EP2569716A1 - Groupage sémantique - Google Patents

Groupage sémantique

Info

Publication number
EP2569716A1
EP2569716A1 EP10717910A EP10717910A EP2569716A1 EP 2569716 A1 EP2569716 A1 EP 2569716A1 EP 10717910 A EP10717910 A EP 10717910A EP 10717910 A EP10717910 A EP 10717910A EP 2569716 A1 EP2569716 A1 EP 2569716A1
Authority
EP
European Patent Office
Prior art keywords
semantic
clusters
graph
cluster
graphs
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
EP10717910A
Other languages
German (de)
English (en)
Inventor
Jean-Marie Henri Daniel Larcheveque
Elizabeth Ireland Powers
Freya Kate Recksiek
Dan Teodosiu
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
VirtuOz Inc
Original Assignee
VirtuOz Inc
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by VirtuOz Inc filed Critical VirtuOz Inc
Publication of EP2569716A1 publication Critical patent/EP2569716A1/fr
Ceased legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G09—EDUCATION; CRYPTOGRAPHY; DISPLAY; ADVERTISING; SEALS
    • G09B—EDUCATIONAL OR DEMONSTRATION APPLIANCES; APPLIANCES FOR TEACHING, OR COMMUNICATING WITH, THE BLIND, DEAF OR MUTE; MODELS; PLANETARIA; GLOBES; MAPS; DIAGRAMS
    • G09B19/00—Teaching not covered by other main groups of this subclass
    • G09B19/04—Speaking
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/35—Clustering; Classification
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/36—Creation of semantic tools, e.g. ontology or thesauri
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/30—Semantic analysis
    • G—PHYSICS
    • G09—EDUCATION; CRYPTOGRAPHY; DISPLAY; ADVERTISING; SEALS
    • G09B—EDUCATIONAL OR DEMONSTRATION APPLIANCES; APPLIANCES FOR TEACHING, OR COMMUNICATING WITH, THE BLIND, DEAF OR MUTE; MODELS; PLANETARIA; GLOBES; MAPS; DIAGRAMS
    • G09B19/00—Teaching not covered by other main groups of this subclass
    • G09B19/06—Foreign languages

