Consider the data from Kwak wala in 7 , which are extensively discussed in Saba Kirchner The contexts for these two mutually exclusive processes are in complementary distribution: when a stem ends in a followed by one or two non-glottalized obstruents, -m u:t is accompanied by lengthening; in all other cases, reduplication occurs. This pattern is summarized in 8. Kirchner analyzes the suffix -m u:t in Kwak wala as containing a non-integrated 3 There are a number of orthogonal processes including allomorphy in affixes, phonotactically driven repairs, and lowering of lenghtened vowels and idiosyncratic exceptions in Kwak wala, none of which devalidate Saba Kirchner s main argument, however.
The former is the default strategy, but whenever lengthening would create a super-heavy syllable a structure that is categorically banned in Kwak wala , copying is chosen as the next-best repair. The Upriver dialect of Halkomelem Central Salishan exhibits an even more dramatic instance of phonologically conditioned reduplicative allomorphy Zimmermann a, a. The continuative forms of verbs has four different types of stem allomorphs: stress shift 9 - a , reduplication 9 -b , vowel lengthening 9 -c , and prefixation of hv- 9 -d.
Stress shift applies to all verbs with main stress on a non-initial syllable in the non-continuative form. The other three strategies apply to stems with initial stress. If the vowel in the first stressed syllable is an epenthetic hv- is prefixed. In all other cases, the first is reduplicated unless the steminitial syllable starts in a laryngeal consonant, in which case the V1 is lengthened and reduplication does not take place.
When overwriting is not possible because the underlying structure is proteced by higher ranked faithfulness constraints, one of the other three repairs apply, depending on which strategy gives the least marked result.
Furthermore, as argued in Zimmermann b , another advantage of MR is that it can offer an account of avoidance and superset effects in multiple reduplication. There remains, however, one obvious gap. An account of over- and underapplication of morphologically induced alternations that uses the basic mechanism of repair-driven copying still constitutes a major desideratum.
In this thesis, I will offer a solution to this problem. I will show how over- and underapplication of segmental mutation can be handled by a phonological theory of reduplication which is compatible with Kirchner s basic insight that copying is triggered by defective prosodic material and not the result of a RED morpheme. The table in 11 gives an overview of the main theoretical ingredients of my proposal. Copying a. Modularity c. Concatenation Copying is a phonological operation available to the grammar for satisfying markedness constraints against defective nodes.
There is a strict division between phonology, morphology, and other modules of the grammar. There are no constraints in the phonological component indexed to specific morphemes. Both mutation and reduplication are the result of non-segmental affixation, and this informs the way these two processes interact with each other. My main claim is that misapplication, as well as alleged cases of reduplication-specific cophonologies, receive a phonological explanation by appealing to the same grammatical building blocks that are independently motivated by other processes in the grammar.
GNLA states that all non-concatenative morphology, including mutation and reduplication, traces back to nonsegmental affixation. Previous research has accumulated an impressive body of evidence that the hypothesis of GNLA indeed holds for isolated morphonological phenomena; for reduplication see Saba Kirchner and Bye and Svenonius ; for morphological manipulation of length see Zimmermann a ; for morphologically triggered segmental alternations see Akinlabi , Wolf , Trommer , and Kimper Accounts of vowel mutation and reduplication in terms of GNLA are illustrated in 12 and 13 , using the Lakota data in 4 and the Kwak wala data in 7.
Each of the proposed analyses will be presented in greater detail in the case studies in chapters 3 5. Overapplication I propose that there are at least three phonological explanations for mutation overapplication that do not rely on diacritic identity relations. The first explanation is deeply entrenched in the assumption of the derivational architecture of grammar. Any cyclic theory of grammar straightforwardly predicts that whenever a string of segments is affected by some process at an earlier level than reduplication, what will be copied is the outcome of that process, rendering the resulting output form non-surface-apparent.
An example of derivationally conditioned overapplication is discussed in Kiparsky Consider the data in When a prefixed form with a derived retroflex is reduplicated, the retroflex appears in the base and the copy, no matter whether or not the initial base C is still in a ruki environment.
Under a derivational approach, this is impossible because the context for applying retroflexation only arises after reduplication has taken place. More cases of overapplication that lend itself to an account in terms of precedence relations are discussed in Mester and Raimy a.
Not all cases of overapplication can be deduced to ordering effects, however. Another grammatical constellation that may give rise to overapplication is the absence of blocking. Blocking effects arise when a process that increases the harmony score of some markedness constraints cannot apply because it also incurs a violation of a higher-ranked faithfulness constraint. Reduplication may create a phonological context in which such a faithfulness restriction is circumvented, either because the relevant constraint only protects underlying material or because copying creates a unique combination of segments that is not proteced but happens not to occur outside of reduplication.
A third possible reason for overapplication is sensitivity to prosodic domains. A crucial prediction of the hypothesis that copying is a phonological repair operation is that there are 8 23 1.
MY PROPOSAL in principle no size restrictions on copied material because the size of the copied portion is determined by the defective node, which may be anything from a to a phrase node. When a copy happens to include a domain which is part of the context definition of some other process e.
