Method for recognizing multiple intentions of user text in man-machine interaction

Abstract

The invention provides a user text multi-intention recognition method in man-machine interaction, and relates to the technical field of natural language processing, a multi-intention segmented text in a cache is compared with a user text, and if the comparison succeeds, a plurality of corresponding single intentions are obtained according to the successfully compared multi-intention segmented text; and if the comparison fails, performing multi-intention rule segmentation on the user text, if the segmentation succeeds, obtaining a plurality of corresponding single intentions, and if the segmentation fails, performing intention recognition by using the TextCNN model, and obtaining the plurality of single intentions in the user text, so that if the plurality of single intentions corresponding to the user text are obtained before intention recognition is performed by using the TextCNN model, the user text is subjected to intention recognition by using the TextCNN model, and the user text is subjected to intention recognition by using the TextCNN model. Compared with intention recognition by directly using a TextCNN model, the method has the advantages that the speed is higher, the accuracy is higher, the problem that existing multi-intention recognition is low in accuracy is solved, and the method is suitable for multi-intention recognition.

Description

Technical Field

[0001] The present invention relates to the technical field of natural language processing, and in particular to a method for identifying multiple intentions of user text in human-computer interaction. Background Art

[0002] Human-computer interaction appears more and more frequently in people's daily work and life, bringing great convenience to people's daily life. For human-computer interaction, whether the machine can accurately understand the user's intention is very critical. Improving the accuracy of intent recognition can help improve the user's interactive experience.

[0003] Currently, smart home appliances on the market, such as TVs and air conditioners, can often only recognize a single user command, but lose the semantics of another part. In real scenarios, a user's sentence may contain multiple intentions, such as turning on the air conditioner to 23 degrees cooling mode. Therefore, it is particularly important to accurately identify multiple intentions from the user's sentence to enhance the user's interactive experience.

[0004] Existing multi-intent recognition, such as the invention patent with Chinese patent application No. 202210739460.0, discloses a multi-intent recognition method and system based on the BERT language model and the TextCNN model. The intent is recognized through the TextCNN model, but the recognition accuracy is poor, and the intent recognition speed for commonly used multi-intent texts is slow. Summary of the invention

[0005] The technical problem solved by the present invention is to provide a method for multi-intent recognition of user text in human-computer interaction to solve the problem of low accuracy of existing multi-intent recognition.

[0006] The present invention solves the above technical problems by adopting a technical solution: a method for identifying multiple intentions of user text in human-computer interaction, comprising the following steps:

[0007] S1. Combining single intents into multi-intent combination data according to combination rules;

[0008] S2, segmenting the multi-intent combination data into single intents according to the combination rule, obtaining multi-intent segmentation texts, and storing them in a cache;

[0009] S3. Compare the user text with the multi-intent segmented text in the cache. If the comparison is successful, obtain the corresponding multiple single intents based on the successfully compared multi-intent segmented text. If the comparison fails, segment the user text according to the multi-intent rules. If the segmentation is successful, obtain the corresponding multiple single intents. If the segmentation fails, use the TextCNN model to perform intent recognition to obtain multiple single intents in the user text.

[0010] Furthermore, the combination rule is an arbitrary combination and the single intentions in the combination do not conflict.

[0011] Furthermore, the single intents in the multi-intent segmentation text are separated by separators.

[0012] Furthermore, the method also includes recording the number of successful comparisons of the multi-intent segmented texts in the cache, and performing the comparisons in sequence according to the number of successful comparisons at the next comparison.

[0013] Furthermore, the multi-intention rule segmentation includes the following steps:

[0014] S31, construct a word segmentation dictionary with part-of-speech tags, and replace entities in the user text with part-of-speech tags based on the entity dictionary to obtain the replaced user text;

[0015] S32. Use regular expressions to segment the replaced user text.

[0016] Furthermore, the entities include film and television titles, song titles, app names, channel names and actors.

[0017] Furthermore, the regular expression includes prefix+entity, entity+suffix and prefix+entity+suffix, the prefix includes open, download, uninstall, delete, exit, query, close, delete and install, and the suffix includes next, a little bit, bigger, a little higher, a little higher, a little lower, a little lower, a little smaller and a little smaller.