Definitions

  • a company may provide a website that includes details about products and/or services of the company. Additionally, the website may include support information, or functionality to purchase products and services from the company. A customer, for instance, may interact with the website to find information about a prospective purchase and later, after the purchase, to find information regarding use of the purchase. Consequently, the amount of information that is made available via these techniques is ever- increasing, which may make it difficult for customers to locate desired information using traditional techniques.
  • search technologies For example, the company may include search technologies on a website to allow customers to hunt for answers to their questions. This may work well for certain types of queries and issues, but may fail as questions become increasingly complex, as issue resolution may require personalized information, and so on. As a result, users may "walk away" from the website frustrated, may make a time-consuming call to a human customer service representative (CS ), and so on. Therefore, traditional search techniques may have a negative impact on user experience with the website and consequently on the user's view of the company as a whole.
  • CS human customer service representative
  • Semantic clustering techniques are described.
  • a conversational agent is configured to accept natural language input from a user ("utterances") and then perform deep linguistic analysis of these utterances.
  • Semantic clustering may be applied to the output of such analysis to provide a variety of functionalities, such as grouping a corpus of utterances into semantic clusters in which each cluster pertains to a similar topic. These semantic clusters may then be leveraged to identify topics and assess their relative importance in order to, for example, prioritize topics that occur frequently or topics whose handling by the conversational agent should be improved.
  • a variety of utterances may be processed using these techniques, such as spoken words or textual descriptions entered via instant messaging, a website interface, SMS, email, a social networking, blogging or micro-blogging service, and so on.
  • FIG. 1 is an illustration of an environment in an example implementation that is operable to perform semantic clustering techniques for a conversational agent.
  • FIG. 2 illustrates an example implementation of a procedure to match a user utterance to likely user intents.
  • FIG. 3 is an illustration of an example semantic graph that is generated by a conversational agent of FIG. 1 for a user utterance.
  • FIG. 4 depicts an example implementation of a system for performing linguistic analysis of a user utterance.
  • FIG. 5 depicts an example implementation of an ontology using a language- independent hierarchy.
  • FIG. 6 depicts an example implementation of a semantic graph pattern that includes graph fragments.
  • FIG. 7 is an illustration of an example implementation showing examples of semantic graphs corresponding to utterances that match the two fragment graphs making up the graph pattern of FIG. 6.
  • FIG. 8 is an illustration of an example implementation of subsuming graphs.
  • FIG. 9 illustrates an example of visualization of a matching percentile of a pattern using a meter.
  • FIG. 10 illustrates an example implementation of information output via a user interface for review by a reviewer.
  • Conversational agent techniques are described, which include semantic clustering and other functionalities that are described in the following sections.
  • conversational agents are implemented using one or more modules to engage in an interactive natural language dialog with a user via a textual chat.
  • use of conversational agents may provide automated assistance to users to help them resolve issues without directly interacting with a human agent (e.g., a customer support representative in a call center). This may help a company to efficiently utilize resources and provide additional functionality to a user that was not available via traditional search techniques.
  • the textual chat may be inputted using a variety of mechanisms, such as transcripts of spoken words (such as telephone calls), text inputs (e.g., instant messages, live chat, email, SMS, blogging and micro-blogging services, and so on), automatic speech recognition, and so forth.
  • the conversational agent may map user inputs (henceforth called “utterances”) to semantic representations.
  • Such representations may be graphs, the nodes of which represent concepts and the edges of which represent semantic roles.
  • Such graphs will henceforth be called “semantic graphs”.
  • the conversational agent may represent a user intent by an intent graph pattern or a plurality of intent graph patterns.
  • a user utterance may be formed into a semantic graph and compared with intent graph patterns (henceforth called "graph patterns"). If there is a match then the utterance likely involves the intent represented by the graph pattern or plurality of graph patterns.
  • Semantic graphs representing utterances may be grouped into semantic clusters using various techniques referred to as semantic clustering techniques. Each of the semantic clusters may pertain to a similar topic. These semantic clusters can then be leveraged to determine dominant or active topics in a corpus of utterances for a variety of purposes. For example, semantic clusters may be used to improve the quality of a conversational agent. Further discussion of semantic clustering techniques and conversational agents may be found in relation to the following sections. [0021] In the following discussion, an example environment is described along with example procedures that may be implemented in the example environment as well as in other environments. Accordingly, the example procedures are not limited to implementation in the example environments and the example environments are not limited to implementation of the example procedures.
  • FIG. 1 illustrates an example environment 100 that is operable to employ semantic clustering techniques for a conversational agent.
  • the illustrated environment 100 includes a service provider 102 having a conversational agent 104 that is accessible to a plurality of client devices 106, 108 over a network 1 10.
  • the client devices 106, 108 may be configured in a variety of ways.
  • the client devices may be configured as a computing device that is capable of communicating over the network, such as a desktop computer as illustrated by client device 106, a mobile station, an entertainment appliance, a set-top box communicatively coupled to a display device, a wireless phone as illustrated by client device 108, a game console, and so forth.
  • the client devices may range from full resource devices with substantial memory and processor resources (e.g., personal computers, game consoles) to a low-resource device with limited memory and/or processing resources (e.g., traditional set-top boxes, hand-held game consoles).
  • the network 1 10 may assume a variety of configurations.
  • the network 1 10 may include a wide area network (WAN), a local area network (LAN), a wireless network, a public telephone network, an intranet, a telephone network, and so on.
  • WAN wide area network
  • LAN local area network
  • the client device 106 configured as a desktop computer and the service provider 102 may be communicatively coupled via the Internet and the client device 108 configured as a wireless phone may be communicatively coupled to the service provider 102 via a telephone network.
  • the client device 106 configured as a desktop computer and the service provider 102 may be communicatively coupled via the Internet and the client device 108 configured as a wireless phone may be communicatively coupled to the service provider 102 via a telephone network.
  • a telephone network A wide variety of other instances are also contemplated.
  • the service provider 102 is illustrated as being implemented by one or more servers (or other computing devices) that are accessible to the client devices 106, 108 via the network 1 10.
  • the conversational agent 104 is illustrated as a module that is implemented by the service provider 102.
  • the conversational agent 104 may include a user experience 1 12 that is accessible via a webpage output by the service provider 102 to the client device 106 configured as a desktop computer.
  • the conversational agent 104 may include a user experience 1 12 that is accessible via a spoken input received by the client device 108 configured as a wireless phone.
  • user experience of the conversational agent 104 may be accessed through a wide variety of techniques. A variety of other examples are also contemplated, such as instant messaging, email, user-generated content in conjunction with a social network, blogging and micro-blogging services, and so on.
  • any of the functions described herein can be implemented using software, firmware, hardware (e.g., fixed logic circuitry), manual processing, or a combination of these implementations.
  • the terms "module” and “functionality” as used herein generally represent software, firmware, hardware, or a combination thereof.
  • the module and/or functionality represents instructions (e.g., program code) that perform specified tasks when executed on a processing system that may include one or more processors or other hardware.
  • the program code can be stored in a wide variety of types and combinations of memory may be employed, such as random access memory (RAM), hard disk memory, removable medium memory, and other types of computer-readable media.
  • the conversational agent 104 is configured to engage in an interactive natural language dialog with a human user via textual chat, to complete a specific task for or on behalf of that user. For example, text entered by a user through interaction with the client device 106 configured as a desktop computer may be provided to the conversational agent 104. In another example, a voice input provided by the client device 108 configured as a wireless phone may be converted to text and processed by the conversational agent 104; the response of the conversational agent 104 can then be converted back to speech before being sent to the client device 108.
  • Tasks may include providing information to the user, answering the user's questions, helping the user solve a problem (support agent), proposing new products and services to the user (sales and/or marketing agent), and so on.
  • the conversational agent 104 may embed complex logic flows to aid interaction with the user using natural language.
  • the conversational agent 104 may also interact with various application programming interfaces (APIs) and backend systems of a company that offers use of the agent, e.g., the service provider 102.
  • APIs application programming interfaces
  • the conversational agent 104 may be offered as a visual avatar on a company web site (or a specific section of the site), but other interaction channels such as instant messaging systems, mobile phones, email, social networking sites, or blogging and micro -b logging services are also contemplated.
  • the conversational agent 104 may respond to user questions and also drive the conversation to solicit specific information to better understand the user's situation.
  • Example Conversation Strategy