Note that the default pattern in parallel evaluations is local application. The reason for this is that the copy mechanism can only select material from the input for copying and does thus not automatically replicate changes induced on base material in a given candidate see discussion in section 2.
Chapter 3 presents one case study for each of the two noncyclic causes of overapplication, i. Chapter 4 discusses sequential ordering as the reason for overapplication of palatalization in Lakota. All three sources of overapplication rule out phonological backcopying. Underapplication Cases of mutation underapplication in reduplicated forms do not follow from a derivational architecture alone.
When a mutation process applies before reduplication, the result is overapplication, while the expected outcome of ordering mutation after reduplication is regular application.
Parallel evaluation of the two processes in non-recursive domains may lead to simple or multiple application depending on the details of the copy mechanism, but never to underapplication. The theory of BMR adopted here explicitly predicts local mutation in such cases. There are at least two more types of underapplication which both arise from scenarios in which reduplication creates a shortage of mutation triggers.
In section 4. When there is an equal number of T s and F s, application of A is optimal. When there are more T s than F s, not applying A is optimal.
TMTP effects arise from conflicting demands on the association of certain nodes in the phonology. A situation which can be characterized in terms of TMTP is sketched in the tableaux in 17 and Mutation here: linking of the floating feature F i to a is triggered by the two synergetic constraints F and F. The latter constraint is ranked rather low, allowing non-integration of a floating F to become a potential winner.
Non-application of mutation in a. Insertion of some other feature F j to satisfy that constraint in c. The candidates in 18 , on the other hand, demonstrate how non-integration of a floating feature becomes optimal when there are more targets than there are triggers.
Since F latter is ranked low, non-integration of the floating F i is optimal when the constraint against line insertion, which is not violated by linking epenthetic nodes, is ranked high enough. These two assumptions together make a predition that has gone largely unnoticed in the literature and that I term the Lossy Copying Hypothesis: 19 The Lossy Copying Hypothesis Let D be a domain stem, word, If parts of D are copied to satisfy a constraint demanding some node N 1 to be associated to a node N 2, the copy will never contain any elements from S and segmental alternations caused by the presence of S in D will not be induced by the copied material.
The lack of these structures in the copied string then mute the copy with respect to whatever morphonological alternation the base is a trigger of. Suppletive allomorphy Apparent cases of suppletive allomorphic relations between base and redupicant present one of the strongest arguments against copying in phonology.
Consider the toy example in 20 , where the allomorph -pek appears following the 1PL marker su- and the allomorph -rimola is used elsewhere.
When the form supek in this hypothetical language is reduplicated, the copy contains the other stem allomorph, -rimola, meaning that what is copied must be a morphological entity and cannot be a phonological string.
If reduplication is nothing more than a construction with two morphologically identical daughters, divergent allomorphy can be accounted for by simply specifying different stem types for each daughter, as shown in The argument for the adequacy of reduplicative constructions as formal objects made in Inkelas and Zoll crucially hinges on the empirical reality of suppletive stem selection in reduplication.
In many case studies discussed by Inkelas and Zoll , however, the allegedly suppletive allomorphs are only minimally distinct, which suggests that they are in fact better analyzed as non-suppletive. Chapter 5 will look at putative cases of suppletive allomorphy and propose reanalyses in terms of phonologically predictable segmental alternations, reconciling them with the hypothesis of phonological copying.
Chapter 2 introduces the critical theoretical ingredients for my account of mutation in reduplication. Following a definition of crucial terms in section 2.
Section 3. I show that there is no need for a grammatical building block tailored to derive backcopying as advertised in Mc Laughlin ; rather, overapplication in Seereer- Siin follows from a simple markedness requirement on the shape of obstruents. A considerable amount of that section is devoted to showing that this approach is consistent with the way reduplication and raising interact with other morpho phonological processes in the language.
Section 4. I argue, contra Saba Kirchner , for a stratal account that captures most of the intricate patterns in the language. I also argue against the view that underapplication is a base-reduplicant identity effect; instead, I propose an analysis in terms of TMTP, i. The case of Kulina in section 4. Section 5. Suppletive allomorphy is one of the empirical pillars of Morphological Doubling Theory Inkelas and Zoll I defuse this argument by showing that mutation in Sye follows from affixation of non-segmental material.
A number of analogous cases are reanalyzed in section 5. Chapter 6 offers further discussion and points at potential theoretical and empirical extensions. Chapter 7 concludes this thesis. While the basic pattern of reduplication is quite straightforward and can be schematized as in 23 , the concrete reduplicative patterns in individual languages and their complex interaction with other grammatical processes often challenge well-established concepts in formal theories of grammar.
In Hinuq emphatic reduplication 24 , a string of segments from the initial onset consonant to the vowel in the second base syllable is copied and the copy is prefixed to the base.
The productivity of this process is evidenced by its extension to loanwords 25 -ad. In Karuk iterative reduplication 26 , the final syllable of a verb stem is reduplicated and an epenthetic vowel is inserted when reduplication creates an illicit consonant cluster 26 -d.
In the case of Hinuq, Bikol, and Karuk, the reduplicant is smaller than the base, making them instances of partial reduplication. Full or total reduplication, on the other hand, creates a copy of all base segments. Examples of full reduplication are given in 27 and The reduplicated root then serves as a base for further derivational and inflectional affixes.
