SYSTEM AND METHOD OF UTILIZING A HYBRID SEMANTIC MODEL FOR SPEECH RECOGNITION

Abstract

A system includes a network interface, a speech input conversion component, and a routing module. Speech input is received in connection with a call. At least a segment of the speech input is transformed into a first textual format. A first list of entries is generated based, at least partially, on consideration of the first textual format. The first list includes at least one action with a corresponding confidence level and at least one object with another corresponding confidence level. An entry of the first list having a higher corresponding confidence level is selected, and a second textual format is output. A second list is generated based, at least partially, on consideration of the selected entry and the second textual format. A routing option is suggested based on the selected entry and a pairing entry in the second list.

Description

CLAIM OF PRIORITY

This application is a continuation application of, and claims priority to, U.S. patent application Ser. No. 11/036,204, filed Jan. 14, 2005, the contents of which are expressly incorporated herein by reference in their entirety.

FIELD OF THE DISCLOSURE

The present disclosure relates generally to speech recognition and, more particularly, to a system and method of utilizing a hybrid semantic model for speech recognition.

BACKGROUND

Many speech recognition systems utilize specialized computers that are configured to process human speech and carry out some task based on the speech. Some of these systems support “natural language” type interactions between users and automated call routing (ACR) systems. Natural language call routing allows callers to state the purpose of the call “in their own words.”

A goal of a typical ACR application is to accurately determine why a customer is calling and to quickly route the customer to an appropriate agent or destination for servicing. Research has shown that callers prefer speech recognition systems to keypad entry or touchtone menu driven systems.

As suggested above, natural language ACR systems attempt to interpret the intent of the customer based on the spoken language. When a speech recognition system partially misinterprets the caller's intent significant problems can result. A caller who is misrouted is generally an unhappy customer. Misrouted callers often terminate the call or hang-up when they realize that there has been a mistake. If a caller does not hang up they will typically talk to an operator who tries to route the call. Routing a caller to an undesired location and then to a human operator leads to considerable inefficiencies for a business. Most call routing systems handle a huge volume of calls and, even if a small percentage of calls are mishandled, the costs associated with the mishandled calls can be significant.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 illustrates a simplified configuration of a telecommunication system;

FIG. 2 is a general diagram that illustrates a method of routing calls;

FIG. 3 is a flow diagram that illustrates a method of processing and routing calls;

FIG. 4 is a table that depicts speech input and mapped synonym terms; and

FIG. 5 is a table illustrating action-object pairs and call destinations relating to the action-object pairs.

DETAILED DESCRIPTION OF THE DRAWINGS

The present disclosure is directed generally to integrating speech enabled automated call routing with action-object technology. Traditional automatic call routing systems assign a correct destination for a call 50% to 80% of the time. Particular embodiments of the disclosed system and method using action-object tables may achieve a correct destination assignment 85 to 95% of the time. In some embodiments, a semantic model may be used to create an action-object pair that further increases call routing accuracy while reducing costs. In particular implementations, the correct call destination routing rate may approach the theoretical limit of 100%. Due to higher effective call placement rates, the number of abandoned calls (e.g., caller hang-ups prior to completing their task) may be significantly reduced, thereby reducing operating costs and enhancing customer satisfaction.

In accordance with the teachings of the present disclosure, a call may be routed based on a selectable action-object pair. In practice, a call is received from a caller and a received speech input is converted into text or “text configurations,” which may be the same as, similar to, or can be associated with, known actions and objects. Generally, objects are related to nouns and actions are related to verbs. The converted text may be compared to tables of known text configurations representing objects and actions. A confidence level may be assigned to the recognized actions and objects based on text similarities and other rules. An action-object list may be created that contains recognized actions and objects and their confidence levels. In some embodiments, the entry (action or object) in the list with the highest confidence level may be selected as a dominant item. If an action is dominant a system incorporating teachings disclosed herein may look for a complementary object. Likewise, if an object is dominant, the system may look for a complementary action.

In some implementations, when an action is dominant, remaining actions may be masked and the confidence level of the complementary objects in the action-object list may be adjusted. Conversely, if an object is dominant, the remaining objects may be masked and the confidence level of complementary actions in the action-object list may be adjusted. An adjustment to an assigned confidence level may be based, for example, on the likelihood that the prospective complement in the action-object list is consistent with the dominant entry. Depending upon implementation details, a call may be routed based on a dominant action and a complementary object or a dominant object and a complementary action.