  • Utterances that are received are parsed by a linguistic analysis module 1 16 of the conversational agent 104 and may be matched by a comparison module 1 18 against a number of possible intents that are part of one or more decision trees 120. Based on the identified intent, the conversational agent 104 may then generate a reply. A conversation between the user and the agent may include one or more of these interactions between the user and the conversational agent 104.
  • a user's intent can be expressed in a variety of ways.
  • the user's intent may be configured as a single information request, may include a set of several potential information requests, and so on.
  • the conversational agent 104 may ask for clarification until a specific information request is identifiable and may be satisfied.
  • conversations are modeled as paths through a set of decision trees 120, which may be configured as circuit-like structures that describe possible conversation flows.
  • the root of each decision tree 120 may describe an initial state, before a user intent has been identified. Leaves of the decision tree 120 may be thought of as answers to a specific request. Accordingly, a path from the root to a leaf of the decision tree 120 may represent a sequence of utterances (e.g., speech acts) that may lead to identification of the information requested by the user and thus completion of the conversational agent's 104 task.
  • the conversational agent 120 may offer increasingly complex dialog strategies that allow the user to switch between tasks (or decision trees) flexibly.
  • the set of intents that can be matched to user utterances at a particular point in time relates to a current position of a conversation in the decision tree 120.
  • a customer of a telecommunications company might initiate a conversation by asking, "Can I access my voice mail from the web?"
  • the conversational agent 104 moves from the decision tree's 120 root node to one of the root's child nodes.
  • the conversational agent 104 may consult the information that is relevant to proceed in the decision tree 120 and respond with a clarifying question, e.g., "What type of phone service do you use for your voice mail?"
  • the conversational agent 104 Assuming the user answers with an utterance that includes sufficient information and is recognized by the agent, the conversational agent 104 has identified the user's intent and moves to a leaf node, which contains an answer to the user's question. It should be noted that a user utterance such as "VOIP" may be associated with a different intent when produced at the beginning of the conversation, at the root of the decision tree 120, as opposed to later in the conversation at the node corresponding to web access to voicemail. [0034] In addition to the current position in the decision tree 120, the conversational agent 104 may have knowledge of pieces of information that were obtained earlier during the conversation.
  • this additional information may be represented as variable-value pairs, which may act to limit the user from asking a same question multiple times, asking for information that was already provided by user, and so on.
  • the conversational agent 104 may implement complex and sophisticated conversation strategies. For example, the conversational agent 104 may proactively ask questions and volunteer related pieces of information based on information known about the user from the conversation or other data collected about the user (e.g., via an API provided by the service provider 102 offering the conversational agent 104).
  • FIG. 2 illustrates an example implementation 200 of a procedure to match a user utterance to a likely intent.
  • User utterance 202 is illustrated as being provided to the linguistic analysis module 1 16 via the user experience 1 12.
  • the linguistic analysis module 1 16 may form a semantic graph 204.
  • the semantic graph 204 may then be leveraged by the comparison module 1 18 to determine a closest matching pattern through a decision tree 120 lookup, and from that, a user's intent 206.
  • the conversational agent 104 is configured to deal with "messy" and "real" user utterances 202.
  • the user utterance may contain a high percentage (over 50%) of misspelled, ungrammatical or incomplete sentences, abbreviations, slang, and so on.
  • FIG. 3 is an illustration of an example semantic graph 300 that is generated by the conversational agent 104 of FIG. 1 for a user utterance.
  • Text from a user's utterance e.g., spoken, written, and so on
  • the nodes of the semantic graph 300 represent concepts; the directed edges are labeled with semantic functions in the figure.
  • concepts related to nodes and semantic functions related to edges will be called semantic concepts or traits (and may also be referred to as attributes, properties, features, and so on).
  • semantic concepts or traits and may also be referred to as attributes, properties, features, and so on.
  • the sentence "I'd like to change my order" may be parsed into the semantic graph 300 shown in FIG. 3.
  • the main trait of a node is the concept it represents.
  • concept traits e.g., "modify” in FIG. 3
  • concept traits are abstracted over lexical variations and spelling mistakes. For example, replacing the word “change” by “alter,” “modify,” or even “moddify” (sic) in the user input does not affect the structure of the semantic graph 300 shown in FIG. 3.
  • representing "my” and “I” by the concept Interlocutor makes the interpretation of the semantic form insensitive to the form used in the user utterance. For example, the utterance "You would like to modify your order” may be parsed by the linguistic analysis module 1 16 to form the graph 300 shown in FIG. 3. Similarly, “We would like to change our order” may also be parsed by the linguistic analysis module 1 16 into the graph 300 shown in FIG. 3.
  • constructions such as "would like to” are represented by a modal trait on the modify node and not a concept trait. Additionally, this particular value may be present on one or more of the utterances "I'd like to”, “I want to”, “I warmtha”, and so on. In this way, a single representation may be provided for a variety of synonymous constructions.
  • use of a dedicated modal trait rather than creating a node with a "want-to" concept trait may help to simplify the semantic graphs and thus facilitate pattern matching, further discussion of which may be found later in the "Pattern Matching" section of the description.
  • the graph edges that are drawn in FIG. 3 may be referred to as "function edges" with the respective labels representing semantic roles.
  • Order is the theme of Modify, i.e., the object acted upon. In FIG. 3, this is denoted by the Theme label on the edges between the two nodes, denoting the "theme” semantic role.
  • Theme label may still fulfill this role even if the same relation were expressed in the passive, e.g., "Has my order been modified?" where "order" is a syntactic subject.
  • semantic abstraction may be used to provide a straightforward identification of common ideas in different utterances, which may increase the efficiency and accuracy of a pattern matching process to be discussed later.
  • function edges and their incident nodes form a tree.
  • the root of the tree may be used as a placeholder that does not represent a particular concept of the utterance.
  • the concept trait may be set to a value "Top,” which is representative of the most general concept.
  • parsing may focus on extracting dependencies between words, which may then be mapped to dependencies between concepts.
  • This approach known generically as a dependency grammar, does not make assumptions on phrase structure. Therefore, incomplete sentences and ungrammatical sentences may be handled and mapped to a semantic graph, much in the way a human may extract meaning from ungrammatical or incomplete sentences.
  • This approach allows a conversational agent to be robust and able to understand "real" user utterances, which are often grammatically incorrect, may contain typos and spelling mistakes, and may use slang words or phrases.
  • FIG. 4 depicts an example implementation of a system 400 for parsing (or analyzing) a user utterance.
  • the system 400 of FIG. 4 illustrates a conversational agent 104 as being portioned into a lexical module 402, a syntactic module 404, and a semantic module 406 which are representative of functionality that may be employed at different levels in the parsing procedure.
  • a user utterance 408 e.g., a sentence
  • each word is matched against entries in a lexicon 410 with the aid of a spell checker.
  • a lexical entry may include several words (for example "credit card” or "bill of sale")
  • the lexical module 402 of the conversational agent 104 may map a word sequence of the user utterance 408 to one or more flexion sequences.
  • a flexion is a lexical entry that includes a lemma (i.e., an uninflected form) and possibly grammatical marks, such as tense, number, person, mood, and so on.
  • the lemma "agent” may have the flexions that include “agent” and "agents.”
  • the lexicon 410 that is used to match words to flexions is language-dependent. Additionally, some of the entries contained therein may be specific to a business area, a conversational agent, and so on. For example, lexical entries may include names of forms specific to a business area or commercial names specific to the conversational agent 104. Accordingly, lexicon 410 lookup may be filtered by word spaces, where a word space characterizes a conversational agent or a business area.
  • This information may include (1) construction information and (2) ontology information.
  • Ontology information pertains to the semantic level; and provides the concept traits which are further mentioned in the "Parsing and semantic representations of input sentences" Section and FIG. 3.
  • Construction information includes possible part- of-speech assignments to a lemma. For example, "format" may be assigned both verb and noun constructions. Construction information may also include syntactic patterns linking the dictionary item to other items. For example, the construction for "format" as a verb may show that the item relates to a noun with a subject link and to another noun with an object link.
  • a unification-based algorithm 414 may be employed to unify available constructions of the lemmata (i.e., a plurality of lemma) in a sequence to yield one or more syntactic graphs.
  • linearity information e.g., in English, a tendency of objects to occur after verbs
  • the confidence assigned to the recognition of particular constructions may be taken into account to score the graphs.
  • a highest-scoring syntactic graph is mapped to a semantic graph 416.
  • a semantic graph 416 having increased abstraction is obtained in which nodes represent ontology concepts and edges represent logical relations between the concepts.
  • Ontology may be represented as a language-independent concept hierarchy. This hierarchy may be represented using a directed graph with two types of edges, "is-a-kind-of ' and "subsumes.”
  • a "password” is a kind of “certificate_or_credentials” and is a kind of "secret_or_arcanum”.
  • secret_or_arcanum also subsumes “esoterica” and “kabbalah”
  • certificate_or_credentials subsumes "login_name", "identity_card”, and "diploma”.
  • FIG. 6 depicts an example implementation 600 of a graph pattern 602 that includes graph fragments for (1) "change password” (with "password” functioning as the theme of "change”); and (2) “how”. These fragments form a graph pattern (which, for purposes of the discussion that follows, may be simply referred to as a "pattern").
