Determining topic tags for a communication transcript based on a trained generative summarization model

Abstract

The disclosure herein describes determining a topic of a communication transcript using a trained summarization model. A first communication transcript associated with a first communication is obtained and divided into a first set of communication segments. Based on the first set of communication segments, a first set of topic descriptions is generated by analyzing each communication segment in the first set of communication segments with a generative language model. The summarization model is trained using the first set of communication segments and the associated first set of topic descriptions as training data. The trained summarization model is then applied to a second communication transcript, and based on applying the trained summarization model to the second communication transcript, a second set of topic descriptions for the second communication transcript is generated. By training the summarization model based on the output of the generative language model, it is enabled to efficiently and accurately generate topic descriptions from communication transcripts.

Description

Background Technology

[0001] Customer Relationship Management (CRM) sessions and other related communications are favorable targets for analysis. Accurately and effectively identifying session segments with high specificity and determining the topics of those segments (e.g., accurately identifying topics from a set of many possible topics) is crucial. Expanding to use many possible topics may require large amounts of labeled data, and topic customization also demands additional time and effort. Summary of the Invention

[0002] This summary is provided to introduce, in a simplified form, the selection of concepts further described below in the detailed description. This summary is not intended to identify key or essential features of the claimed subject matter, nor is it intended to help determine the scope of the claimed subject matter.

[0003] A computerized method for determining topics in communication transcripts is described. A first communication transcript associated with a first communication is obtained, and the first communication transcript is divided into a first communication segment set. Based on the first communication segment set, a first topic description set is generated by analyzing each communication segment in the first communication segment set using a generative language model. A summarization model is trained using the first communication segment set and the associated first topic description set as training data. The trained summarization model is then applied to a second communication transcript, and a second topic description set for the second communication transcript is generated based on the application of the trained summarization model to the second communication transcript. Attached Figure Description

[0004] This instruction manual will be better understood by reading the following detailed description with reference to the accompanying drawings, in which:

[0005] Figure 1 This is a block diagram illustrating a system configured to determine topic tags for communication transcription according to an embodiment;

[0006] Figure 2 This is a schematic diagram illustrating the generation of topic descriptions from training communications using a generative language model, according to an embodiment.

[0007] Figure 3 This is a schematic diagram illustrating the training of a summarization model based on segmented description tuples, according to an embodiment.

[0008] Figure 4 This is a schematic diagram illustrating the mapping from topic descriptions to topic tags in a topic directory, according to an embodiment.

[0009] Figure 5 This is a flowchart illustrating a computerized method for training a summarization model to generate topic descriptions from communication transcription, according to an embodiment.

[0010] Figure 6 This is a flowchart illustrating a computerized method for providing segmented topic tags for communication transcription using a summarization model, according to an embodiment.

[0011] Figure 7 This is a schematic diagram illustrating a graphical user interface (GUI) displaying topic tags associated with communication segments according to an embodiment; and

[0012] Figure 8 A computing device according to an embodiment is shown as a functional block diagram.

[0013] In all the accompanying drawings, the corresponding reference numerals indicate the corresponding parts. Figures 1 to 8 The system is shown as a schematic diagram. The accompanying drawings may not be drawn to scale. Detailed Implementation

[0014] This disclosure provides computerized methods and systems for training a summarizing model and using the trained summarizing model to identify topic tags for segments of communication transcripts. Training communication transcripts associated with some sessions or other communications (e.g., communications indicating the type of communication to be analyzed) are acquired and the transcripts are segmented into communication segments. A generative language model (GLM) (such as GPT-3) is then used with each segment to generate associated topic descriptions. Pairs of segments and topic descriptions are used as training data to generate a summarizing model, such that the summarizing model is configured to generate topic descriptions when applied to communication segments. The trained summarizing model can then be applied to other communication transcripts, and topic descriptions for those transcripts are generated based on the applied training model. Additionally, a topic catalog can be defined as including a set of topic tags and supporting examples (e.g., example topic descriptions) associated with each of those topic tags. The topic descriptions output from the summarizing model can be compared with the supporting examples in the topic catalog, and the topic descriptions are mapped to associated topic tags based on the identified matching supporting examples. Those topic tags can be provided as output in response to received communication transcripts.

[0015] This disclosure operates in an unconventional manner by using a GLM such as GPT-3 to train a smaller, more focused generative model, and then normalizing the model's output to a fixed set of topics (e.g., topic tags from a topic catalog). Using a large GLM to train a more focused model results in improved performance and reduced resource requirements when generating topic descriptions at runtime, since larger, more general GLMs can have significantly greater resource and processing requirements. Furthermore, normalizing the output of the trained model by mapping it to differentiated topic tags via a topic catalog reduces or eliminates the illusion and / or vulnerability to adversarial attacks that rely solely on GLMs to generate topic descriptions.

[0016] Furthermore, this disclosure enables the generation of customized, centralized summarization models and customized topic directories, thereby allowing for fine-tuning of the final topic tagging system and methodology for analyzing specific types of communications. Users can provide customized sets of topic tags and associated support examples. Additionally, this disclosure can be configured to provide recommended support examples to users in response to customized topic tag sets. The use of the described topic directories and associated mapping methods allows users to edit and / or add to a set of topic tags that can be identified relative to the input communications transcription, reducing human intervention and improving the flexibility of the described communications analysis system and methodology.

[0017] Figure 1 This is a block diagram illustrating a system 100 configured to determine topic tags 130 of a communication transcript 122 according to an embodiment. The system includes a summarization model trainer 102 configured to train a summarization model 120, and a topic tagging engine 104 configured to use the summarization model 120 to generate topic tagging results 132 from the communication transcript 122. In some examples, the system 100 resides on, is stored on, and / or executes on a computing device such as a personal computer, server device, tablet computing device, or other mobile device. For example, a server device may be configured to perform the operation of the summarization model trainer 102 and the topic tagging engine 104.