Reduplication of complex stems is observed in Mapuzungun, where a reduplicated verbal stem in combination with the suffix -ye denotes event repetition over an extensive period of time a. A reduplicated nominal or adjectival stem including a number suffix is used in constructions meaning different kinds of b.
QUOT bubbles began to come up out of the water Stonham 28 Complex stem reduplication in Mapuzungun isolate 14 a. Replicative processes are ubiquitous in non-linguistic natural systems such as biology and chemistry, but they also play a key role in mathematics, engineering, and sociology Hein et al. It is therefore not surprising to find replicative processes in various guises in natural language, too, be it as spreading of features, copying of syntactic constituents, or accomodation in discourse.
In the spirit of Hurch a and Frampton , I conceive of reduplication as copying of segmental material to express a morphological category. The data in 29 32 below show examples of repetition of linguistic elements of different sizes that do not fall within the scope of the present work because they do not qualify as morphologically meaningful reduplication. This vowel is not the exponent of a morphological feature but a mere phonotactic filling see also Kawahara It is self-evident that in such cases, no copying of segments has taken place; rather, the [ATR] feature spreads outwards from the root to the suffix independently of the height and frontness specifications in the root and the affix vowels.
Another line that needs to be drawn is one between morphological reduplication and replication in syntax. Copying of syntactic constituents is found in a wide array of contexts in many languages.
This includes verb doubling with and without fronting a , 32 , wh-copying b , and DP-internal determiner doubling c.
Fully inflected adjectives in Sardinian can be reduplicated to express intensification Stolz et al. Echo formation with fixed segments in Dravidian involves copying of large prosodic constituent corresponding to the vp see the examples in and discussion below. Moving away even further from morphology, repetition at the level of discourse may occur spontaneously without being the exponent of a grammatical feature Inkelas and Zoll , Hyman Obviously, such repetitions do not fall under the definition of reduplication, either.
Lirkod, Gil lo yirkod baxayim. Hebrew b. Romani c. Anderson and Harrison In this study, I also set aside the question of pseudoreduplication, i. Due to the lack of a meaningful base and their low productivity, the reduplicative status of expressive pseudoreduplicated onomatopoeias and ideophones such as English sh-sh-sh, Hindi thundering sounds or Mizo olep-olep sticky is dubious, although they are sometimes subsumed under the term reduplication as well Abbi Zimmermann argues for a representational account by which pseudoreduplicants are present underlyingly but have 7 Abbi defines reduplication as a word formation process and distinguishes between morphological and lexical reduplication.
The former refers to the highly lexicalized class of pseudoreduplicated expressives, all of which contain a repetition of one or more syllables but are not morphologically analyzable and cannot be used in isolation.
The latter refers to reduplication that expresses an inflectional or derivational category. Conceiving of reduplication as copying of phonological material raises the question of how to treat morphologically meaningful segmental lengthening and gemination.
At a terminological level, I am drawing a strict line between reduplication and length-manipulating morphology, contra Rubino At a formal level, the two processes have in common that they are both triggered by defective prosodic nodes. The crucial difference is that a change in phonological length results from a mere manipulation of association lines while reduplication involves the creation of a copy of phonological material.
The formal proximity of the two processes is informed by the empirical argument of reduplicative allomorphy and is one of the cornerstones of purely phonological approaches to reduplication within Optimality Theory that refute the notion of a RED morpheme.
This argument will be discussed in more detail in section Mutation Mutation is another well-known non-concatenative process by which a morphological feature is not exclusively expressed by segmental affixation but by a modification of phonological material. I adopt the broad definition of mutation morphology in Wolf and Trommer and treat any morphologically triggered phonological alternation as an instance of mutation morphology. The examples in 33 show the broad major classes of mutation morphology.
Initial consonant hardening in Nivkh a. Gradation in Pite Saami b. Ablaut in Hidatsa c. Umlaut in German d. Plural in Ngbandi preterite verbs e. Somali f. The imperative in Gidabal g. The data in 33 -e and 33 -f are instances of suprasegmental mutation morphology see i. Myers , Donohue , Yip The examples in 33 -g and 33 -h present cases of Morphological Length Manipulation MLM , an empirical field that is extensively documented and discussed in Zimmermann a.
MLM is of theoretical relevance to the current study because it and reduplication derive from the same basic mechanism, viz. It is, however, not within the empirical scope of this dissertation. Rather, I shall confine myself to cases of segmental mutation: initial, medial, and final consonant and vowel mutation, as illustrated by 33 -a 33 -d.
Xa- shoot Nivkh, Paleosiberian Shiraishi 83 cx1f q h a- shoot a bear b. Hidatsa, Sioux Park ma-ruwiira-g I twisted it and GNLA aims to derive all productive cases of non-concatenative morphology from the concatenative affixation of phonologically defective material such as floating tones or empty foot nodes Trommer and Zimmermann The fundamental architectural assumption underlying GNLA is that morphological exponence is inherently item-based and not procedural Bye and Svenonius The central innovation that I propose here is that over- and underapplication of mutation in reduplicated forms, often believed to be a paradigm cases of morpheme-specific subgrammars, are entirely phonological in nature and follow from the exact same principles as the two processes in isolation Phonologically defective representations Prosodically Defective Morphemes The most basic background assumption about the organization of prosodic structure is the Prosodic Hierarchy The Prosodic Hierarchy states that prosodic nodes are organized in 18 33 2.