  • An utterance is considered as matching this pattern if each of the pattern fragments occurs in a semantic graph of the utterance. It should be noted that this is a condition that is considered sufficient for matching; further discussion of this and other conditions may be found in relation to the following section.
  • examples of semantic graphs that match the two graph fragments 602 of FIG. 6 are shown.
  • the conversational agent 104 has been created to explain how to change credentials (i.e., user ID and/or password) rather than a password, specifically. Accordingly, a pattern may be defined to match questions about how to change one's password as well as a user ID or other credentials.
  • a pattern may be defined to match questions about how to change one's password as well as a user ID or other credentials.
  • information to be provided by the conversational agent 104 may be described by a general concept that subsumes a number of more specific concepts that are likely to occur in user utterances.
  • the conversational agent 104 may deliver generic information about connecting an Internet router, but requests for this information are likely to mention specific router brands and models.
  • Concept subsumption may provide flexibility to the conversational agent 104.
  • the conditions that are to be met for a match to be considered between a semantic graph and a pattern are stated as follows: A pattern matches a semantic graph if and only if a subgraph of the semantic graph subsumes the pattern.
  • graph subsumption would be the semantic graph for "change credentials" as subsuming the graph for "change password,” an example of which is shown in the implementation 800 of FIG. 8.
  • Graph 1 subsumes Graph 2 in this implementation 800. More generally, a graph gl subsumes a graph g2 if and only if g2 can be transformed into gl by zero or more applications of the following operations:
  • Trait subsumption has been illustrated in FIG. 8 for concept traits. However, it should be noted that trait subsumption may be defined on a variety of traits, including function labels on edges. Here are other examples:
  • the modal value “MAY or MUST” subsumes the value “MUST”.
  • the generic edge label “Attribute” subsumes each other edge label (for example, “Agent,” “Theme,” “Location,” etc.).
  • modal values are based on sets of possible values. Either a trait takes its value in a hierarchy (e.g., edge labels, ontology concepts) or in a collection of sets. For example, the modal value "MUST" is really a singleton set that includes a single instance of "MUST.”
  • matching In addition to capturing stylistic variations on a question, matching also helps capture logically distinct but equivalent ways of expressing the same intent. For example, a user might ask how she can change her password by typing, "How can I change my password?" or by typing, "Can you help me change my password?” Therefore, a single intent is not usually captured by a single graph pattern. Accordingly, several graph patterns may be used. This set of patterns forms a logical disjunction, meaning that, in order to match the intent, a user utterance matches at least one of the patterns.
  • a set of possible intents may be associated with each position in a conversational agent's 104 decision tree 120. This set is the union of the intents of the child nodes at that position in the decision tree 120.
  • Each of the possible intents at the current decision tree 120 position is represented by a set of graph patterns. The set of patterns collectively representing each of the possible intents at a current position are referred to as the active patterns in the following discussion.
  • the conversational agent 104 may perform the following steps to determine user intent:
  • step 2 If no successful match can be found in step 2 above, we say that the utterance is unmatched. In such a case, the conversational agent may not have the linguistic knowledge to assign an intent to this utterance.
  • a number of metrics may be used to measure a distance between a graph of an utterance and a matching graph pattern in the conversational agent's knowledge. These metrics may combine one or more of the following quantities algebraically:
  • a metric formula is used to compute a distance that decreases with quantity 1 and increases with quantities 2 and 3. These constraints correspond to the semantic "closeness" of the semantic graph and the pattern it matches.
  • the amount of information in trait values may be measured in a number of ways:
  • a node's traits collectively count as a constant (typically equal to 1). Therefore, quantity 1 does not vary for a given pattern and quantity 3 is simply a number of nodes that are not involved in the match.
  • trait values may be considered to have increasing specificity as the number of occurrences in utterances decreases. This may be captured by computing an inverse of a proportion of utterances containing a given trait value. This measure may also be referred to as inverse frequency.
  • the matching distance between two trait values (quantity 2, henceforth called subsumption distance) may be computed as a function of:
  • the number of graph nodes is used as a measure of information and the trait distance is proportional to the number of hierarchy levels linking two concepts or two edge labels. Indirect patterns
  • the conversational agent 104 may also leverage indirect patterns that are assigned low confidence and may be used in cases when the conversational agent 104 is not "sure" of the user's intent. Exclusive, or direct, patterns may take precedence over non-exclusive, or indirect, patterns when identifying a user's intent by the conversational agent 104. If the user's utterance does not match one or more direct patterns, each indirectly matching intent may be considered as potentially relevant. The conversational agent 104 may then offer the user a list of question rewordings or a list of potentially relevant topics or questions. This may occur when a user has entered several keywords but not a full sentence or phrase that more fully describes what is being requested.
  • a user may type "cashback” which might mean “How does cashback work?" or "I never received my cash back.”
  • a designer of the conversational agent 104 may address this situation in a variety of ways, examples of which include the following:
  • the first method may be useful in specific situations for conversational agents where several keywords or ideas are used throughout by the agent in a wide variety of contexts. Therefore, more precise information is to be gathered to differentiate between them.
  • the second method (that relies on indirect patterns) makes it possible to deal with intent ambiguity with minimal demands on designer time.
  • Semantic clustering may be performed to provide a variety of functionality. For example, semantic clustering may be performed to group a corpus of user utterances into zero or more semantic clusters where each cluster corresponds to utterances that pertain to the same or to similar topics. This similarity of utterances inside each cluster may be used for a variety of purposes, further description of which may be found below. Furthermore, this similarity may be used as an input to a human reviewer, to a machine learning algorithm, to an A/B testing algorithm to improve performance of the conversational agent, or to an alerting system which may identify or react to new or existing user issues, among other applications.
  • a designer of the conversational agent 104 may improve the quality of a conversational agent 104 by examining the user utterances that are not matched to a direct intent pattern and/or that get matched to indirect patterns exclusively. Review of these utterances by the designer may indicate whether new intent patterns or new topics are to be added to the agent's content. This review may be performed in a variety of ways, further discussion of which may be found in relation to the following sections.
  • Erroneous intent identification In some instances, user utterances may be matched to a wrong intent. Typically, this occurs because patterns that were previously inserted into the conversational agent's configuration are too general or are simply wrong.
  • Traditional ad-hoc techniques that were developed to address these challenges would rely on reviews or testing of randomly selected utterances or conversations, and detection of manifestations of disinterest, disapproval or hostility from the user. This ad-hoc approach suffers from lack of prioritization and amount of time required. Further, many users ask their question, receive an answer, and simply close the conversational agent, thereby giving no indication whether the answer provided really answered their question.
  • Identification of salient topics in a body of utterances may be employed in a variety of areas. For example, such identification may be conducted for marketing purposes, for identifying strong and weak points of a service, for identifying new topics that are being discussed by users, for analyzing the evolution of topics over time, for creating a new conversational agent using content from pre-existing sources, and so on.
  • the utterances can originate from text-based chats with the conversational agent 104, text-based chats with live people (e.g., live chat interactions or instant messaging), from emails, SMS, blogging or micro-blogging sites, social networking sites, phone transcripts, or any other form of communication that may result in the production of a body of utterances that may be parsed into semantic graphs.
  • One such area that may employ the identification techniques is to identify evolution of a topic profile of a communication channel over time. In particular, topic evolution may be tracked in near-real time and be used to assess the significance of topics as topics come into prominence in order to react to the situation that caused the topics to become prominent.
  • Another such area is to compare two or more distinct utterance corpora with respect to their respective major topics, e.g., to provide identification of common and distinct topics.
  • quantitative information regarding the occurrence frequency or the occurrence time of each topic may also be provided.
  • topic identification offers the possibility of using analytics data to improve a conversational agent.
  • an analytics tool such as a web analytics tool or BI tool
  • may produce various metrics such as measures of conversation success based on the combination of elementary criteria (e.g. whether the user thanked the agent at the end of the conversation, or whether the interaction resulted in a sale), and filter these results by time of the day or time period.
  • elementary criteria e.g. whether the user thanked the agent at the end of the conversation, or whether the interaction resulted in a sale
  • analytics results are generally not actionable for improving a conversation agent, because the results may be aggregated across a multitude of conversations and topics without an indication of content areas that are to be improved.
  • semantic clustering makes it possible to identify topics having a relatively higher priority for pattern-matching improvement. This identification may be performed based on the values of the aggregated metrics and on various conversational agent-specific criteria.
  • Cluster attachment can provide this information in addition to the particular utterance at which the reviewer happens to be looking. In this particular example, cluster attachment may help the reviewer understand at a glance that "I would like to sell my model car" is not a frequent topic for this conversational agent.