[0018] Alternatively, in other examples, system 100 is distributed across multiple computing devices, such that components, elements, and / or parts of system 100 can reside on and / or execute on different computing devices that communicate with each other (e.g., via one or more communication networks, such as an intranet, the Internet, etc.). For example, system 100 may be configured to store data associated with the operation of the summary model trainer 102 and / or the topic tagging engine 104 on one or more distributed storage devices, and / or system 100 may be configured to execute the summary model trainer 102 and / or the topic tagging engine 104 on one or more distributed computing devices (e.g., the summary model trainer 102 executes on a first server device, and the topic tagging engine 104 executes on a second server device). In other examples, other arrangements of computing devices may be used to implement system 100 without departing from this specification.

[0019] The summarization model trainer 102 includes hardware, firmware, and / or software configured to train the summarization model 120 based on training communication transcription 106. The summarization model trainer 102 includes: a transcription segmentation engine 108 configured to divide the training communication transcription 106 into communication segments 110; a generative language model 112 (GLM) that generates segmented description tuples 114 from the communication segments 110; and a model training engine 118 that uses the segmented description tuples 114 to train and / or tune the summarization model 120, as described herein.

[0020] In some examples, training communication transcripts 106 include natural language text data of the language used during the communication, such as telephone calls, video calls, instant messaging chat logs, and / or other forms of conversation between the parties. In relevant examples, one-way communications such as voicemails can be analyzed as described herein without departing from this specification. Training communication transcripts 106 may be manually generated by a transcriptist who listens to or otherwise observes the associated communications, and / or they may be automatically generated using speech-to-text or speech-to-text analysis of audio streams or other recordings of the communications. Additionally or alternatively, training communication transcripts 106 may include data indicating the words and phrases used during the communication and / or other data associated with the communication, such as punctuation used, timing data associated with the communication (e.g., when words are spoken, pause lengths between sentences, etc.).

[0021] Transcription segmentation engine 108 includes hardware, firmware, and / or software configured to divide a transcription of communication (such as training communication transcription 106) into communication segments 110. In some examples, transcription segmentation engine 108 is configured to divide transcription 106 into communication segments 110 by identifying approximately coherent topic portions of the communication (e.g., each segment 110 includes communication data of the communication associated with a single topic, and each segment 110 may be associated with different topics from other segments 110 of transcription 106). For example, transcription segmentation engine 108 may be configured to identify each sentence in transcription 106 and vectorize the identified sentences (e.g., using a bidirectional encoder representation from a Transformer (BERT) technique, etc.). The sentence vectors of transcription 106 can then be grouped based on similarity (e.g., sentence vector groups can be determined based on maximizing accumulated weighted cosine similarity using text segmentation implementations, etc.). The resulting communication segments 110 comprise groups of sentences from the transcripts 106 being analyzed, grouped such that all sentences within each group are related to a specific topic. It should be understood that in other examples, other techniques may be used to segment the training communication transcripts 106 into communication segments 110 of sentences grouped by topic, without departing from the description herein.

[0022] In some examples, communication segments 110 from training communication transcription 106 are provided to GLM 112 (e.g., Generative Pre-trained Transformer 3 (GPT-3)) and / or GLM 112 is applied to communication segments 110. Each communication segment 110 can be processed individually using GLM 112 as described herein. GLM 112 includes hardware, firmware, and / or software configured to interpret the language of communication segments 110 and generate topic descriptions 116 associated with communication segments 110, which are related to... Figure 1 The communication segments 110 in the segmented description tuple 114 shown are associated. It should be understood that in other examples, other data arrangements or organizations may be used to represent the association between the communication segments 110 and the topic description 116 without departing from the description herein.

[0023] Figure 2 This is a schematic diagram 200 illustrating, according to an embodiment, the generation of topic descriptions 216 from training communication transcription 206 using GLM 212. In some examples, schematic diagram 200 shows a system (such as...) Figure 1The process is performed in system 100. Training communication transcription 206 is divided into communication segments 210 (e.g., segments 210A, 210B, and 210C), and each of those segments 210 is provided as input to GLM 212. The result is topic description 216, which includes a description of each segment in segment 210 in the form of topic descriptions 216A, 216B, and 216C. In some examples, GLM 212 generates one topic description 216 for each communication segment 210 (e.g., description 216A for segment 210A, description 216B for segment 210B, and description 216C for segment 210C). Alternatively, GLM 212 can be configured to generate one or more topic descriptions 216 for each communication segment 210 (e.g., GLM 212 can identify two possible topic descriptions 216 for segment 210 and provide both as descriptions 216 for that segment 210).

[0024] In some examples, GLM 212 is trained to use deep learning techniques, such as how to configure a GPT-3 model, to interpret and / or generate human-like natural language text. GLM 212 can be configured to perform natural language processing (NLP) on communication segments 210 and generate a natural language human-like topic description 216 for the communication segments 210 (e.g., phrases, sentences, or paragraphs describing the topics of the communication segments 210). In some examples, in addition to providing GLM 212 with communication segments 210 for processing, GLM 212 is prompted to provide a topic description of the provided communication segments 210 (e.g., questions can be asked to GLM 212, such as "What topics are discussed in this segment?" etc.). In such examples, GLM 212 is configured and / or trained to interpret both the segments and the associated prompts and respond to the prompts with human-like responses.

[0025] return Figure 1 The segmented description tuple 114 generated using the topic description 116 from GLM 112 is used as training data by model training engine 118. Model training engine 118 includes hardware, firmware, and / or software configured to train summarizing model 120 using the segmented description tuple 114 as input. Model training engine 118 can be configured to train summarizing model 120 using machine learning techniques to accurately assign topic descriptions to input communication segments. Thus, the trained summarizing model 120 generates topic descriptions from communication segments that are substantially similar to the topic description 116 generated by GLM 112 from the same communication segments.