[0018] Beneficial effects of the present invention: The present invention provides a method for multi-intent recognition of user text in human-computer interaction, by comparing the multi-intent segmented text in the cache with the user text, if the comparison is successful, the corresponding multiple single intents are obtained based on the successfully compared multi-intent segmented text; if the comparison fails, the user text is segmented according to the multi-intent rules, if the segmentation is successful, the corresponding multiple single intents are obtained, if the segmentation fails, the TextCNN model is used to perform intent recognition to obtain multiple single intents in the user text, so that before using the TextCNN model for intent recognition, multiple single intents corresponding to the user text are obtained, compared with directly using the TextCNN model for intent recognition, the speed is faster and the accuracy is higher, which solves the problem of low accuracy of existing multi-intent recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a flow chart of a method for identifying multiple intentions of user text in human-computer interaction provided by the present invention. DETAILED DESCRIPTION

[0020] The present invention aims to solve the problem of low accuracy of existing multi-intention recognition and provides a method for multi-intention recognition of user text in human-computer interaction. Figure 1 As shown, the following steps are included:

[0021] S1. Combining single intents into multi-intent combination data according to combination rules;

[0022] Specifically, the combination rule is any combination and the single intents in the combination do not conflict. For example, the two single intents of watching a movie and listening to music cannot be combined.

[0023] S2, segmenting the multi-intent combination data into single intents according to the combination rule, obtaining multi-intent segmentation texts, and storing them in a cache;

[0024] Specifically, the single intents in the multi-intent segmentation text are separated by separators. For example, the multi-intent "someone's movie with louder sound" is segmented into single intents, and the multi-intent segmentation text obtained is "someone's movie | louder sound".

[0025] S3. Compare the user text with the multi-intent segmented text in the cache. If the comparison is successful, obtain the corresponding multiple single intents based on the successfully compared multi-intent segmented text. If the comparison fails, segment the user text according to the multi-intent rules. If the segmentation is successful, obtain the corresponding multiple single intents. If the segmentation fails, use the TextCNN model to perform intent recognition to obtain multiple single intents in the user text.

[0026] Specifically, for multi-intent segmented texts in the cache, the number of successful comparisons of the multi-intent segmented texts in the cache can also be recorded. During the next comparison, the comparisons are performed in sequence according to the number of successful comparisons. In this way, for commonly used user texts, multiple single intents can be quickly obtained with an accuracy rate of 100%.

[0027] For user texts that fail the comparison, multi-intention rule segmentation is performed, and the multi-intention rule segmentation includes the following steps:

[0028] S31. Construct a word segmentation dictionary with part-of-speech tags, and replace entities in the user text with part-of-speech tags based on the entity dictionary to obtain the replaced user text; specifically, the entities include movie and TV names, song names, app names, channel names, actors and sounds, etc.

[0029] S32. Use regular expressions to segment the replaced user text. Specifically, the regular expressions include prefix+entity, entity+suffix and prefix+entity+suffix. The prefixes include open, play, download, uninstall, delete, exit, query, close, delete and install. The suffixes include next, a little bit, bigger, higher, higher, lower, lower, smaller and smaller, etc.

[0030] The judgment criteria for successful multi-intent rule segmentation are: the regular expression can completely cover the user text, and there are at least two segmentation texts, each of which corresponds to a single intent.

[0031] For example: if the user text is "Open QQ Music and play someone's Infernal Affairs with the volume turned up", after entity replacement, it becomes "Open @appname and play @figure's @video@sound turned up", and the regular expression is used to segment it into "Open @appname, play @figure's @video, @sound turned up" to obtain three single intents, that is, the regular expression can completely cover the user text, and there are at least two segmented texts, each of which corresponds to a single intent, where @appname is the part-of-speech tag of QQ Music, @figure is the part-of-speech tag of someone, @video is the part-of-speech tag of Infernal Affairs, and @sound is the part-of-speech tag of sound.