Under certain grammars, defective prosodic nodes react with other phonological material and cause sometimes dramatic changes, in a similar fashion as segmentally defective features are involved in mutation morphology. Zimmermann a discusses segmental lengthening, shortening, deletion, epenthesis, stress shift, reduplication, and blocking of independent processes as possible outcomes of such reactions.
The only contribution of the morphology is providing the phonological material for optimization. That such material may contain defective structures is independently predicted by ROTB. The plural suffix -Vk triggers deletion of a final short stem vowel a. Some stems, however, resist the subtractive effect of -Vk c. Stems that are immune are equipped with an additional mora at their right edge which is also defective.
That mora is usurped by the prosodically defective suf- 19 34 THEORY fix vowel without causing a visible change in the stem form. However, research on specific artificial intelligence related tasks has made a huge progress, for example in the fields of image processing, knowledge engineering and especially the field of computational linguistics and natural language processing.
Recent computer systems are now able to process large amounts of unstructured data, such as text, and transform relevant information into structured data. Speech recognition and textto-speech software enabled communication interfaces that allow humans to communicate with a computer in natural language. Other technological development, such as the appearance of the internet and mobile devices, opened new perspectives for communication systems such as conversational agents. A conversational agent or dialogue system is a computer program that uses natural language to engage humans in a conversation, either spoken or written.
There are plenty of text based dialogue systems used in practice for domain specific tasks such as sales, marketing, FAQ 1 or customer support. Also, a few are used for e- learning and tutoring purposes. Still, the majority of these task-oriented systems used 1 Frequently asked questions 1 16 Chapter 1. Introduction in practice apply rule-based approaches [Chiticariu et al. Most of the information retrieval systems, in contrast, apply statistical or machine learning methods to process natural language.
Personal assistants and information retrieval systems for mobile devices, such as Apple s Siri, Google Now or Mircofts Cortana, have found their way into everyday life by providing knowledge and information from the web.
Nevertheless, there is still a gap between natural language interfaces for information retrieval and truly intelligent, cognitive agents. Holding and keeping up a coherent conversation, even in written conversation, is still a big challenge.
Conversational agents for entertainment, called chat-bots, are still not able to trick humans into believing they are talking to a human at all. Personal coaching, for example, is a promising yet challenging field: The client-centric coaching maxim of counseling without advice [Radatz, ] involves the guidance of individuals by stimulating questions, rather than by concrete instructions or correct answers.
Therefore, a computer based coaching system does not necessarily need to understand the semantic details of the client s problem domain. We see a chance for a new dialogue system for the purpose of coaching. Such a coaching dialogue system aims to hold a professional, structured conversation instead of entertaining or tricking humans into believing it is human.
Thus, we see a perspective for a new dialogue system for the purpose of coaching. So far, conversational agents have not been developed for personal coaching. Furthermore, an introduction into computer based coaching and related work presented.
Theoretic background includes information about dialogue act and grounding theory. Background is followed by the implementation details of the dialogue act classifier, the sub-dialogue model and communication and turn-taking strategy of VPINO. In Chapter 4, we introduce an implementation of VPINO as a training transfer coach, followed by a user-study on the evaluation of its effectiveness. An overview of the software architecture and the authoring tool developed in the course of this thesis is presented in Chapter 8 The thesis is closed with a summary of conclusions and findings, and give an outlook on future work in Chapter 9.
What is Coaching? In a professional setting, coaching has successfully been applied across many areas such as human resource development, team-building, decision-making or improvement of various individual skills such as leadership, communication or sales, just to name a few.
However, there is no coherent definition for the term coaching so far. Coaching or counselling is often described as a profession rather than a research discipline, although modern approaches originate cognitive and behavioural science i. In a survey on coaching in research and practice, [Grant et al.
This section we will give a glance over the common definitions for the term coaching and the resulting process. The term coaching is also used interchangeably with counselling, consulting and others.
For reasons of consistency, we will keep using the term coaching throughout this work. More detailed information about the specific coaching approaches applied in this work, i. Therefore, the responsibility of a coach in this process is to 1 discover, clarify, and align with what the client wants to achieve, 2 encourage client self-discovery, 3 elicit client-generated solutions and strategies and 4 hold the client responsible and accountable. According to [Hamlin et al. The role of the coach is to facilitate the clients movement through this selfregulatory cycle by helping the client to develop specific action plans and then to monitor and evaluate progression towards those goals 18 Chapter 1.
Introduction While many definitions of coaching focus on the roles of the client and the coach, others define coaching as a process for improving problem work performance [Fournies, ], or a process of guidance, encouragement, and support to the learner [Redshaw, ] A Goal Oriented Process Whereas the definitions and models emerge from different application areas, it is plausible that they differ. However, all of them agree on a view of coaching as a collaborative relationship formed between coach and client for the purpose of goal attainment.
