal intent category, effectively improving the ability to extract temporal information and narrowing the matching range of text with time nodes.

[0041] S4, the method for extracting effective time of user mentions in multi-round interactions is as follows Figure 2 As shown in the figure, the text information of the time intent category is divided into rule matching and named entity extraction to extract the time nodes. The specific steps are as follows:

[0042] 1) Build a time delay vocabulary to determine whether the user's text contains time delay logic. If so, the IsDelay flag is set to 1; otherwise, the flag is set to 0. For example, "Wait for 5 minutes and call back," where "wait" is a keyword for time delay logic.

[0043] 2) Build a Bert-BiLSTM-CRF named entity extraction model. Input text feature information, use the BiLSTM layer to obtain bidirectional semantic features, use the CRF layer to calculate the probability distribution of entity occurrences, and output the time entities present in the current text, such as "5:00 PM: time entity," thereby identifying the set of time named entities in the text.

[0044] 3) Use the HanLP framework to segment the text and mark the part of speech, match the quantifiers, number words, and time words in the text, and splice the part of speech time node text collection that appears in the text;

[0045] 4) Take the intersection of the time named entity set and the time nodes shared by the part-of-speech time node set, remove the repeated inclusion relationship time nodes, and obtain the single-round interaction time node text set.

[0046] S5. Construct a time text rule calculation method. The calculation process is as follows: Figure 3 As shown in the figure, it realizes the conversion of text time to digital time and date. The specific steps are as follows:

[0047] 1) Create an afternoon vocabulary list of ["evening", "afternoon", "noon", "afternoon"], and determine whether the text contains afternoon vocabulary. If so, set the IsPM flag to 1; otherwise, set the flag to 0.

[0048] 2) Formulate a time mapping relationship as shown in Table 1, determine whether a time mapping relationship exists, and store the elapsed time according to the time mapping flag PassTime;

[0049] Table 1 Part of the time mapping relationship

[0050] Time Description Mapping Values today +0day tomorrow +1day the day after tomorrow +2day next month +30day

[0051] 3) Formulate regular extraction rules for time units to match the year, month, day, hour, minute, and second information that appears in the text, and set the time unit to be T2 (Y, m, d, H, M, S);

[0052] 4) Calculate the time and date based on the time information extracted from the text. If IsDelay is 0, the time and date are calculated as follows:

[0053] T=today().replace(T2)+IsDelay*T2+IsPM*H 12 +PassTime

[0054] If IsDelay is 1, the time and date are calculated as follows:

[0055] T=today()+IsDelay*T2+IsPM*H 12 +PassTime

[0056] In the above formula, T is the final date and time, H 12 Indicates 12 hours. The today function obtains the current date and time T1. The calculation method is to determine whether the shift logic IsDelay uses the replace function to replace the time unit of T1 with the time unit extracted to T2, and finally add the time offset IsPM*H 12 , PassTime, IsDelay*T2 to get the specific time and date.

[0057] S6. Develop effective time verification rules for outbound call business scenarios as shown in Table 2 below. Determine whether the converted time and date meet the effective time. If so, the extraction is successful, and the outbound call process continues to broadcast according to the normal process. Otherwise, it means that the precise time point cannot be extracted. The user is asked again for the specific time through the dialogue, the text time node is extracted, the time and date are converted, and the scheduled outbound call task is performed according to the time point to realize the ability to remind the user on a regular basis.

[0058] Table 2 Validity time verification rules

[0059] Serial number rule 1 Is the time unit accurate to the hour? 2 Whether the time is historical time 3 Does the time match the outbound call time period (8:00-20:00)? 4 Does the time meet the maximum allowed span date (two months from now)?

[0060] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A method for extracting and calculating time text in multi-round interactions of intelligent outbound calls, characterized in that: The following steps are involved: S1. Use ASR algorithm to convert the real-time audio of the user's speech during the intelligent interaction process into text information for multiple rounds of interaction; S2. Build a stop word library to remove irrelevant vocabulary, filter valid text information through preprocessing, and perform text feature engineering and feature representation on the text information; S3. Use the Bert-Attention intention recognition model to obtain the temporal feature textual expression information in the interaction and divide the textual features in the intelligent interaction into temporal feature intentions; S4. In the method for extracting effective time mentioned by users in multiple rounds of interaction, the text information classified into time intent categories by intent recognition is extracted through rule matching and named entity extraction. The specific steps are as follows: 1) Build a delay time vocabulary to determine whether the user's text contains time delay logic. If it does, record the IsDelay flag as 1; otherwise, record the flag as 0; 2) Build a Bert-BiLSTM-CRF named entity extraction model, input text feature information, output the time entities existing in the current text, and clarify the set of time named entities in the text; 3) Use the HanLP framework to segment the text and mark the part of speech, match the quantifiers, number words, and time words in the text, and splice the part of speech time node text collection that appears in the text; 4) Take the intersection of the time nodes shared by the time named entity and the part-of-speech time node text set, remove the repeated inclusion relationship time text, and obtain the single-round interaction time node text set; S5. Construct a time text rule calculation method to convert text time into digital time and date. The specific steps are as follows: 1) Determine whether there is a word for afternoon. If so, set the IsPM flag to 1; otherwise, set the flag to 0. 2) Determine whether a time mapping relationship exists and store the elapsed time according to the time mapping flag PassTime; 3) Formulate regular extraction rules for time units to match the year, month, day, hour, minute, and second information that appears in the text, and set the time unit to be T2 (Y, m, d, H, M, S); 4) Calculate the time and date based on the time information extracted from the text. If IsDelay is 0, the time and date are calculated as follows: T=today().replace(T2)+IsDelay*T2+IsPM*H 12 +PassTime If IsDelay is 1, the time and date are calculated as follows: T=today()+IsDelay*T2+IsPM*H 12 +PassTime In the above formula, T is the final date and time, H 12 Indicates 12 hours. The today function obtains the current date and time T1. The calculation method is to determine whether the shift logic IsDelay uses the replace function to replace the time unit of T1 with the time unit extracted to T2, and finally add the time offset IsPM*H 12 , PassTime, IsDelay*T2 to get the specific time and date; S6. Establish effective time verification rules to determine whether the converted time and date meet the effective time. If yes, the extraction is successful and the outbound call continues with the original process. Otherwise, multiple rounds of inquiries are required to clarify the specific time mentioned by the user, and the time and date are extracted and converted again.

Images