[0026] However, the trained summarizing model 120 has the advantage of being smaller and less resource-intensive than GLM 112, which is configured to interpret and generate natural language with a more general ability. The trained summarizing model 120 is fine-tuned to simulate the output of GLM 112 for a much narrower set of possible outputs (e.g., the summarizing model 120 may only assign topic descriptions previously generated by GLM 112 in segmented description tuples 114). These differences between the summarizing model 120 and GLM 112 result in the summarizing model 120 having a narrower set of potential uses, but being more resource and / or time efficient on those potential uses than the more general GLM 112.

[0027] Figure 3 This is a schematic diagram 300 illustrating, according to an embodiment, training a summarizing model 320 based on segmented description tuples 314. In some examples, the process shown in schematic diagram 300 is in a system (such as...) Figure 1 The process is executed in system 100. Tuples 314 (including tuples of communication segment 310A and topic description 316A, tuples of communication segment 310B and topic description 316B, and tuples of communication segment 310C and topic description 316C) are used as example inputs and outputs to train the summarization model 320 to generate similar topic descriptions 324 (e.g., descriptions 324A, 324B, and 324C) and / or assign similar topic descriptions 324 (e.g., descriptions 324A, 324B, and 324C) to similarly provided communication segments (e.g., segments of communication transcription 122).

[0028] In some examples, machine learning techniques (e.g., via a summarizing model trainer 318) are used to train the summarizing model 320. The summarizing model 320 can be trained using segmented description tuples 314, which indicate an accurate set of topic descriptions based on associated communication segments. Such training data can include data mapping to many different kinds of communication segments that are associated with accurate topic descriptions. When communication segments are provided, the segmented description tuples 314 can be used to train the summarizing model 320 to identify accurate topic descriptions (e.g., from a set of different topic descriptions provided in the segmented description tuples 314).

[0029] In some examples, the training of the summarizing model 320 includes machine learning techniques, such as using trained regressors like random decision forests, directed acyclic graphs, support vector machines, convolutional neural networks, or other neural networks, or another trained regressor. It should further be understood that when applying machine learning techniques and / or algorithms, the training of the summarizing model 320 can utilize segmented description tuples 314 as training data pairs. Millions of training data pairs can be stored (e.g., in machine learning data structures of system 100) for training the summarizing model 320.

[0030] Additionally or alternatively, training the summarizing model 320 may include fine-tuning the summarizing model 320 to a distilled summarizing model such as distil-BART 12-6, such that the summarizing model 320 is configured to generate short, abstract topic descriptions that approximate the topic descriptions generated by GLM 112.

[0031] return Figure 1 After the summarization model 120 has been trained by the model training engine 118 as described herein, the model 120 is provided to the topic tagging engine 104 portion of the system. The topic tagging engine 104 is then configured to receive and / or otherwise acquire communication transcripts 122, provide segments of those transcripts 122 to the summarization model 120 to generate associated topic descriptions 124, map those topic descriptions 124 to topic tags 130 of the topic directory 126, and provide those topic tags as topic tagging results 132 in response to the communication transcripts 122.

[0032] In some examples, the topic tagging engine 104 is configured to segment the communication transcript 122 into communication segments in a manner substantially similar to that of the transcription segmentation engine 108 described herein. Additionally or alternatively, the transcription segmentation engine 108 of system 100 may be used in both the summarization model trainer 102 and the topic tagging engine 104. Alternatively, each of the summarization model trainer 102 and the topic tagging engine 104 may be configured to include the transcription segmentation engine 108 for segmenting transcripts into segments, as described herein.

[0033] Furthermore, the topic tagging engine 104 is configured to provide segments of the communication transcription 122 to the summarization model 120, and / or otherwise apply the summarization model 120 to the segments of the transcription 122. As a result, the summarization model 120 generates topic descriptions 124 and assigns them to the processed transcription segments.

[0034] After generating topic descriptions 124, they are used in conjunction with a topic catalog 126 to identify topic tags 130 best suited for the segments of the communication transcription 122. In some examples, the topic catalog 126 includes supporting examples 128 associated with topic tags 130. Each topic tag 130 may be associated with a set of supporting examples 128, which are similar to the topic descriptions 124 that can be generated from the summarization model 120. Alternatively or additionally, supporting examples 128 may include words, sentences, or other phrases that may indicate the associated topic tags 130. Figure 4 A more detailed description of the subject catalog.

[0035] Figure 4 This is a schematic diagram 400 illustrating the mapping from topic descriptions 424 to topic tags 430A in a topic directory 426 according to an embodiment. In some examples, the process shown in schematic diagram 400 is in a system (such as...) Figure 1 The topic description 424 is used in conjunction with the topic directory 426 to determine whether the topic tag 430A “quote” is the best match for the topic description 426, such that “quote” is provided as the topic tag result 432.

[0036] In some examples, topic description 424 is compared with some or all of the supporting examples in topic catalog 426 (e.g., supporting example 428A mapped to topic tag 430A or otherwise associated with topic tag 430A, and other supporting examples mapped to other topic tags within the topic catalog). Based on this comparison, the system (e.g., topic tagging engine 104) can identify one or more supporting examples that are most closely compared to topic description 424 and determine the topic tag associated with the identified supporting example as topic tag 432 for topic description 424. Because topic description 424 is also associated with a communication segment (e.g., communication segment 110) as described herein, the determined topic tag 432 can be assigned to or otherwise associated with that communication segment. When some or all of the communication segments in a communication transcript (e.g., communication transcript 122) are assigned topic tags as topic tag result 432, topic tag result 432 can be provided in association with the communication transcript, enabling consumers of result 432 to more quickly determine and / or interpret the content of the communication transcript.

[0037] Additionally or alternatively, topic tag results 432 can be used in conjunction with associated communication transcripts to enable classification and other data processing to be applied to communication transcripts (e.g., in a database where transcripts associated with topic tag results 432 can be searched based on those topic tag result associations and otherwise analyzed).