The relationship is characterized as client-centric, meaning the client is valued as the expert that plays an active part in the coaching process.
The goal-oriented, goal-focused or solution-driven paradigm sees the primary function of coaching in fostering the client s self-regulation [Ives, ]. According to [Grant, ] Coaching is essentially about helping individuals regulate and direct their interpersonal and intra-personal resources to better attain their goals. The coaching process involves four consecutive processes: 1 goal setting, 2 examination of real conditions, 3 finding options, alternative strategies or courses of action and 4 implementation of the objectives.
In an integrative review of ten coaching models, [Carey et al. A detailed survey of research on the effectiveness and the state of the art in coaching can be found in [Grant et al. The work presented in this thesis will focus on the sub-disciplines of rational decision coaching and training transfer coaching employing the paradigm of client-centric goaloriented coaching Virtual Coaching and E-Coaching Research on coaching in the e-learning context, also called e-coaching, is still a relatively young discipline.
Technological innovations and the increasing demand for highly available coaching that is independent of time or space restrictions were the main reasons for this development.
However, coaching was not yet applied in an entirely computer based variant. Although most of the e-coaching approaches try to employ modern media for coaching, nearly all of them are limited to technical substitution of face-to-face communication through synchronous or asynchronous text chat, phone coaching or video commu- 4 19 1.
Although increasingly supported by technology, these forms of e-coaching always use a human coach in the background. To our knowledge, there are no natural language systems that focus on decision support or coaching in general. Their system is based on a system-directed dialogue planner DTask [Bickmore et al. Therefore, it does not support natural language communication.
The process of goal attainment includes setting of goals, developing action plans and putting these plans into practice. The client-centric view of coaching gives a good perspective for a computer based coach: A coach does neither have to be an expert in the client s problem domain, nor does he necessarily need to understand the user s problem or goals at all. The role of the coach is to help the client to structure the problem and guide the process to client-generated solutions and strategies.
Accordingly, the client-centric approach achieves goal attainment by utilizing the client s own problem solving abilities, not by giving advice. The verbalization of the clients thoughts, i. This positive effect of verbalization, or self-explanation, is also often used as a learning strategy in the educational domain [Ericsson and Simon, , Aleven and Koedinger, ].
Therefore, detailed understanding of the underlying problem is not necessarily required. As a consequence, we argue that there is no need of deep artificial intelligence in order to implement an effective coach. Many of the coaching definitions, such as the GROW model [Whitmore, ], describe a somehow formal process for generic problem solving.
In this thesis we will show that this generic process can be transformed into a plan for structured conversation. A formal structure may not be suitable for each and every possible coaching scenario, but it can be successfully applied to rational decision coaching or training transfer coaching. In this section, the idea of adapting the principles of coaching with a conversational agent is discussed.
A conversational agent is a computer system that is capable of holding a structured, coherent conversation with a human user in natural language, either spoken, written or embodied i. Introduction There are a lot of synonyms or closely related terms for dialogue systems such as chat bots, chatter bots, virtual assistants, chat bot, conversational agents or virtual agents. In this work, we will keep using the terms conversational agent or dialogue system.
Also, this work focuses on text based chat- communication. In the following, we introduce our dialogue system for coaching and describe the basic conceptual principles. More background information and implementation details will be presented in the following chapter Vpino In this work, we introduce VPINO, a text based natural language dialogue system specifically developed for the purpose of holding structured, goal directed coaching conversations.
It is a mixed initiative system, which means that both conversation partners are allowed to take the initiative at any time. VPINO thereby refers to cognitive skills often attributed to the animal fox Latin: Vulpes , such as smart, clever, intelligent, cunning or sly. Socratic questioning or Socratic dialoguing , also referred to as maieutics, is a general conversation technique that is attributed to the Greek philosopher Socrates. The goal of a Socratic conversation is to explore implicit knowledge, uncover assumptions and follow logical implications by systematic and disciplined questioning.
Socratic dialoguing is known to improve critical thinking on the problem subject [Paul and Binker, ]. Although often characterized as a pedagogic method, Socratic questioning is also used in the field of coaching and psychotherapy [Neenan and Palmer, ]. In accordance with modern coaching approaches, the client is seen as the expert on the problem subject. Hence, coaching with Socratic questions does not intend to give advice, or push the client into any certain direction, but rather help the expert i.
An ideal Socratic coach would do this by asking targeted questions only. A Conversational Agent as a Coach Figure 1. Vpino starts the conversation with a greet and asks if the user is ready to go on.
The user agrees, and Vpino moves on to ask for the user s expectations on the coaching conversation Proactive Behaviour A major task of VPINO is to guide the client through the coaching process. Therefore, in order to keep control, VPINO needs to be aware of the state and direction of the conversation at all times. According to [L Abbate et al. Thereby, the complexity for the tasks to understand the user responses and to keep track of the conversation state is reduced.
Introduction Natural Language Understanding Since a coach does not necessarily need to understand the exact details of a client s utterances, shallow natural language understanding capabilities based on dialogue acts are sufficient for directing the course of a coaching conversation.