  • cluster attachment enables a reviewer to assess the seriousness of failure by the agent to correctly respond to a particular utterance. Without the ability to reliably carry out such assessment, contingent feedback from reviewers and other interested parties are prone to lead to futile or even detrimental improvement efforts. The latter occur when remedying an issue in a minor area may cause a regression in an important area.
  • cluster attachment In addition to providing frequency information about the topic underlying an utterance, cluster attachment also leads to topic identification, which has useful applications for displaying utterances in a UI and allowing an agent to respond helpfully to sentences whose intent has not been identified by other means than topic identification.
  • Clustering in general may involve an array of techniques that use a distance metric in order to subdivide sets of objects into one or more groups of closely related objects.
  • Semantic clustering is an approach that is based on the clustering of semantic graphs that are produced by a linguistic analysis process such as the one previously described in the Matching User Input Utterances to User Intents section, using one or more of the distance metrics described below or elsewhere.
  • the objects that are clustered in this context are the semantic graphs produced by linguistic analysis.
  • the clustering of semantic graphs induces a clustering of the corresponding utterances that generated these graphs; thus use the term "semantic clustering" may refer to both clustering of semantic graphs and clustering of utterances.
  • semantic clustering amounts to grouping utterances by user intent or topic, even though users may describe the same user intent or topic in different utterances using different words and grammatical structures.
  • semantic clustering When applied to a corpus of user utterances that have failed to be matched to a specific direct and/or indirect intent (henceforth called "unmatched utterances"), semantic clustering enables reviewers to:
  • a user intent or topic may contain tens, hundreds, or even thousands of user utterances.
  • cluster sizes i.e., the number of user utterances in each cluster
  • a corpus having a size between 10,000 to 100,000 unmatched user utterances is used for semantic clustering.
  • fewer unmatched utterances may be used for cases where the volume of data is relatively small, and larger numbers of unmatched utterances may be clustered for conversational agents (or other text corpora) that have a sufficient volume of conversations.
  • the number of utterances may also be chosen depending on whether the semantic clustering is to be done on a near real-time basis.
  • Sentences and other user utterances are represented by their semantic graphs, which may be a result of the parsing process described in the Matching User Input Utterances to User Intents section previously. It should be noted that this in itself may provide an initial level of clustering. For example, in a corpus of 50,000 utterances, the same semantic graph may be used to represent 10 to 20 different utterance forms.
  • Semantic clustering may be used to create groups of utterances that, for a human reviewer, are intuitively close to each other because they are related to the same or to similar topics.
  • a number of proximity metrics may be defined to compare two or more semantic graphs. For example, concept-set proximity and graph-intersection proximity may be used. Both types involve taking a ratio of an aggregate specificity of semantic objects common to the two graphs to the aggregate specificity of each of the semantic objects found in the two graphs. In other words, these metrics represent specificity of common concepts divided by overall specificity.
  • Specificity will be further discussed in the following section. However, generally, the less frequently a term occurs in a reference corpus, the more specific the term may be considered. Therefore, the inverse of a term's frequency may be used as a reliable measure of specificity.
  • the aggregate specificity of a number of different concepts may be computed by taking the sum of their specificities.
  • the concept-set metric may be used to map each of the semantic graphs to a set of concepts and metaconcepts that it contains.
  • the concepts are the concept traits of the graph nodes.
  • Metaconcepts are the values found for modal traits and "referent" traits (e.g., affirmative, negative, interrogative, interrogative-negative).
  • the concept-set proximity of two graphs gl and g2 may be defined as the ratio of the aggregate specificity of common concepts and metaconcepts to the aggregate specificity of each of the concepts and metaconcepts found in gl and g2.
  • the graph-intersection metric assigns a proximity to a pair of graphs gl and g2 in the following way:
  • intersection heuristics use a concept of "ontology distance" between two concepts.
  • the ontology distance between two concepts cl and c2 is defined as a pair (d, h), where d is the length of the path in the concept hierarchy from cl to c2 via their closest common ancestor, and h is the larger of (the distance between cl and the closest common ancestor) and (the distance between c2 and the closest common ancestor).
  • the ontology distance between a concept and itself is (0, 0). Between a concept and a parent concept, the ontology distance is (1, 1), while between two sibling concepts, the ontology distance is (2, 1). [0096] In addition, for the purpose of defining the intersection heuristics in the following discussion, the following may hold true:
  • a node in a semantic graph includes a set of trait valuations and an identity.
  • the set of trait valuations in a node is a total function over the set of available trait names.
  • a graph node n subsumes two nodes cl and c2 as closely as possible if for all valuations (t, v) in n, the corresponding valuations (t, ul) and (t, u2) in cl and c2 are such that v subsumes ul and u2 as closely as possible.
  • value v subsumes ul and u2 as closely as possible if and only if v is the closest common ancestor of ul and u2.
  • Edge label 1 subsumes labels ml and m2 if and only if ml, m2 and 1 are equal or 1 is Attribute.
  • Edge label 1 subsumes labels ml and m2 as closely as possible if 1 subsumes ml and m2 and either 1 is not Attribute or ml is not equal to m2.
  • the subsumption specificity of two trait values ul and u2 and a subsuming value v has been defined as the product of:
  • Intersection Heuristic 1 The following is an example definition of IntersectionHeuristicl(nodel, node2, resultNode):
  • IntersectionHeuristic 1 (c 1 , c2, n).
  • IntersectionHeuristic 1 (rootl, root2, resultRoot), where rootl is the root of gl, root2 is the node of g2 and resultRoot is a newly created node with concept trait Top (i.e., the most general concept in the concept hierarchy).
  • Rootl is the root of gl
  • root2 is the node of g2
  • resultRoot is a newly created node with just one trait, which is the concept Top.
  • IntersectionHeuristic2 (node 1 , node2, resultNode) is now described as follows: Compute a matrix M of proximities between each child of node 1 and each child of node2. Each entry M[cl, c2] of this matrix equals the concept-set proximity between cl and c2, e.g.,:
  • minProx be a value such that, if M[cl, c2] ⁇ minProx, then any node n subsuming cl and c2 is considered to contain no useful information.
  • the first pair (cl, c2) is the pair maximizing M[cl, c2]
  • the second pair (c3, c4) is such that (c3 ⁇ cl and c2 ⁇ c2 and c4 ⁇ cl and c4 ⁇ c2) and M[c3,c4] is the pair maximizing M[c3, c4] with values thus constrained, and similarly for subsequent pairs.
  • the length of the list L will thus be equal to the smallest of (the number of children of nodel) and (the number of children of node2); every child of nodel or node2 will occur at most once in L.
  • Intersection heuristic 1 yields a maximally specific intersection except in cases where two very dissimilar nodes dominate very similar lists of nodes. In so far as, in semantic graphs, the semantics of a node determines the semantics of the nodes it can dominate, this problem does not occur frequently.
  • intersection heuristic 2 takes into account each of the concepts in sub-trees dominated by 2 nodes c 1 and c2 prior to pairing them and is therefore robust with respect to the risk of finding similar node lists under dissimilar nodes.
  • intersection heuristic 2 may be considered close to the optimal intersection algorithm, which would rank pairs based on their graph- based proximities. But while the running time of the optimal algorithm, even if implemented with dynamic programming, is exponential in the product of the depth of the shallower graph and the largest child list, heuristic 2 is linear in the number of graph nodes. Therefore, from a computational complexity point of view, heuristic 2 is suitable for computing an intersection of a set (or corpus) of graphs rather than just two graphs, as described in the "Precision Feedback" section of the discussion.
  • Use of inverse frequency as a specificity metric generally provides a sufficient approximation to measure the specificity of a concept, i.e., its ability to characterize a specific topic.
  • topic selectivity may be measured in a more direct way, namely by comparing a unigram distribution of concepts in a corpus with a conditional distribution of terms in utterances containing the concept to measure.
  • a neutral concept like "interlocutor" typically occurs in a subcorpus whose concept distribution does not essentially differ from the distribution of terms in the whole corpus.
  • clustering may be based primarily on a concept set- based proximity metric; in an alternate implementation, clustering may be based on a graph-proximity metric.
  • two-level hierarchical clustering as described in the "Two-Level Hierarchical Clustering" section, may make use of the same or different proximity metrics for each level; for instance, in one implementation, the first-level clustering may use concept-set proximity, while the second level may use graph proximity.
  • intersection heuristic 1 is primarily used. However, in an alternate implementation the results of both heuristics may be compared for added precision, as described in the Precision Feedback section.
  • the semantic clustering algorithm is based on an algorithm named Quality Threshold (or QT) clustering. It can be outlined as follows: Let SG be the set of semantic graphs corresponding to the user utterances to be clustered.
  • Cluster fragmentation is defined as the scattering across several clusters of utterances pertaining to the same topic.
  • Cluster homogeneity is defined as the extent to which different utterances that have been clustered together actually pertain to the same topic. For example, when clusters are not sufficiently homogeneous, a smaller number of clusters may be generated; however, the clusters may contain user utterances that are not highly related. On the other hand, if there is a high level of cluster fragmentation, a relatively larger number of clusters may be generated that are not easily reviewable, since similar utterances may not be grouped together in the same cluster.
  • SG be the set of semantic graphs corresponding to the user utterances to be clustered.
  • minClusterSize be a parameter of the algorithm denoting the minimum cluster size, such that minClusterSize > 1.
  • C be the set of generated clusters. Initially C is the empty set. For each semantic graph s € SG, associate with s a candidate cluster containing just s. Let Candidate(s) denote this candidate cluster. For each semantic graph si € SG,