[0038] Using the topic catalog 426 to map topic descriptions 424 from the summarizing model to specific topic tag results 432 offers several advantages. Some models, such as summarizing models, may face "illusions" and / or they may be vulnerable to adversarial attacks. By including the additional step of mapping descriptions to topic tags, the system is able to filter out or otherwise avoid the effects of problems such as those encountered by summarizing models.

[0039] The comparison of topic description 424 with the supporting examples in topic directory 426 can be performed in several ways. In some examples, a "maximum similarity" procedure can be used, in which topic description 424 is compared with each supporting example to obtain a value indicating the similarity between the two entries, and after calculating this similarity value for each supporting example, the topic tag associated with the supporting example that has the highest similarity value in topic description 424 is selected.

[0040] Alternatively or additionally, an "average embedding" process can be used, in which the average embeddings of the set of supporting examples for each topic tag (e.g., vectorized data of letters / words / sentences of the supporting examples) are averaged and compared with the embeddings of the topic description to calculate an associated similarity value. The set of supporting examples with the average embeddings most similar to the embeddings of the topic description indicates which topic tag will be selected for association with the topic description.

[0041] Furthermore, the selection of topic tags associated with a topic description can be based on a defined threshold. For example, a similarity value between a topic description and (e.g., 80%) of the supporting examples(s) associated with the topic tag could be required to be higher than a defined threshold for the topic tag selected for that topic description. In some examples, establishing such a threshold might result in some topic descriptions not being associated with the topic tag. Such topic descriptions and associated communication segments could be associated with the "no topic" or "unknown topic" tags. Alternatively, such topic descriptions could be associated with the best-matching topic tag, but they could also be flagged to indicate that the match with the topic tag is more likely to be inaccurate. Other methods can be used to handle topic descriptions that do not strongly match the topic tag without deviating from the description.

[0042] In some examples, comparing the topic description with supporting examples is based on vectors generated from the words and / or phrases of the topic description and supporting examples. This use of vectors allows for faster and / or more accurate identification of similarity between descriptions and examples compared to other methods that directly compare letters, words, or phrases. For example, the average embedding process described above can first perform a word embedding process on the topic description and supporting examples to form real-valued vectors of their words, which encode the meaning of the words such that words that are closer in the vector space are expected to be similar in meaning. In other examples, other types of vector-based comparison methods can be used without departing from this specification.

[0043] Furthermore, in other examples, more or different comparison methods can be used to map topic descriptions to topic tags without departing from this specification. For example, instead of using a maximum similarity process or an average embedding process, the system can be configured to perform both processes and then select topic tags based on the results of these two processes, where some logic is used to make a decision between the different results from the two processes. Alternatively or additionally, other comparison methods can be used.

[0044] return Figure 1 In some examples, topic catalog 126 may be defined to include a default set of topics and associated support examples, and / or to include a custom set of topics and associated support examples defined by, for example, a customer using system 100. For example, if a user of a version of topic tagging engine 104 is primarily interested in generating topic tagging results 132 from sales calls, that user can define topic catalog 126, which includes topic tags 130 based on sales calls and associated support examples 128 describing those topic tags 130.

[0045] Furthermore, such users can submit a collection of past sales call transcripts to system 100 to be used as training communication transcripts 106, thereby enabling the summarization model 120 to be specifically trained on the user-submitted sales call communications. As a result, a topic tagging engine 104 using the summarization model 120 and a customized topic directory 126 is fine-tuned for analyzing the sales call transcripts that the user wishes to analyze.

[0046] In some examples, the definition of supporting example 128 and / or associated topic tag 130 can be done automatically or otherwise based on automatically generated recommendations that the user can view. For example, the user can define the topic tag 130 to use, and the topic tag engine 104 or system 100 can more generally be configured to automatically obtain recommended supporting examples based on those defined topic tags 130. The user can view the recommended supporting examples and accept or reject them, making the resulting supporting example 128 customizable by the user. Alternatively, the recommended supporting examples can be automatically included in the topic directory 126 for mapping descriptions to topic tags 130.

[0047] Additionally or alternatively, system 100 may be configured to prompt a user, for whom topic tagging engine 104 is configured to view some or all of the topic descriptions 116 used to train summary model 120. While viewing topic descriptions 116, the user may be prompted to categorize those topic descriptions 116 into one or more topic tags in topic tags 130 within topic directory 126. As a result of the user's categorization, some or all of the categorized topic descriptions 116 may be included in topic directory 126 as supporting examples 128, which are associated with topic tags 130 based on the user's categorization.

[0048] Figure 5 This is a flowchart illustrating a computerized method 500 for training a summarization model (e.g., summarization model 120) to generate topic descriptions (e.g., topic description 124) from communication transcripts (e.g., communication transcripts 122) according to an embodiment. In some examples, method 500 is used in systems (such as those described herein). Figure 1 The process is executed on system 100. At 502, a first communication transcript associated with the first communication is obtained. In some examples, the first communication transcript is the training communication transcript 106 as described above. The first communication transcript may be obtained from a user or client within a communication transcript set representing the communication to be analyzed to obtain associated topic tags.

[0049] At 504, the first communication transcript is divided into a first communication segment set. In some examples, the first communication transcript is divided into segments using a transcription segmentation engine 108, such as the summarization model trainer 102 described herein.

[0050] At 506, based on the first set of communication segments, a first set of topic descriptions is generated by analyzing each communication segment in the first set of communication segments using a GLM. In some examples, the generation of the first set of topic descriptions is based on a GLM 112 of the summarizing model trainer 102 (e.g., the GLM is GPT-3, etc.). Furthermore, generating topic descriptions from the GLM involves providing communication segments to the GLM and then using plain language prompts to prompt the GLM to provide topic descriptions for the provided communication segments.

[0051] At 508, the summarizing model is trained using the first communication segment set and the associated first topic description set as training data. In some examples, the summarizing model is trained by the model training engine 118 of the summarizing model trainer 102 using segment description tuples 114, as described herein.