With pro-actively targeted formulations, the communicative function of user utterances, the dialogue act an utterance performs, is sufficient for VPINO to continue the conversation properly. Therefore, the general natural language understanding capability is based on dialogue act classification See Sect.
Based on the classification of a user utterance, VPINO determines a possible reaction to that specific dialogue act. The user s responses are parsed for semantic information only when needed for continuing a fruitful conversation, depending on the current context and state of the conversation. VPINO is not intended as an information retrieval or question answering system. Therefore, it will never answer user questions or requests related the problem. Conversational agents are a rather wide field with a lot of sub-disciplines, such as question answering or information retrieval, chatbots, and others.
Holding a structured coaching conversation with a dialogue system differs from many other application scenarios. Thus, in order to avoid misunderstandings, we will point out the differences to these scenarios and existing systems. We distinguish between holding a structured coaching conversation and other application areas with respect to domain closure, dialogue coherence and dialogue objective.
This section begins with a brief introduction into the history of artificial intelligence and conversational agents. After that, a selection of chatbots and information retrieval systems is presented. Finally, this section presents more relevant work from the tutoring domain and other areas Artificial Intelligence The topic of conversational agents is always closely related to artificial intelligence AI. One of the most important milestones in the history of AI was the proposition of the Turing Test by the British mathematician Alan Turing in [Turing, ].
The intention of the Turing Test was to determine if a machine has thinking skills equivalent to humans. Turing suggested that if humans were unable to distinguish the computer s responses from that of a human, the computer could be said to be thinking and thus can be called intelligent.
The Turing Test had a strong influence on the definition of AI. Nowadays, literature distinguishes between strong and weak AI. A computer system, that is capable of full human 8 23 1. Background and Related Work cognitive skills is called strong AI. Weak AI systems, in contrast, are limited to solving specific problems or reasoning tasks.
However, the proposition of the Turing Test led to philosophical questions on strong AI, which are discussed until today. Regardless of the criticism on the Turing Test, Hugh Loebner established the first formal instantiation of a Turing Test in , called the Loebner Prize 3.
However, in contrast to the large number of conversational agents participating in this competition, VPINO is not intended to pass the Turing-test.
Needless to say, in a professional setting such as decision coaching, the client should always be aware that his partner is not a human Chatbots While the existence of strong AI is still discussed, a large number of conversational agents have been developed.
ELIZA is a simulation of a Rogerian psychotherapist, implemented using a rather simple pattern matching techniques. ELIZA s responses are generated by substitution of key words, which are extracted from the user s utterances, into predefined phrases. Moreover, it was rather intended as a technology demonstration than a therapist. Weizenbaum chose this scenario for rather simple, practical reasons: He argued that Therefore, from a programming perspective, Weizenbaum argued that one would not assume that the therapist knew nothing about boats, but that he had some purpose in directing the conversation in that direction.
Even after explaining the mechanism behind ELIZA, some of them were hard to convince of the fact that they have been talking to a machine. Nevertheless, with the role as a coach, VPINO also profits from this assumed credibility by the user. Given the 3 9 24 Chapter 1. Introduction transcripts of conversations, human psychiatrists were not able to distinguish between human patients and PARRY. In [Saygin et al. ELIZA directs the conversation away from herself by asking questions. Like ELIZA, it makes use of a pattern matching approach and follows a simple stimulus-response principle.
AIML basically consists of a large knowledge collection, a set of rules. These rules are represented as pattern-template pairs rules : Patterns match the users text input with a subset of regular expression syntax.
The pattern language consists of words, spaces and wildcard symbols. Templates define the AIML response in case of a pattern match. The simplest response for a matched pattern is plain text.
Additionally, templates are able to set and read simple unary attributes and invoke other programs, e. Templates can also recursively call the pattern matcher with a text defined inside the template. Recursive substitution is probably the most important feature of AIML. It is used to simplify user language, normalize and correct spelling and grammar, detect keywords or synonyms, and transform user input into categories and concepts.
Although very powerful, this feature of AIML is also considered its biggest weakness. The possibility writing recursively self-modifying scripts makes AIML systems hard to author and even harder to maintain.
Tracing a user response or debugging is nearly impossible. Apart from the attributes, which are mainly used to store names or topics, AIML conversations are stateless. The attribute values cannot be accessed by the patterns. Also, AIML systems do not maintain and work with a conversation history.
Wallace [Wallace, ]. However, A. Its original version contains about , units of knowledge that were manually authored by a human. According to [Wallace, ], AIML implements a form of supervised learning where a person he calls botmaster plays a crucial role in training the bot. Thereby, it leads to a large reduction to the number of rules compared to AIML.
These unsupervised learning systems are crowdsourcing the bot content to the users. Unsupervised learning has its own drawbacks, for example that the The trade-off between supervised and unsupervised methods might be summarized as creative writing vs.
Chatbots follow the stimulus-response principle at all times, and therefore lack intentionality. As the term chatbot suggests, the quality of a conversation with a chatbot can be characterized more as chat or smalltalk instead of holding a coherent conversation.
Consequently, AIML is restricted to local dialogue coherence which makes insufficient for coaching conversation. The implementation of a coach requires to hold a coherent, goal oriented conversation over more than two consequent turns with the user.