[0032] The user text is "someone's movie sound is louder", and after entity replacement, it becomes "@figure's @video@sound is louder". Using regular expressions for segmentation, there is only one "entity + suffix", that is, the regular expression can completely cover the user text, but there is only one cut text. Therefore, segmentation using regular expressions fails, and the TextCNN model is needed for intent recognition.

[0033] The user text is "I want to watch a movie starring someone in 2024 with the volume turned up by 20 decibels". After entity replacement, it becomes "I want to watch a movie starring @figure in 2024 with @sound turned up by 20 decibels". Regular expressions are used for segmentation, but there are no prefixes and suffixes for regular expressions. Therefore, segmentation using regular expressions fails, that is, regular expressions cannot completely cover user text. The TextCNN model is needed for intent recognition, where @figure is the part-of-speech tag for the actor and @sound is the part-of-speech tag for the sound.

[0034] The user text is "Turn on the air conditioner", and after entity replacement, it becomes "Turn on @ac". When segmented using regular expressions, there is only one "prefix + entity + suffix". Therefore, segmentation using regular expressions fails. That is, the regular expression can completely cover the user text, but there is only one segmented text. The TextCNN model is needed for intent recognition, where @ac is the part-of-speech tag for air conditioner.

[0035] Therefore, for user texts that fail cache comparison, multi-intent rules are used to segment them. If the segmentation is successful, corresponding multiple single intents are obtained with an accuracy of 100%. If the segmentation fails, the TextCNN model is used for intent recognition with an accuracy of the TextCNN model. Overall, the overall accuracy of this technical solution is higher than the accuracy of intent recognition using only the TextCNN model.

[0036] In particular, when in use, the user can set specific multi-intent combination data according to his or her own habits, and increase the word frequency, which corresponds to the number of successful cache comparisons, so that multi-intent recognition can be quickly completed for specific user texts without the need for multi-intent rule segmentation and TextCNN models.

Claims

1. A method for identifying multiple intentions of user text in human-computer interaction, characterized in that: The following steps are involved: S1. Combining single intents into multi-intent combination data according to combination rules; S2, segmenting the multi-intent combination data into single intents according to the combination rule, obtaining multi-intent segmentation texts, and storing them in a cache; S3, comparing the user text with the multi-intent segmented text in the cache, and if the comparison is successful, obtaining the corresponding multiple single intents based on the successfully compared multi-intent segmented text; If the comparison fails, the user text is segmented according to the multi-intent rules. If the segmentation is successful, the corresponding multiple single intents are obtained. If the segmentation fails, the TextCNN model is used to perform intent recognition to obtain multiple single intents in the user text.

2. The method for identifying multiple intentions of user text in human-computer interaction according to claim 1, characterized in that: The combination rule is an arbitrary combination and the single intentions in the combination do not conflict.

3. The method for identifying multiple intentions of user text in human-computer interaction according to claim 1, characterized in that: The single intents in the multi-intent segmentation text are separated by separators.

4. The method for identifying multiple intentions of user text in human-computer interaction according to claim 1, characterized in that: It also includes recording the number of successful comparisons of the multi-intent segmented texts in the cache, and performing the comparisons in sequence according to the number of successful comparisons during the next comparison.

5. The method for identifying multiple intentions of user text in human-computer interaction according to any one of claims 1 to 4, characterized in that: The multi-intention rule segmentation includes the following steps: S31, construct a word segmentation dictionary with part-of-speech tags, and replace entities in the user text with part-of-speech tags based on the entity dictionary to obtain the replaced user text; S32. Use regular expressions to segment the replaced user text.

6. The method for identifying multiple intentions of user text in human-computer interaction according to claim 5, characterized in that: The entities include film and television titles, song titles, app names, channel names, and actors.

7. The method for identifying multiple intentions of user text in human-computer interaction according to claim 5, characterized in that: The regular expression includes prefix+entity, entity+suffix and prefix+entity+suffix, the prefix includes open, download, uninstall, delete, exit, query, close, delete and install, and the suffix includes next, a little bit, bigger, a little higher, a little higher, a little lower, a little lower, a little smaller and a little smaller.

Images