[0052] In step 510, the trained summarization model is applied to the second communication transcript, and in step 512, a second set of topic descriptions for the second communication transcript is generated based on the applied trained summarization model. In some examples, the trained summarization model is applied to the second communication transcript, and a second set of topic descriptions is generated in the topic tagging engine 104, as described herein. Furthermore, in some examples, the second communication transcript is divided into communication segments, and the trained summarization model is applied to those communication segments. Then, for each communication segment, a topic description is generated by the summarization model.

[0053] In some examples, a topic directory (e.g., topic directory 126) is defined, which includes example topic descriptions (e.g., supporting example 128) associated with topic tags (e.g., topic tag 130). A second set of topic descriptions is compared with the example topic descriptions in the topic directory, and based on this comparison, the second set of topic descriptions is mapped to a set of topic tags in the topic directory. A set of topic tags is provided in response to a second communication transcription.

[0054] In addition, in some examples, the comparison between the topic description and the example topic description is based on maximum similarity comparison or average embedding comparison, as described in this paper.

[0055] Additionally or alternatively, the comparison of the topic description with the example topic description includes comparing a vectorized version of the topic description with a vectorized version of the example topic description, as described herein.

[0056] In some examples, the topic directory can be defined based on a default set of topic tags and topic descriptions or a custom set of topic tags and topic descriptions. Furthermore, recommended topic descriptions can be generated based on topic tags, as described in this article. Approval of recommended topic descriptions allows those approved topic descriptions to be included in the topic directory.

[0057] Figure 6 This is a flowchart illustrating a computerized method 600, according to an embodiment, for providing segmented topic tags (e.g., topic tagging results 132) of communication transcripts (e.g., communication transcript 122) using a summarization model. In some examples, method 600 is implemented in systems (such as...) Figure 1 Executed on system 100). In 602, a topic directory (e.g., topic directory 126) is defined using example topic descriptions (e.g., support example 128) and topic tags (e.g., topic tag 130).

[0058] At 604, the communication transcripts associated with the communication are obtained (e.g., communication transcript 122). At 604, the communication transcripts are partitioned into a set of communication segments.

[0059] At 606, a summary model (e.g., summary model 120) is applied to the communication segment set. Based on the application of the summary model, at 610, a set of topic descriptions (e.g., topic description 124) is generated from the communication segments.

[0060] In section 612, the set of topic descriptions is compared with example topic descriptions from the topic directory, and based on this comparison, the set of topic descriptions is mapped to a set of topic tags. In some examples, the comparison and mapping are further based on a defined similarity threshold. For example, for each pair of topic descriptions from the set of topic descriptions and example topic descriptions from the topic directory, a similarity value is calculated for that pair. Those similarity values ​​are compared with a defined similarity threshold, and mapping associated topic tags to topic descriptions requires that the associated similarity values ​​exceed the defined similarity threshold.

[0061] In 616, in response to the acquired communication transcripts, a set of topic tags is provided. In some examples, providing the set of topic tags includes displaying a GUI that includes a timeline of the communication transcripts associated with the displayed transcripts, and includes topic tags for the displayed set of topic tags, which are associated with the displayed timeline (e.g., ...). Figure 7 (GUI).

[0062] Figure 7 This is a schematic diagram illustrating a GUI 700 that displays topic tags 730A, 730B, 730C, and 730D associated with segments of communication, according to an embodiment. In some examples, the GUI 700 is displayed as a system (such as...) Figure 1 It is part of the operation of System 100. GUI 700 includes visualization of sessions or similar communications between two parties (first party and second party).

[0063] The participation of the first party is displayed as timeline 702, with bars along the timeline indicating the duration of the first party's speech. The participation of the second party is displayed as timeline 704, with bars along the timeline indicating the duration of the second party's speech. The entire communication is further represented as timeline 706, which allows the user to interact with the playback of the communication.

[0064] In some examples, when playing back a recording of a communication, indicator 708 indicates the current point in the communication playback (e.g., as shown, playback at 1:01 of the communication). Playback of the recording can be controlled using interface 710, which includes a "Play" button, a "Fast Forward" button, a "Rewind" button, a "Track End" button, and a "Track Start" button. The "Play" button can be configured to play the communication from the current point at normal speed. The "Fast Forward" button allows indicator 708 to advance rapidly through the communication. The "Rewind" button moves indicator 708 backward through the communication. The "Track End" button moves indicator 708 to the end of the communication. The "Track Start" button moves indicator 708 to the beginning of the communication. In other examples, more, fewer, or different types of interfaces can be displayed in GUI 700 without departing from this specification. For example, it may include buttons that allow indicator 708 to jump to the beginning of the next topic tag's segment and / or to jump indicator 708 to the beginning of the previous topic tag's segment.

[0065] Exemplary operating environment

[0066] This disclosure may be used as a basis for... Figure 8 The computing device 818 is operated according to an embodiment of the functional block diagram 800. In an embodiment, components of the computing device 818 may be implemented as part of an electronic device according to one or more embodiments described herein. The computing device 818 includes one or more processors 819, which may be a microprocessor, a controller, or any other suitable type of processor for processing computer-executable instructions to control the operation of the electronic device. Alternatively or additionally, the processor 819 may be any technology capable of executing logic or instructions, such as a hard-coded machine. Platform software including an operating system 820 or any other suitable platform software may be provided on the device 818 to enable application software 821 to execute on the device. According to embodiments, the training and use of the summarization model as described herein to provide topic tags for communication transcription can be implemented through software, hardware, and / or firmware.