Referring now to FIG. 1, an illustrated communications system 100 that includes a call routing support system is shown. Communications system 100 includes a speech-enabled call routing system (SECRS) 118, such as an interactive voice response system having a speech recognition module. Communications system 100 also includes a plurality of potential call destinations. Illustrative call destinations shown include service departments, such as billing department 120, balance information 122, technical support 124, employee directory 126, and new customer service departments 128. In practice, communication network 116 may receive calls from a variety of callers, such as the illustrated callers 110, 112, and 114. In a particular embodiment, communication network 116 may be a public telephone network, a wireless telephone network, a voice over Internet protocol (VoIP) type network, or some other network capable of supporting communication. As depicted, SECRS 118 may include components, such as a processor 142, memory 143, a synonym table 144, and a routing module 140. Depending upon implementation details, SECRS 118 may be coupled to and may route calls to various destinations across a LAN, an Intranet, an extranet, the Public Internet, and/or some other communication link or network, as shown. In addition, SECRS 118 may route calls to an agent, such as the illustrated live operator 130.

An illustrative embodiment of SECRS 118 may be a call center having a plurality of agent terminals attached. Thus, while only a single operator 130 is shown in FIG. 1, it should be understood that a plurality of different agent terminals or types of terminals may be coupled to SECRS 118, such that a variety of agents may service incoming calls. Moreover, and as indicated above, SECRS 118 may be operable as an automated call routing system.

In a particular embodiment, action-object routing module 140 includes an action-object lookup table for matching action-object pairs to desired call routing destinations. This process may be better understood through consideration of FIG. 2. Referring to FIG. 2, an illustrative block diagram of SECRS 118 is depicted. In this particular embodiment, processor 142 in SECR 118 includes an acoustic processing model 210, semantic processing model 220, and action-object routing table 230. In a first conversion, acoustic model 210 may receive speech input 202 and provide text as its output 204. Semantic model 220 may receive text 204 directly or indirectly from acoustic model 210 and produce an action-object table. The action(s) and object(s) in the action-object table may be ordered or ranked according to a confidence level. The confidence level may be used to indicate how likely a given action or object reflects a correct and useable customer instruction.

When a speech input conversion creates a dominant action (e.g., an action has the highest confidence level in the action-object list), a system like SECRS 118 of FIG. 1 may initiate a secondary conversion that creates an object list from the initial speech input. The call may then be routed based on several criteria, such as the overall highest confidence level in the action-object list (a dominant list entry) and the highest confidence level complimentary term from the secondary conversion (a complement to the dominant entry).

In practice, the secondary conversion or a second list can be generated that may take the initial speech received from the caller and processes the initial speech a second time. During the second conversion the semantic model 220 may look specifically for consistent objects while ignoring actions if an action had the highest overall confidence level. In such a case, the high scoring action may have been selected, the actions may have been masked, and objects that are inconsistent with the selected action may be tagged as invalid. Examples of invalid action-object combinations can be understood by referring to FIG. 5, where objects are listed on the left of the chart, and actions are listed across the top of the chart. For example, if the action of “acquire” has the highest confidence level in the action-object list then during the secondary conversion, objects such as “bill,” “payment,” “other providers,” “coupon specials” “name/number” and “store locations” may be masked or tagged as invalid selections.

If the speech input conversion creates a dominant object, a secondary conversion may be initiated to create an action list to assist in selecting a complementary action. The secondary conversion may take the initial speech received from the caller and processes the initial speech a second time. It may also rely on an output from the processing performed in connection with the earlier conversion. During the second conversion, semantic model 220 may look specifically for actions while ignoring objects. The confidence levels of actions may also be adjusted based on actions that are inconsistent with the selected object. Thus, in either case a call may be routed based on a dominant entry and a valid complement to the dominant entry.

The results of a reiterative speech recognition process may be provided to action-object routing table 230. Routing table 230 may receive action-object pairs 206 and produce a call routing destination 208. Based on the call routing destination 208, a call received at a call routing network like SECRS 118 may be routed to a final destination, such as the billing department 120 or the technical support service destination 124 depicted in FIG. 1. In a particular embodiment, the action-object routing table 230 may be a look up table or a spreadsheet, such as a Microsoft Excel™ spreadsheet.