  • Candidate(S) be the largest candidate cluster. Let
  • cluster size provides different ways of identifying the largest cluster candidate.
  • the size of a cluster is the aggregate frequency of its members in the set of utterances.
  • the size of a cluster is the number of unique lexically normalized utterances (normalization is performed for case, spacing, etc.).
  • the size of a cluster is the number of unique semantic representations for its members.
  • Step A stipulates a maximum distance to one distinguished cluster member.
  • Step B the following variation (called transitive clustering) may be made to Step B:
  • Candidate(sl) denote the cluster candidate for si .
  • the constant QT2 may be chosen to be a lower threshold than QT.
  • Transitive clustering strategy may reduce cluster fragmentation without impairing cluster homogeneity.
  • clusters of unmatched utterances may include utterances for whose intent no direct or indirect pattern has been defined
  • an addition to the semantic clustering algorithm may be made to add functionality to automatically generate patterns for a cluster.
  • the patterns can be presented for inspection to a human reviewer (e.g., the designer of the conversational agent 104), who may then validate them and determine the desired intent for these utterances before adding the patterns to the conversational agent.
  • the agent designer may modify the patterns before adding them, may decide to discard them altogether, and so on.
  • the association between pattern and intent can be performed automatically, e.g., without user intervention through execution of software such as machine learning or A/B testing.
  • the pattern generation algorithm may yield patterns in order of decreasing specificity. For example, it may be easier to modify specific graphs, e.g., graphs with a multitude of nodes and traits, by the designer so as to capture the desired intent with the appropriate level of specificity.
  • a pattern that is too generic is essentially useless because it can be easily overtaken by more specific patterns or matched incorrectly to utterances for which no appropriate patterns exist. Therefore, a designer reviewing a cluster's patterns in decreasing specificity order may find a point beyond which it is not desirable to continue the review. Once that point has been reached, the agent designer may decide to discard the subsequent suggestions.
  • members of m are marked for exclusion from further intersections to avoid generating patterns that only include insignificant differences.
  • a further application of generating patterns for clusters is that the pattern projections form homogeneous sub-clusters, thereby providing useful hierarchical clustering.
  • clusters individually represent one or more closely related topics, while sub-clusters may each focus on a single topic. Viewing instances of closely related topics (as identified by sub-clusters) close to each other as part of a single cluster may be useful to assess missing content in the agent or risks of interaction between patterns identifying similar intents.
  • the reviewer may choose between a variety of options: • Associate the suggested pattern with a specified intent, and add the pattern to the configuration of the conversational agent.
  • a graphical user interface may be used by the reviewer to modify the pattern manually.
  • the reviewer decides to exclude semantic graphs, the following approach may be used to generate a new pattern. If the two semantic graphs that have been intersected to generate the original pattern are still in the reviewer's selection, the process stops without a pattern being generated. If, however, one or both of the generating utterances have been excluded, the following algorithm may be launched to compute a more precise match than previously:
  • intersection operations are heuristic approximations to the optimum, this is not guaranteed; but using intersection heuristic 2, which is close to optimal, provides a relatively high probability that the results will be bracketing-invariant.
  • This approach to pattern re-computation viz., an intersection heuristic may help to avoid reducing the precision of traits. In particular, unequal concepts may result in a closest common ancestor in the generated pattern.
  • intersection heuristic 2 rather than the slightly faster but less robust intersection heuristic 1, which is used for the initial match generation in an implementation.
  • functionality may be implemented to evaluate, based on an existing corpus of utterances, the impact of adding a pattern to the conversational agent's 104 decision tree 120. This impact is determined by the matching extent of the pattern, which will be described below.
  • the pattern can be simply matched against the semantic graphs of these utterances. Estimating the matching extent provides agent reviewers with heuristics to predict whether the most optimal patterns are being added to improve the agent's performance efficiently and without adding unneeded patterns.
  • matching extent estimation can be performed for suggested patterns. In an alternate implementation, matching extent estimation can be performed on either suggested patterns, suggested patterns that have been modified by the reviewer, or patterns that have been generated through other means (for instance, manually).
  • Matching extent of a pattern [00133] The matching extent of a pattern p is the subset S of semantic graphs of the utterance corpus so that for every s e S:
  • the utterance corpus contains the matching distance for each matched utterance.
  • the identities of patterns involved in direct matches may be stored together with each semantic graph in the corpus and the matching distance may be computed "on the fly.”
  • the set L(p) may be constructed as follows:
  • ext(p) denote the matching extent of pattern p.
  • the pattern (or patterns) with the largest matching extent has (or have) percentile 100%; on the other hand, the patterns with the smallest matching extent have a matching percentile of 100 x s/
  • FIG. 9 illustrates an example 900 of visualization of a matching percentile of a pattern using a meter 902.
  • the meter 902 illustrates matching percentile coverage of a current pattern.
  • a segment illustrated to the left in the meter 902 represents a relatively low matching percentile and a segment to the right in the meter 902 represents a relatively high matching percentile. If in the meter 902 the reading 904 is in one of these segments, this is to be interpreted as a warning that the current pattern coverage is overly restrictive or too general, respectively.
  • the potential matching percentile of a prospective pattern p can be estimated as described in the section "Matching extent of a pattern" above. This matching percentile may be shown using appropriate visualization such as the meter 902 in FIG. 9 for a designer to assess the impact of adding the prospective pattern p to the decision tree 120 of the conversational agent 104.
  • the extents of existing patterns may be recomputed to take into account cases in which an utterance potentially matches p instead of the existing pattern p' that it was matching with previously.
  • the potential matching extent of the prospective pattern p may be presented to the reviewer by displaying a small subset of potentially matching utterances (for instance, on the order of 10).
  • the subset includes utterances with the shortest matching distance to the new pattern.
  • a selectable control e.g., via a user interface
  • this set may be expanded to show the whole matching extent, if desired by the reviewer.
  • the graphical user interface presents the following information for reviewers:
  • the following information 1000 may be shown to the reviewer, an example of which is illustrated in FIG. 10: • The number of matched utterances 1004, with respect to the total number of utterances in the corpus 1002 (both matched and unmatched).
  • the number of unmatched utterances 1006 i.e., utterances not matching directly, including both utterances that the agent does not understand and utterances that trigger reformulations), with respect to the total number of utterances in the corpus 1002.
  • This number or proportion 1008 provides a quantitative estimate of the potential improvement of the matching performance of the conversational agent that can be achieved during the current semantic clustering round.
  • the minimum aggregate frequency k is a parameter of the configuration.
  • the aggregate frequency of a cluster is the total number of occurrences in the corpus of the utterances it contains.
  • semantic clusters that are less frequent than k are not worked on by a reviewer due to providing minimal leverage, although other implementations are also contemplated.
  • the contents of a cluster may be epitomized by displaying an utterance corresponding to the semantic graph having the least specificity in the cluster.
  • the reason for this choice is that such an utterance primarily contains information that played a role while building the cluster. Choosing an utterance with a richer semantic graph may run a risk of displaying incidental information that is not representative of the cluster's topic.
  • the cluster center may be defined as a most central utterance, where centrality is measured by the average of proximities to each member of a cluster. In case there is more than one most central utterance, the most frequent utterance among such most central utterances is selected to represent the cluster's topic.
  • Clustering criteria includes the quality threshold, the threshold criterion (radius vs. diameter), a building method (e.g., transitive vs. intransitive), and a proximity metric.
  • Tuning semantic clustering includes the choosing of clustering criteria such that, in most cases, utterances matched by different patterns in a specified agent will go to different clusters.
  • clustering criteria such that, in most cases, utterances matched by different patterns in a specified agent will go to different clusters.
  • clustering criteria there are a variety of practices for defining graph patterns in agents; so that it is necessary to tune clustering criteria to match the specificity level of a particular conversational agent.
  • two-level hierarchical clustering includes applying a QT algorithm to a corpus of semantic graphs to obtain clusters and then applying the QT with more stringent clustering criteria on the set of semantic graphs in each cluster obtained in the first step to obtain subclusters.
  • semantic clustering technique may be applied to a corpus representing a set of utterances that have been assigned an intent by the agent, e.g., a set of matched utterances.
  • semantic clustering attempts to classify by topic utterances that have already been assigned precise topics (namely, intents in the decision tree 120) by the conversational agent.
  • Two cases may arise, either each member of a cluster is assigned the same intent by the agent (convergent assignment) or members are assigned multiple intents (divergent assignment).
  • the terms convergent cluster and divergent cluster may also be used to describe these cases.
  • Divergent assignment may mean that a cluster spans several different intents (heterogeneous cluster) and those intents have been correctly identified by the agent, in which case no corrective action is desirable. However, provided the quality threshold of the clustering process is tuned to the level of distinction the agent makes between intents, a majority of cases of divergence may arise when a heterogeneous cluster contains some utterances with incorrect intent assignment. Therefore, each case of divergent assignment may be presented for review.