[0067] Computer-executable instructions may be provided using any computer-readable medium accessible by computing device 818. Computer-readable media may include, for example, computer storage media such as memory 822 and communication media. Computer storage media such as memory 822 include volatile and non-volatile, removable and non-removable media implemented using any method or technique for storing information such as computer-readable instructions, data structures, program modules, etc. Computer storage media include, but are not limited to, RAM, ROM, EPROM, EEPROM, permanent memory, phase-change memory, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical storage, magnetic tape cassettes, magnetic tape, disk storage, shingled disk storage or other magnetic storage devices, or any other non-transmission medium that may be used to store information accessible to the computing device. Conversely, communication media may embody computer-readable instructions, data structures, program modules, etc., in modulated data signals or other transmission mechanisms such as carrier waves. As defined herein, computer storage media does not include communication media. Therefore, computer storage media should not be construed as the propagation of signals themselves. The propagation of signals themselves is not an example of computer storage media. Although the computer storage medium (memory 822) is shown as being within the computing device 818, those skilled in the art will understand that the storage may be distributed or remotely located and may be accessed via a network or other communication link (e.g., using communication interface 823).

[0068] The computing device 818 may include an input / output controller 824 configured to output information to one or more output devices 825, such as a display or speaker, which may be decoupled from or integrated into the electronic device. The input / output controller 824 may also be configured to receive and process input from one or more input devices 826 (e.g., a keyboard, microphone, or touchpad). In one embodiment, the output device 825 may also act as an input device. An example of such a device could be a touch-sensitive display. The input / output controller 824 may also output data to devices other than the output devices, such as locally connected printing devices. In some embodiments, a user may provide input to (or receive output from) the input devices 826 and / or receive output from the output devices 825.

[0069] The functions described herein can be performed at least in part by one or more hardware logic components. According to an embodiment, computing device 818, when executed by processor 819, is configured by program code to perform embodiments of the described operations and functions. Alternatively or additionally, the functions described herein can be performed at least in part by one or more hardware logic components. Illustrative types of hardware logic components that can be used, such as but not limited to, include field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), program-specific standard products (ASSPs), system-on-a-chip (SoCs), complex programmable logic devices (CPLDs), and graphics processing units (GPUs).

[0070] At least some of the functions of the various elements in the figure can be performed by other elements in the figure or entities not shown in the figure (e.g., processors, network services, servers, applications, computing devices, etc.).

[0071] Although described in conjunction with exemplary computing system environments, the examples of this disclosure can be implemented with many other general-purpose or special-purpose computing system environments, configurations, or devices.

[0072] Examples of well-known computing systems, environments, and / or configurations suitable for use with any aspect of this disclosure include, but are not limited to, mobile or portable computing devices (e.g., smartphones), personal computers, server computers, handheld (e.g., tablets) or laptop devices, multiprocessor systems, game consoles or controllers, microprocessor-based systems, set-top boxes, programmable consumer electronics, mobile phones, wearable or accessory form factors of mobile computing and / or communication devices (e.g., watches, glasses, headsets or headphones), network PCs, minicomputers, mainframes, distributed computing environments including any of the systems or devices described above, and so on. In general, this disclosure can operate with any device having processing capabilities, enabling it to execute instructions such as those described herein. Such systems or devices can accept input from a user in any manner, including input devices such as keyboards or pointing devices, input via gestures, proximity input (such as by hover), and / or input via voice.

[0073] Examples of this disclosure can be described in the general context of computer-executable instructions, such as program modules, that are executed by one or more computers or other devices as software, firmware, hardware, or a combination thereof. Computer-executable instructions can be organized into one or more computer-executable components or modules. Generally, program modules include, but are not limited to, routines, programs, objects, components, and data structures that perform a particular task or implement a particular abstract data type. Aspects of this disclosure can be implemented with any number and organization of such components or modules. For example, aspects of this disclosure are not limited to the specific computer-executable instructions or specific components or modules shown in the accompanying drawings and described herein. Other examples of this disclosure may include different computer-executable instructions or components having more or fewer functions than those shown and described herein.

[0074] In examples involving general-purpose computers, aspects of this disclosure, when configured to execute the instructions described herein, transform a general-purpose computer into a special-purpose computing device.

[0075] An example system for determining the topics of a communication transcript includes: at least one processor; and at least one memory including computer program code, the at least one memory and the computer program code being configured together with the at least one processor such that the at least one processor: acquires a first communication transcript associated with a first communication; divides the first communication transcript into a first communication segment set; generates a first topic description set based on the first communication segment set by analyzing each communication segment in the first communication segment set using a generative language model; trains a summarization model using the first communication segment set and the associated first topic description set as training data; applies the trained summarization model to a second communication transcript; and generates a second topic description set of the second communication transcript based on the application of the trained summarization model to the second communication transcript.

[0076] An example computerized method for determining the topics of a communication transcript includes: a processor acquiring a first communication transcript associated with a first communication; the processor dividing the first communication transcript into a first communication segment set; the processor generating a first topic description set based on the first communication segment set by analyzing each communication segment in the first communication segment set using a generative language model; the processor training a summarizing model using the first communication segment set and the associated first topic description set as training data; the processor applying the trained summarizing model to a second communication transcript; and the processor generating a second topic description set of the second communication transcript based on the application of the trained summarizing model to the second communication transcript.

[0077] One or more computer storage media having computer-executable instructions for determining the topics of communication transcripts, the computer-executable instructions, when executed by a processor, causing the processor to at least: define a topic directory including example topic descriptions associated with topic tags; acquire communication transcripts associated with the communication; partition the communication transcripts into a set of communication segments; apply a summary model to the set of communication segments; generate a set of topic descriptions based on applying the summary model to the set of communication segments, wherein each segment in the set of communication segments is associated with a topic description in the set of topic descriptions; compare the set of topic descriptions with example topic descriptions in the topic directory; map the set of topic descriptions to a set of topic tags in the topic directory based on the comparison of the set of topic descriptions with example topic descriptions in the topic directory; and provide a set of topic tags in response to acquiring the communication transcripts.

[0078] Alternatively, or in addition to the other examples described herein, examples include any combination of the following:

[0079] - The application of the trained summarization model to the second communication transcription includes: the processor dividing the second communication transcription into a second communication segment set; the processor applying the summarization model to the second communication segment set; and the processor generating a second topic description set based on the application of the summarization model to the second communication segment set, wherein each segment in the second communication segment set is associated with a topic description in the second topic description set.