Referring to FIG. 3, an illustrative embodiment of a method of processing a call using an automated call routing system such as the system of FIG. 1 is illustrated. The method starts at 300 and proceeds to step 302 where a speech input signal, such as a received utterance, is received or detected. Using phonemes or some other effective techniques, the received speech input may be converted into a plurality of word strings or text in accordance with an acoustic model, as shown at steps 304 and 306. In a particular embodiment, probability values may be assigned to word strings based on established rules and the content and coherency of the word string. At step 308, the word strings may be parsed into objects and actions. Objects generally represent nouns and adjective-noun combinations, while actions generally represent verbs and adverb-verb combinations. The actions and objects are assigned confidence values or probability values based on how likely they are to reflect the intent of the caller. In a particular embodiment a probability value or confidence level for the detected action and the detected object is determined utilizing a priority value of the word string used to create the selected action and the selected object.

In some cases, many possible actions and objects may be detected or created from the word strings. A method incorporating teachings of the present disclosure may attempt to determine and select a most probable action and object from a list of preferred objects and actions. To aid in this resolution, a synonym table such as the synonym table of FIG. 4 may be utilized to convert detected actions and objects into actions and objects that the system expects and/or is configured to “listen for.” Thus, detected objects and actions may be converted to expected actions and objects and assigned a confidence level. The process may also utilize the synonym table, for example, to adjust confidence levels of the actions and objects. The synonym table may store natural language phrases and their relationship with a set of actions and objects. In practice, natural language spoken by the caller may be compared to the natural language phrases in the table. Using the synonym table, the system and method may map portions of the natural phrases to detected objects and maps portions of the natural spoken phrase to detected actions. Thus, the word strings can be converted into expected objects and actions, at step 308. In summary, at step 310 multiple actions and multiple objects can be detected and provided with a confidence level according to the likelihood that a particular action or object identifies a customer's intent and thus will lead to a successful routing of the call.

The confidence level may be assigned to an action and/or an object based on many criteria, such as the textual similarities, business rules, etc., in step 310. Confidence levels may also be assigned based on a combination of factors, and some of these factors may not involved speech recognition. For example, in a particular example, if a caller does not currently have service, a caller's number (caller ID) may be utilized to assign a high confidence level to the action “acquire” and a low confidence value the actions “change” or “cancel.” In the event that a confidence level for an action-object pair is below a predetermined level, the call may be routed to a human operator or agent terminal.

An action-object list may be utilized at step 312 to select a dominant entry. If an action is selected as the dominant entry at step 334, other actions in the action-object list may be masked and objects that are inconsistent with the selected action may be tagged as invalid at step 336. The process of invalidating objects based on a dominant action can be further explained by referring to FIG. 5 where objects are listed on the left side of the chart and actions are listed across the top of the chart. For example if the action of “cancel” has the highest confidence level in the action-object list, the objects described as “bill,” “payment,” “other providers,” “coupon specials” “name/number” and “store locations” may be masked or tagged as invalid selections because a caller would not likely want to, for example, “cancel-store locations.” Thus, the method may ignore objects and invalid actions when a dominant object has been selected. The entries at the intersection of valid action-object illustrate routing destinations or phone extension where a call is routed when the system determines a dominant entry and it's complement.

Based on a dominant action, the confidence level of the objects can be adjusted at step 338. The caller's input of the utterance may be sent through the acoustic model, again in step 340, and the acoustic model may create and store word strings, as shown in step 342. Word strings may be parsed into objects using the semantic model in step 344, and an object list may be formed where each object in the list is assigned a confidence level in step 346. When a list is sufficiently complete, the object having the highest confidence level may be selected to complement the dominant action and an action-object pair may be created at step 330.

If at step 312 it is determined that an object has the highest confidence level or is dominant then a search for a complementary action may be conducted. Objects remaining in the action-object list and action that are inconsistent with the selected object may be masked or tagged as invalid, as shown in step 316. Thus such a method may ignore objects and invalid actions in the search for a complementary action when a dominant object has been elected.

Based on the dominant object, the confidence level of listed actions may be adjusted at step 318. The original caller input may be sent through the acoustic model, again in step 320 and the acoustic model may create and store word strings as in step 322. Words strings may then be parsed into objects using the semantic model in step 324 and an actions list may be formed where actions in the list is assigned a confidence level at step 326. The action having the highest confidence level (at step 328) may be selected to complement the dominant object and an action-object pair may be passed at step 330. The call may then be routed at step 331, the process ending at 332.