  • Clustering is performed hierarchically at two levels. Sub-clusters are obtained using clustering criteria tuned to the level of specificity of the agent's graph patterns.
  • Semantic clustering based on natural language parsing and the proximity metrics outlined in the "Proximity Metrics" section may be used to feed clustering results to a graphical user interface.
  • the graphical user interface may be employed by a reviewer to examine clusters, and subclusters if desirable, in order to understand what are the most popular topics for a corpus of utterances (e.g., user inputs performed via live chat, IM, SMS, email, social networking sites, blogging and micro-blogging services, etc.), as well as to obtain a quantitative indication of the prevalence of those topics in the corpus.
  • the following lists techniques that may be employed to support this functionality.
  • a cluster can be represented by its center, as defined in the "Cluster Center” section. Additionally, a user selectable (i.e., clickable) user interface control associated with the cluster center may be used to expand the cluster to display its content.
  • a user selectable (i.e., clickable) user interface control associated with the cluster center may be used to expand the cluster to display its content.
  • Proximity to the center may be computed using a variety of metrics, examples of which were described in the Proximity Metrics section, e.g., concept-set proximity.
  • clusters may be referred to as "same-topic" clusters in the following discussion. It should be noted that a simple comparison of cluster member utterances (or semantic graphs) may not be sufficient in some instances to determine if two clusters relate to the same topic. This is because the actual utterances and semantic graphs making up a cluster may vary from one corpus to the next.
  • concept-set proximity provides the ingredients for defining a vector space.
  • each concept may be considered as a dimension and the coordinates may be set as follows: • If concept c is in the concept set of g, the coordinate for dimension c is set to the specificity of c.
  • centroid of a cluster as the point whose coordinates are the averages of the coordinates of each semantic graph in the cluster.
  • the distance between the centroids of two clusters CI and C2 may be computed and a decision made that CI and C2 represent the same topic if this distance is below a specified threshold.
  • the center of a cluster may be defined as its least specific member, an example of which was previously discussed in the Cluster Center section.
  • a decision may be made, for example, that two clusters CI and C2 represent the same topic if and only the set-based proximity of the respective centers is above a specified threshold.
  • the threshold is set between 0.95 and 1.0 although other implementations may also be considered.
  • the center-based heuristic may be used in place of the centroid-based heuristic insofar as (1) the centroid has increased sensitivity to the presence of nonessential concepts than the center and (2) concept-set proximity is perceptibly more accurate than Euclidean distance. Cluster identification in binary comparison of cluster sets
  • inter(Cl), inter(C2) and inter(Cl, C2) may be computed as follows.
  • inter(Cl), inter(C2) and inter(Cl, C2) are initially empty.
  • the sets inter(Cl) and inter(C2) contain same-topic clusters from Cl and C2, respectively, determined based on the proximity identification criterion (as described above). However, for the purpose of displaying cluster members, they may be used as follows, respectively:
  • the set of clusters specific to Cl is the set:
  • the set of clusters specific to C2 is the set:
  • inter(Cl, C2) may be shown. In an alternate implementation, if it is desirable to show more recent clusters, inter(C2) may be used instead.
  • Clustering of a relatively large number of utterances may allow reviewers to identify prominent topics, based on the size of the resulting semantic clusters.
  • This may have a variety of applications such as to make sense of large volumes of utterances efficiently and automatically.
  • an implementation may contain one or more of the following applications of semantic clustering: "Voice of the Customer” analysis, display of "hottest topics” for the benefit of social networking, online, or micro-blogging communities, identification of the top issues on support channels or real-time user feeds such as from social networks or micro-blogging sites, and so on.
  • a variety of techniques may be employed to provide this analysis, an example of two of which is described as follows.
  • a company may seek knowledge of "what is utmost on its customers' minds" and therefore filter a corpus of utterances (consisting of messages or product opinions) to retain those utterances that contain references to specific products or topics, prior to identifying dominant topics in the retained corpus using semantic clustering.
  • Automated topic tracking may be used to filter a corpus of utterances (consisting of messages or product opinions) to retain those utterances that contain references to specific products or topics, prior to identifying dominant topics in the retained corpus using semantic clustering.
  • a technique may be employed to cluster corpora built during time windows that may be close and/or overlapping. For instance, an implementation may use corpora of utterances built each day, by taking into consideration all utterances that have occurred over the last 5 days.
  • a customer support department may be interested in knowing the top issues encountered by customers in real time so as to efficiently allot customer support operators. If this department offers a high-throughput automated chat, a one -hour sliding window may be used to track the main topics on an hourly basis. Topic appearance and disappearance may be brought to the attention of specialists by using automated topic tracking (i.e., without manual topic marking or a predetermined list of topics) the specialists may then decide whether these topics correspond to an underlying issue or to an issue resolution, respectively.
  • Automated topic tracking can be achieved by means of the approach described in the section "Cluster identification in binary comparison of cluster sets.” If CI is the set of clusters from a previous time window and C2 is the set of clusters from the current time window, then CI - inter(Cl) may represent outdated clusters, while C2 - inter(C2) may represent new clusters.
  • Automated topic tracking is also applicable across corpora that are not successors in a time series.
  • a company may be interested in comparing the dominant topics in user utterances obtained through different channels, such as blog contents, blog comments, micro-blogging sites, SMS, or conversations with humans and conversations with conversational agents, during a given time period.
  • a company with operations or subsidiaries in several countries may be interested in identifying common topics across these operations from customer-support logs.
  • the relevant output of the algorithm described in the section "Cluster identification in binary comparison of cluster sets" will be inter(Cl, C2), which yields common topics.
  • An analytics tool offers the possibility of associating other quantitative data beside frequency data with utterances in a cluster. For instance, the results of semantic clustering could be joined against the metrics derived by an analytics system (such as a web analytics solution or a business intelligence solution).
  • an analytics system such as a web analytics solution or a business intelligence solution.
  • analytics data can be aggregated per cluster. For example, it could be of interest to a company to know which topics are most or least conducive to commercial orders (in this case, checkout completion and/or shopping cart value could be the relevant analytics metrics that are joined with the result of semantic clustering), or on the contrary cancelations (in which case shopping cart abandonment may be the relevant metric).
  • Some analytics measures for example measures of conversation success, correlate with the quality of an agent's content and pattern matching. Accordingly, aggregating such measures per semantic cluster and displaying the results to a reviewer provides straightforward clues to areas the reviewer needs to act upon in order to improve a conversational agent.
  • DiameterProximity(s, C) average proximity of s to each member of C where proximity may be one of the proximity metrics that have been defined previously.
  • is defined as follows:
  • TopicFrequency(s, C) DiameterProximity(s, C) *
  • an utterance s is said to be attached to cluster C € S if and only if C is the cluster that maximizes DiameterProximity(s, C) for any C € S. If several clusters satisfy this definition, the cluster C among these several clusters that maximizes TopicFrequency(s, C) may be defined to be the attachment cluster.
  • center of C is defined as in the Cluster Center section, i.e. as the cluster member with least specificity.
  • the diameter method may be used for analytical applications, as in assessing the relative importance of issues reported in the conversational agent 104, whereas the radius method may be used when cluster attachment is used in the conversational agent 104 to identify the intent of a user utterance.
  • a hybrid approach is used by the conversational agent 104. In this approach, diameter proximity is used, except for clusters whose size exceeds a certain threshold, e.g., 500.
  • the above techniques may be leveraged in a variety of different applications. For example, the importance of an individual case of bad-quality response may be accessed.
  • the ability to sort utterances by attachment frequency makes it possible to report an individual issue not as a request to remedy that particular issue; but as information on what topics may be improved; or to decide that improvement is not desirable should it be found that this particular utterance does not correspond to a (frequent) topic. This approach allows a prioritized approach to agent improvement.
  • corpora may be displayed by topic. While nontrivial clusters (i.e. clusters of more than 1 utterance) may account for a varying proportion of a given corpus, cluster attachment may make it possible to consider each single utterance in a corpus as a member of an enlarged cluster. This makes it possible to use summarization techniques that use the center of a cluster or the k most central utterances of an enlarged cluster to represent potentially hundreds or thousands of sentences, thus enabling efficient browsing of a corpus. Such efficient browsing may be further facilitated by sorting enlarged clusters by the sum of the attachment frequencies of their members, thus giving most prominence to the most popular topics.
  • online exploitation of cluster attachment may be performed. Determining the attachment cluster of a single sentence may be performed with sufficient speed to be used as part of a conversational agent's matching process. In cases where a conversational agent 104 cannot determine a specific intent for a user input, identifying its topic through cluster attachment can enable the agent to:
  • topic determination is an inference the agent made and ask for confirmation before answering on this topic