[0080] - Further includes: defining a topic directory by the processor, the topic directory including example topic descriptions associated with topic tags; comparing a second set of topic descriptions with the example topic descriptions in the topic directory by the processor; mapping the second set of topic descriptions to a set of topic tags in the topic directory by the processor based on the comparison of the second set of topic descriptions with the example topic descriptions in the topic directory; and providing the set of topic tags by the processor in response to receiving a second communication transcription.

[0081] - The comparison of the second set of topic descriptions with the example topic descriptions includes at least one of the following: comparison based on maximum similarity and comparison based on average embedding.

[0082] - The comparison between the second set of topic descriptions and the example topic descriptions includes: comparing the vectorized version of the second set of topic descriptions with the vectorized version of the example topic descriptions.

[0083] - The definition of the topic directory includes: generating recommended example topic descriptions based on the topic tags in the topic directory; and including approved example topic descriptions in the topic directory based on received approval of the recommended example topic descriptions.

[0084] - The first topic description set is generated by analyzing each communication segment in the first communication segment set using a generative language model based on the first communication segment set, including: providing communication segments to the generative language model; and using ordinary language prompts to prompt the generative language model to provide topic descriptions of the provided communication segments.

[0085] - The comparison of the set of topic descriptions with the example topic descriptions includes at least one of the following: comparison based on maximum similarity and comparison based on average embedding.

[0086] - The comparison between the set of topic descriptions and the example topic descriptions includes: comparing the vectorized version of the set of topic descriptions with the vectorized version of the example topic descriptions.

[0087] - The definition of the topic directory includes: generating recommended example topic descriptions based on the topic tags in the topic directory; and including approved example topic descriptions in the topic directory based on received approval of the recommended example topic descriptions.

[0088] The comparison of the set of topic descriptions with the example topic descriptions in the topic directory includes: for each pair of topic descriptions from the set of topic descriptions and example topic descriptions from the topic directory, calculating a similarity value for the pair; comparing the calculated similarity value with a similarity threshold; and mapping the pair of topic descriptions to topic tags associated with the pair of example topic descriptions based on the calculated similarity value exceeding the similarity threshold.

[0089] - Providing a set of topic tags in response to acquiring communication transcripts includes: displaying a graphical user interface (GUI) including a timeline of communication associated with the communication transcripts and topic tags including the set of topic tags displayed, the displayed topic tags being associated with the displayed timeline.

[0090] Any range or device value given herein may be extended or changed without loss of the desired effect, as will be apparent to those skilled in the art.

[0091] While no personally identifiable information is tracked in any aspect of this disclosure, examples have been described with reference to data monitored and / or collected from users. In some examples, users of the data collection may be notified (e.g., via a dialog box or preferences) and given the opportunity to give or refuse consent to the monitoring and / or collection. Consent may take the form of opting in or opting out of consent.

[0092] Although the subject matter has been described using language specific to structural features and / or methodological actions, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are disclosed as exemplary forms of implementing the claims.

[0093] It should be understood that the above benefits and advantages may relate to one embodiment or several embodiments. Embodiments are not limited to those that solve any or all of the stated problems or have any or all of the stated benefits and advantages. It will be further understood that a reference to "a" means one or more of those items.

[0094] The embodiments shown and described herein, as well as embodiments not specifically described herein but within the scope of the claims, constitute exemplary components for a processor to acquire a first communication transcript associated with a first communication; exemplary components for a processor to divide the first communication transcript into a first communication segment set; exemplary components for a processor to generate a first topic description set based on the first communication segment set by analyzing each communication segment in the first communication segment set using a generative language model; exemplary components for a processor to train a summarization model using the first communication segment set and the associated first topic description set as training data; exemplary components for a processor to apply the trained summarization model to a second communication transcript; and exemplary components for a processor to generate a second topic description set of the second communication transcript based on applying the trained summarization model to the second communication transcript.

[0095] The term "comprising" is used in this specification to mean including (multiple) features or (multiple) actions that follow, without excluding the presence of one or more additional features or actions.

[0096] In some examples, the operations illustrated in the figures can be implemented as software instructions encoded on a computer-readable medium, implemented in hardware programmed or designed to perform the operations, or both. For example, aspects of this disclosure can be implemented as a system on a chip or other circuit system comprising multiple interconnected conductive elements.

[0097] Unless otherwise specified, the order in which the operations in the examples of this disclosure are performed or executed is not essential. That is, unless otherwise specified, operations may be performed in any order, and the examples of this disclosure may include more or fewer operations than those disclosed herein. For example, it is contemplated that performing or executing a particular operation before, simultaneously with, or after another operation is within the scope of various aspects of this disclosure.

[0098] When describing elements of this disclosure or its examples, the articles “a,” “an,” “the,” and “the” are intended to mean the presence of one or more of these elements. The terms “comprising,” “including,” and “having” are intended to be inclusive and mean that additional elements may be present besides those listed. The term “exemplary” is intended to mean “an example of…”. The phrase “one or more of the following: A, B, and C” means “at least one of A, and / or at least one of B, and / or at least one of C.”

[0099] Having described in detail various aspects of this disclosure, it will be apparent that modifications and variations are possible without departing from the scope of the aspects of this disclosure as defined in the appended claims. Since various changes can be made to the above-described constructions, products, and methods without departing from the scope of the aspects of this disclosure, all that is contained in the above description and shown in the accompanying drawings is intended to be illustrative rather than restrictive.