In practice, it may be beneficial to convert word strings such as “I want to have” to an action such as “get.” This substantially reduces the size of the action and object tables. As shown in FIG. 4, differently expressed or “differently spoken” inputs that have the same or similar caller intent may be converted to a single detected action-object, and/or action-object pair. Further, improper and informal sentences as well as slang may be connected to an action-object pair that may not bear phonetic resemblance to the words uttered by the caller. With a mapped lookup table such as the table in FIG. 4, speech training and learning behaviors found in conventional call routing systems may not be required. The tables in the present disclosure may be updated easily, leading to a lower cost of system maintenance.

The above-disclosed subject matter is to be considered illustrative, and not restrictive, and the appended claims are intended to cover all such modifications, enhancements, and other embodiments that fall within the true spirit and scope of the present invention. Thus, to the maximum extent allowed by law, the scope of the present invention is to be determined by the broadest permissible interpretation of the following claims and their equivalents, and shall not be restricted or limited by the foregoing detailed description.

Claims

1. A system, comprising:

a network interface configured to receive a speech input in connection with a call;

a speech input conversion component configured to: transform at least a segment of the speech input into a first textual format; generate a first list of entries based, at least partially, on consideration of the first textual format, the first list comprising at least one action having a corresponding confidence level and at least one object having another corresponding confidence level; select an entry of the first list having a higher corresponding confidence level; output a second textual format; generate a second list based, at least partially on consideration of the selected entry and the second textual format; and

a routing module configured to suggest a routing option for the call based on the selected entry and an associated pairing entry in the second list.

2. The system of claim 1, wherein the speech input conversion component is further configured to re-process the speech input to create an object list when an action is the selected entry.

3. The system of claim 1, wherein the speech input conversion component is further configured to re-process the speech input to create an action list when an object is the selected entry.

4. The system of claim 2, wherein the speech input conversion component is further configured to:

include an associated confidence level with an object entry in the object list; and

select an object as the pairing entry based on the associated confidence level.

5. The system of claim 3, wherein the speech input conversion component is further configured to:

re-process the speech input to produce the action list with confidence levels; and

select an action based on the confidence levels in the action list.

6. The system of claim 1, wherein the speech input conversion component is further configured to compare the first textual format to a list of word strings and to assign a probability to at least one word string included in the list of word strings.

7. The system of claim 6, wherein the speech input conversion component is further configured to assign an appropriate confidence level to the at least one word string.

8. The system of claim 1, wherein the entry selected is one of a verb and an adverb-verb combination.

9. The system of claim 1, wherein the entry selected is one of a noun or an adjective-noun combination.

10. The system of claim 1, wherein the speech input conversion component is further configured to utilize a synonym table to assist in converting the speech input into action and objects.

11. A method, comprising:

receiving a speech input in connection with a call;

processing the speech input to generate a first action list and an object list;

assigning a first confidence level to each action of the first action list and to each object of the object list;

selecting a particular object with a high confidence level from the object list; and

removing at least one action from the first action list, wherein the at least one action is inconsistent with the particular object.

12. The method of claim 11, further comprising assigning a second confidence level to each remaining action of the first action list based on the particular object.

13. The method of claim 12, further comprising:

re-processing the speech input to generate a second action list;

assigning a second confidence level to each action of the second action list;

selecting a particular action with a high confidence level from the second action list; and

suggesting a routing option for the call based on the particular action and the particular object.

14. The method of claim 13, further comprising routing the call to a destination.

15. The method of claim 13, wherein the first confidence level and the second confidence level are assigned based on a predetermined likelihood of reflecting an intent of a caller.

16. A method, comprising:

receiving a speech input in connection with a call;

processing the speech input to generate a first object list and an action list;

assigning a first confidence level to each object of the first object list and to each action of the action list;

selecting a particular action with a high confidence level from the action list; and

removing at least one object from the first object list, wherein the at least one object is inconsistent with the particular action.

17. The method of claim 16, further comprising assigning a second confidence level to each remaining object of the first object list based on the particular action.

18. The method of claim 17, further comprising:

re-processing the speech input to generate a second object list;

assigning a second confidence level to each object of the second object list; and

selecting a particular object with a high confidence level from the second object list;

suggesting a routing option for the call based on the particular action and the particular object.

19. The method of claim 18, further comprising routing the call to a destination.

20. The method of claim 16, wherein the first confidence level and the second confidence level are assigned based on a predetermined likelihood of reflecting an intent of a caller.