Nevertheless, AIML is a widely spread amongst conversational agents used in practice. However, most practical application scenarios e. Apart from entertaining chatbots, a lot of conversational agents have been implemented for practical application scenarios, many of them using well established AIML.
However, most of the systems used in practice focus on providing domain specific information. The story of the game develops by chatting with the artificial characters Will and Grace who strive about their marriage. The player 26 Chapter 1. Introduction modifies the relationship to the characters, and their internal emotional states e.
Instead, the player s text input is interpreted as discourse acts, a representation of the general meaning of the player s action, similar to VPINO s dialogue acts. However, conversation focuses on discourse, with potentially more than just two communication partners. It is a text-based, mixed initiative system that simulates the discourse patterns and pedagogical strategies of a typical human tutor in an introductory computer literacy course. It is also equipped with a talking head embodiment, capable of performing gestures.
Based on a set of topics from computer literacy domain, AutoTutor provides four different types of questions, also called topic formats: 1 Question - Answer, 2 Didactic information - Question - Answer, 3 Graphic display - Question - Answer, 4 Problem - Solution. Each topic format includes a main question that is presented to the learner. To determine the type of the students reaction, responses are classified into speech acts.
In order to evaluate the quality of the students answers to reasoning questions, AutoTutor makes use of Latent Semantics Analysis [Landauer and Dumais, ], a technique for analysing relationships between semantic concepts in text documents.
Based on the evaluation results, it also provides feedback on the students answers and possibly pumps for more information if needed. Background and Related Work Conversational Agents for Coaching As already mentioned in the section on e-coaching See Sect , professional coaching has not been extensively used with conversational agents.
In the healthcare domain, Sim- Coach is an animated virtual assistant with the focus on promoting access to specific health care information [Rizzo et al. SimCoach is an embodied, mixed initiative system.
Unlike the name suggests, it does not actually coach the users, at least not in our sense of coaching. Moreover, SimCoach provides information and advice for veterans with the focus on health care problems e. IR and QA systems aim to provide answers either in natural language or as structured information to user requests in natural language.
IR systems make use of advanced natural language processing and machine learning techniques to analyse, extract and generate new semantic information based on given explicit knowledge base. This knowledge base is usually provided as a large set of unstructured information i. Therefore, we will not get into the detailed mechanisms of information retrieval systems. With their state of the art information retrieval system Watson [Ferrucci et al.
Cognitive computing systems help human experts make better decisions by penetrating the complexity of Big Data. VPINO, on the other hand, supports rational decision-making with a client-centric approach. Although QAs use text or speech based natural language interfaces, they are not intended for holding a coherent conversation. Introduction 1. We implemented a rational decision coach and a training transfer coach for communication skills.
Our artificial coach intends to support clients with reflection on their goal accomplishment process. We apply the technique of Socratic questioning. Instead of suggesting or instructing the client, we ask targeted questions. These coaching questions, as well as all other possible responses by the system, are precisely preformulated.
By keeping the leading role and guiding the client, the system is able to keep track of the conversation context to a reasonable extent.
Thereby, our dialogue system does not need to fully understand the semantic content of the client s responses in detail.
Instead, the system classifies the client s reactions as dialogue acts. With a model for common dialogue act sequences, VPINO keeps track of the state and direction of the conversation.
Of course, the goal of the work presented in this thesis is not to replace real human coaches in general; just like Joseph Weizenbaum was sure that his famous chatbot ELIZA is not a way to replace psychotherapists. Our vision is a low threshold computer-based transfer coaching for situations where a personal human coach is simply not affordable or available. This thesis is a first step in that direction. We conducted a series of studies to examine whether humans would accept such a system and consider it a useful tool.
We evaluated our tool in a proof of concept and applied it to the concrete coaching scenarios training transfer coaching and rational decision coaching Proof of Concept In order to find out whether our approach is applicable on professional coaching in general, we conducted a pilot study and tested a prototype as a proof of concept.
We implemented a general Socratic coach on 10 students See Chapt. Subject of the conversations were career related problems and decisions. The results of this pre-study convinced us that our approach is promising for more specific professional tasks such as rational decision coaching. The study also revealed the necessity for more human-like conversations and grounding techniques See 2.
In addition, these first results suggested that the clients personality and attitude have an impact on the quality and efficiency of our tool Scenario I: Training Transfer We used our dialogue system as a training transfer coach for communication skills. We conducted a user study where the clients received an online communication training on 14 29 1. Contributions and Findings weekend followed by a week of daily coaching sessions with our dialogue system. The clients daily work and progress on their communication skills were the topic of the conversations.
Our computer based coaching was tested against a conventional transfer method with respect to effectiveness and influence of the users personality and attitude. As a result of the study, we found that by using our tool clients could successfully improve their communication skills. In accordance with our first qualitative impressions from the pre-study, we found that the personality trait openness of the client has an impact on the outcome of the transfer coaching: More open users that received a daily coaching intervention by our system particularly improved their communication skills.
We also implemented a version of VPINO specifically tailored to support humans making decisions on a distinctly rational basis.