Claims

1. A system comprising: processor; as well as A memory including computer program code, the memory and the computer program code being configured, together with the processor, such that the processor: Obtain the first communication transcript associated with the first communication; The first communication transcription is divided into a first communication segment set; Analyze each communication segment in the first set of communication segments using a generative language model; Based on the analysis, a first topic description set based on the first communication segment set is generated; The first communication segment set and the associated first topic description set are used as training data to train the summary model; The trained summarization model is then applied to the second communication transcription; A second set of topic descriptions of the second communication transcript is generated by applying the trained summarization model to the second communication transcript. Define a topic directory, which includes sample topic descriptions associated with sample topic tags; Map the second set of topic descriptions to the example topic descriptions in the topic directory; Use the example topic tags associated with the example topic description to generate recommended topic tags; as well as The generated recommended topic tags are provided on a graphical user interface (GUI) as a timeline displayed on the GUI, the timeline including the generated recommended topic tags for the second communication transcription.

2. The system of claim 1, wherein applying the trained summarization model to the second communication transcription comprises: The second communication transcription is divided into a second communication segment set; The trained summary model is then applied to the second set of communication segments; as well as The second topic description set is generated by applying the trained summarization model to the second set of communication segments, wherein each segment in the second set of communication segments is associated with a topic description in the second set of topic descriptions.

3. The system of claim 1, wherein mapping the second set of topic descriptions to the example topic descriptions includes at least one of the following: comparison based on maximum similarity and comparison based on average embedding.

4. The system of claim 1, wherein mapping the second set of topic descriptions to the example topic descriptions comprises: Compare the vectorized version of the second set of topic descriptions with the vectorized version of the example topic description.

5. The system of claim 1, wherein defining the topic directory includes: The processor generates recommended example topic descriptions based on the topic tags in the topic directory; as well as Acceptance of the recommended example topic descriptions will be granted, and the approved example topic descriptions will be included in the topic directory.

6. The system of claim 1, wherein generating the first topic description set based on the first communication segment set by analyzing each communication segment in the first communication segment set using the generative language model comprises: Provide communication segments to the generative language model; as well as The generative language model is prompted with common language cues to provide a topic description of the provided communication segment.

7. A computerized method, comprising: The processor obtains the first communication transcript associated with the first communication; The processor divides the first communication transcription into a first communication segment set; The processor analyzes each communication segment in the first communication segment set using a generative language model; Based on the analysis, the processor generates a first topic description set based on the first communication segment set; The processor uses the first set of communication segments and the associated first set of topic descriptions as training data to train the summary model; The processor applies the trained summarization model to the second communication transcription; The processor generates a second set of topic descriptions of the second communication transcript by applying the trained summary model to the second communication transcript. The processor defines a topic directory, which includes example topic descriptions associated with example topic tags; The processor maps the second set of topic descriptions to the example topic descriptions in the topic directory; The processor uses the example topic tags associated with the example topic description to generate recommended topic tags; as well as The generated recommended topic tags are provided on a graphical user interface (GUI) as a timeline displayed on the GUI, the timeline including the generated recommended topic tags for the second communication transcription.

8. The computerized method of claim 7, wherein applying the trained summarization model to the second communication transcription comprises: The processor divides the second communication transcription into a second communication segment set; The processor applies the trained summary model to the second set of communication segments; as well as The processor generates a second set of topic descriptions by applying the trained summary model to the second set of communication segments, wherein each segment in the second set of communication segments is associated with a topic description in the second set of topic descriptions.

9. The computerized method of claim 7, wherein mapping the second set of topic descriptions to the example topic descriptions includes at least one of the following: comparison based on maximum similarity and comparison based on average embedding.

10. The computerized method of claim 7, wherein mapping the second set of topic descriptions to the example topic descriptions comprises: Compare the vectorized version of the second set of topic descriptions with the vectorized version of the example topic description.

11. The computerized method of claim 7, wherein defining the subject catalog comprises: Based on the topic tags in the topic directory, generate recommended example topic descriptions; as well as Acceptance of the recommended example topic descriptions will be granted, and the approved example topic descriptions will be included in the topic directory.

12. The computerized method of claim 7, wherein generating the first topic description set based on the first communication segment set by analyzing each communication segment in the first communication segment set using the generative language model comprises: Provide communication segments to the generative language model; as well as The generative language model is prompted with common language cues to provide a topic description of the provided communication segment.

13. One or more computer storage media having computer-executable instructions, which, when executed by a processor, cause the processor to at least: Obtain the first communication transcript associated with the first communication; The first communication transcription is divided into a first communication segment set; Analyze each communication segment in the first set of communication segments using a generative language model; Based on the analysis, a first topic description set based on the first communication segment set is generated; The first communication segment set and the associated first topic description set are used as training data to train the summary model; The trained summarization model is then applied to the second communication transcription; The trained summarization model is used to generate a second set of topic descriptions of the second communication transcript. Define a topic directory, which includes sample topic descriptions associated with sample topic tags; Map the second set of topic descriptions to the example topic descriptions in the topic directory; Use the example topic tags associated with the example topic description to generate recommended topic tags; as well as The generated recommended topic tags are provided on a graphical user interface (GUI) as a timeline displayed on the GUI, the timeline including the generated recommended topic tags for the second communication transcription.

14. One or more computer storage media according to claim 13, wherein comparing the set of topic descriptions with the example topic descriptions includes at least one of the following: comparison based on maximum similarity and comparison based on average embedding.

15. One or more computer storage media according to claim 13, wherein comparing the set of topic descriptions with the example topic descriptions comprises: Compare the vectorized version of the set of topic descriptions with the vectorized version of the example topic description.

16. One or more computer storage media according to claim 15, wherein defining the subject directory includes: Based on the topic tags in the topic directory, generate recommended example topic descriptions; as well as Acceptance of the recommended example topic description is granted, and the approved example topic description is included in the topic directory.

17. One or more computer storage media according to claim 13, wherein comparing the set of topic descriptions with the example topic descriptions of the topic directory includes: For each pair of topic descriptions from the topic description set and example topic descriptions from the topic directory, a similarity value is calculated for the pair; The calculated similarity value is compared with a similarity threshold; as well as The topic description of the pair is mapped to a topic tag associated with the example topic description of the pair based on the calculated similarity value that exceeds the similarity threshold.

Images