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Business, Economics & Management (AREA)
  • General Engineering & Computer Science (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Educational Administration (AREA)
  • Educational Technology (AREA)
  • Data Mining & Analysis (AREA)
  • Databases & Information Systems (AREA)
  • Artificial Intelligence (AREA)
  • Machine Translation (AREA)

Abstract
EP10717910A 2010-03-26 2010-03-26 Groupage sémantique Ceased EP2569716A1 (fr)

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/US2010/028898 WO2011119171A2 (fr) 2010-03-26 2010-03-26 Groupage sémantique

Publications (1)

Publication Number Publication Date
EP2569716A1 true EP2569716A1 (fr) 2013-03-20

Family

ID=43431931

Family Applications (1)

Application Number Title Priority Date Filing Date
EP10717910A Ceased EP2569716A1 (fr) 2010-03-26 2010-03-26 Groupage sémantique

Country Status (2)

Country Link
EP (1) EP2569716A1 (fr)
WO (1) WO2011119171A2 (fr)

Families Citing this family (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9378202B2 (en) 2010-03-26 2016-06-28 Virtuoz Sa Semantic clustering
US8676565B2 (en) 2010-03-26 2014-03-18 Virtuoz Sa Semantic clustering and conversational agents
US8694304B2 (en) 2010-03-26 2014-04-08 Virtuoz Sa Semantic clustering and user interfaces
US9524291B2 (en) 2010-10-06 2016-12-20 Virtuoz Sa Visual display of semantic information
US10666735B2 (en) 2014-05-19 2020-05-26 Auerbach Michael Harrison Tretter Dynamic computer systems and uses thereof
US9742853B2 (en) 2014-05-19 2017-08-22 The Michael Harrison Tretter Auerbach Trust Dynamic computer systems and uses thereof
US10305748B2 (en) 2014-05-19 2019-05-28 The Michael Harrison Tretter Auerbach Trust Dynamic computer systems and uses thereof
US11003671B2 (en) 2014-11-26 2021-05-11 Vettd, Inc. Systems and methods to determine and utilize conceptual relatedness between natural language sources
US11048879B2 (en) 2017-06-19 2021-06-29 Vettd, Inc. Systems and methods to determine and utilize semantic relatedness between multiple natural language sources to determine strengths and weaknesses
JP6912714B2 (ja) * 2017-07-21 2021-08-04 富士通株式会社 情報処理装置、情報処理方法および情報処理プログラム
US20200175078A1 (en) * 2017-09-11 2020-06-04 Hitachi, Ltd. Multiple intent interpreter and recommender
US11948582B2 (en) 2019-03-25 2024-04-02 Omilia Natural Language Solutions Ltd. Systems and methods for speaker verification
EP4038538A1 (fr) * 2019-10-04 2022-08-10 Omilia Natural Language Solutions Ltd. Induction non supervisée d'intentions d'utilisateur à partir de corpus de services de clients conversationnels
US11126793B2 (en) 2019-10-04 2021-09-21 Omilia Natural Language Solutions Ltd. Unsupervised induction of user intents from conversational customer service corpora
CN111353028B (zh) * 2020-02-20 2023-04-18 支付宝(杭州)信息技术有限公司 用于确定客服话术簇的方法及装置

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2006031609A2 (fr) * 2004-09-10 2006-03-23 Microsoft Corporation Apprentissage automatique
WO2006113970A1 (fr) * 2005-04-27 2006-11-02 The University Of Queensland Mise en groupe de concepts automatique

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2006031609A2 (fr) * 2004-09-10 2006-03-23 Microsoft Corporation Apprentissage automatique
WO2006113970A1 (fr) * 2005-04-27 2006-11-02 The University Of Queensland Mise en groupe de concepts automatique

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
ANDREW E SMITH: "Automatic extraction of semantic networks from text using leximancer", HUMAN LANGUAGE TECHNOLOGY, ASSOCIATION FOR COMPUTATIONAL LINGUISTICS, N. EIGHT STREET, STROUDSBURG, PA, 18360 07960-1961 USA, 27 May 2003 (2003-05-27), pages 23 - 24, XP058108715, DOI: 10.3115/1073427.1073439 *
See also references of WO2011119171A2 *

Also Published As

Publication number Publication date
WO2011119171A2 (fr) 2011-09-29

Similar Documents

Publication Publication Date Title
US10360305B2 (en) Performing linguistic analysis by scoring syntactic graphs
US9275042B2 (en) Semantic clustering and user interfaces
US9196245B2 (en) Semantic graphs and conversational agents
EP2569716A1 (fr) Groupage sémantique
US9524291B2 (en) Visual display of semantic information
US11301632B2 (en) Systems and methods for natural language processing and classification
US10593350B2 (en) Quantifying customer care utilizing emotional assessments
Gao et al. Emerging app issue identification from user feedback: Experience on wechat
US8719192B2 (en) Transfer of learning for query classification
US9043285B2 (en) Phrase-based data classification system
US20200143265A1 (en) Systems and methods for automated conversations with feedback systems, tuning and context driven training
US9424354B2 (en) Providing crowdsourced answers to information needs presented by search engine and social networking application users
WO2020123323A1 (fr) Applications d'apprentissage automatique pour événements liés temporellement
Kubiak et al. An overview of data-driven techniques for IT-service-management
US11757807B2 (en) Interactive chatbot for multi-way communication
Mezouar et al. Are tweets useful in the bug fixing process? an empirical study on firefox and chrome
US20150149463A1 (en) Method and system for performing topic creation for social data
WO2012047214A2 (fr) Affichage visuel d'informations sémantiques
WO2020139865A1 (fr) Systèmes et procédés pour des conversations automatisées améliorées
US20220318681A1 (en) System and method for scalable, interactive, collaborative topic identification and tracking
US11243916B2 (en) Autonomous redundancy mitigation in knowledge-sharing features of a collaborative work tool
WO2023060045A1 (fr) Développement de question-réponse
Tarnowska et al. Sentiment analysis of customer data
Sarkar et al. NLP algorithm based question and answering system
Zhang et al. Labelling issue reports in mobile apps

Legal Events

Date Code Title Description
PUAI Public reference made under article 153(3) epc to a published international application that has entered the european phase

Free format text: ORIGINAL CODE: 0009012

17P Request for examination filed

Effective date: 20121025

AK Designated contracting states

Kind code of ref document: A1

Designated state(s): AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HR HU IE IS IT LI LT LU LV MC MK MT NL NO PL PT RO SE SI SK SM TR

DAX Request for extension of the european patent (deleted)
17Q First examination report despatched

Effective date: 20180327

REG Reference to a national code

Ref country code: DE

Ref legal event code: R003

STAA Information on the status of an ep patent application or granted ep patent

Free format text: STATUS: THE APPLICATION HAS BEEN REFUSED

18R Application refused

Effective date: 20200302