The method for rational decision-making is an adoption of Benjamin Franklin s famous Pros and Cons. Franklin s approach was extended with more modern concepts such as goal setting. Our dialogue system employs a client-centric coaching approach where the client is seen as the real domain expert.
Therefore, the system does not make a decision for the client, nor does it suggest a particular option. Instead, the dialogue system guides and leads the user through the chat conversation while hiding the details of the underlying methods on decision-making.
User Study on Decision-Making We conducted a user study to measure the clients goal attainment when using our coach, and factors that moderate success of the conversation. The participants were free to choose a rational decision problem of their choice and speak to the coach about that problem. Overall, the decision coach was evaluated quite positively. We were able to help a large number of participants, either with clarification of their situation or guidance on making a decision.
The results suggest that the clients personality, motivation, cooperation and their usual approach to decisions affect success of the conversation.
Participants with a rather hypervigilant approach on making decisions profit in particular from holding a conversation with our dialogue system. However, a successful conversation requires a reasonable level of motivation and cooperation from the user. Clients who do not take the system seriously enough or try to challenge it will not enjoy target-aimed dialogues. Introduction Follow-up User Study on Decision-Making We further improved our decision-making dialogue plan based on the insights from the first study.
Previous results suggested that showing off intelligent, human like behaviour, motivate participants for more cooperative behaviour.Also, it supports modularization in different files or blocks and the invocation of dictionaries, which is a huge organisational advantage for authoring and maintaining rules. One of the most important milestones in the history of AI was the proposition of the Turing Test by the British mathematician Alan Turing in [Turing, ]. Their system is based on a system-directed dialogue planner DTask [Bickmore et al. I could not have made it without any of your guys support. It is also not clear what rules out an analysis that assumes the plural morpheme contains two identical segments with different anchor points. MY PROPOSAL in dissertation no size labor supply business plan on copied material because the des of the copied portion is. More precisely, the rules are used to create new annotations corresponding to the dialogue act of a text determined by the defective node, which may be anything from a to a phrase node. My flexibility zur incredibly subpar and I easily wore magazines found in the prefaces to comparison contrasting essay a topic to write on.
However, in contrast to the large number of conversational agents participating in this competition, VPINO is not intended to pass the Turing-test.
This knowledge base is usually provided as a large set of unstructured information i.
Nevertheless, the appearance of machines with human like skills seems to be behind schedule. The table in 11 gives an overview of the main theoretical ingredients of my proposal. Stems that contain only neutral vowels count as front. The list of additional conditions specifies additional constraints that the matched text or annotations need to fulfil to perform the actions defined in the list of actions.
This argument will be discussed in more detail in section Mutation Mutation is another well-known non-concatenative process by which a morphological feature is not exclusively expressed by segmental affixation but by a modification of phonological material.
This massively reduces the number of possible candidates that need to be evaluated and at the same time calls for specific versions of constraints on association relations between phonological nodes. MR does not stipulate a constraint that could compel identity between base and reduplicant. Furthermore, an introduction into computer based coaching and related work presented. The interesting fact is that there is a fourth group of speakers for whom only the initial syllable is a possible target, and if laxing harmony applies, it skips the high vowel in the penult.
Nevertheless, the appearance of machines with human like skills seems to be behind schedule. The clients daily work and progress on their communication skills were the topic of the conversations. Instead, the system followed the cooperative principle [Grice, ] and the concept of adjacency pairs [Schegloff, ].
Section 3. I will provide arguments for a phonological account of reduplication by showing that over- and underapplication of mutation are not only compatible with, but in fact a logical consequence of a theory that assumes non-segmental affixation to be the sole trigger of non-concatenative morphology.
Finally, I would like to express my deepest gratitude to Tatjana, for always being there and cheering me up when needed. Background and Related Work Conversational Agents for Coaching As already mentioned in the section on e-coaching See Sect , professional coaching has not been extensively used with conversational agents. And why should a mutation operation that normally applies only once do so multiple times if the phonology is blind to whether a given process simultaneously affects a base and a reduplicant? Hence, coaching with Socratic questions does not intend to give advice, or push the client into any certain direction, but rather help the expert i. DEP militates against inserted association lines and MAX is a faithfulness constraint 12 protecting underlying association lines.
The work presented in this thesis will focus on the sub-disciplines of rational decision coaching and training transfer coaching employing the paradigm of client-centric goaloriented coaching Virtual Coaching and E-Coaching Research on coaching in the e-learning context, also called e-coaching, is still a relatively young discipline. IR systems make use of advanced natural language processing and machine learning techniques to analyse, extract and generate new semantic information based on given explicit knowledge base.
As opposed to morphological backcopying, this particular prediction is far from uncontroversial, and none of the few reported cases of phonological backcopying are very convincing. However, participants with a generally less structured decision making approach particularly benefit from using VPINO. The constraint formulations in 54 show this difference in very general terms Trommer 66; see also Albright In all other cases, the first is reduplicated unless the steminitial syllable starts in a laryngeal consonant, in which case the V1 is lengthened and reduplication does not take place. At the same time, the shape of the reduplicant usually obeys very general phonological principles such as avoidance of marked structures and prosodic templates, which suggests that reduplicative morphemes are also partially phonological in nature.
The set of classification rules is applied to pre-processed text.