Smart Enough to Talk With Us? Foundations and Challenges for Dialogue Capable AI Systems
Barbara J. Grosz · Computational Linguistics · 2018
I am deeply grateful for the honor of this award, all the more so for its being completely unexpected. I am especially pleased by the recognition this award gives to our early attempts to build computational models of dialogue and to develop algorithms that would enable computer systems to converse sensibly with people. ACL was my first academic home. I presented my first paper, the paper that laid out the basics of a computational model of discourse structure, at the 1975 ACL meeting. Then, after several decades of research centered on dialogue systems, my research focus shifted to modeling collaboration. This shift was driven in part by the need for computational models of collaborative activities to support dialogue processing, a topic which I explore briefly below, and in part by limitations in speech processing and semantics capabilities. Research in these areas has advanced significantly in the last decade, enabling advances in dialogue as well, and I have recently returned to investigating computer system dialogue capabilities and challenges. I am glad to be back in my intellectual home.The use of language has been considered an essential aspect of human intelligence for centuries, and the ability for a computer system to carry on a dialogue with a person has been a compelling goal of artificial intelligence (AI) research from its inception. Turing set conversational abilities as the hallmark of a thinking machine in defining his “imitation game,” more commonly referred to as “The Turing Test.” In the 1950 Mind paper in which he defines this game, Turing conjectures thus, “I believe that at the end of the century the use of words and general educated opinion will have altered so much that one will be able to speak of machines thinking without expecting to be contradicted.” (Turing 1950, p. 442). Though the Turing Test remains an elusive (and now debatable) goal, this conjecture has proved true. Natural language processing research has made great progress in the last decades, with personal assistant systems and chatbots becoming pervasive, joining search and machine translation systems in common use by people around the world. Although it is commonplace to hear people talk about these systems in anthropomorphic terms — using “she” more frequently than “it” in referring to them — the limitations of these systems’ conversational abilities frequently leads people to wonder what they were thinking or if they even were thinking. As the dialogue in Figure 1 illustrates, at root these systems lack fundamental dialogue capabilities. The first assignment in a course I have been teaching, “Intelligent Systems: Design and Ethical Challenges,” is to test phone-based personal assistants. Although some systems might handle this particular example, they all fail similarly, and the range of dialogue incapabilities students have uncovered is stunning. As I have argued elsewhere, the errors that systems make reveal how far we still have to go to clear Turing’s hurdle or to have systems smart enough to (really) talk with us (Grosz, 2012).The first generation of speech systems, which were developed in the 1970s, divided system components by linguistic category (e.g., acoustic-phonetics, syntax, lexical and compositional semantics), and dialogue challenges of that era included identifying ways to handle prosody and referring expressions, along with nascent efforts to treat computationally language as action. In contrast, current spoken language research takes a more functional perspective, considering such issues as sentiment analysis, entrainment, and deception detection. Typical current dialogue challenges include such functions as turn-taking, clarification questions, and negotiation. The database query applications of earlier decades, which lacked grounding in dialogue purpose, have been replaced by template filling and chat-oriented applications that are designed for relatively narrowly defined tasks.Various findings and lessons learned in early attempts to build dialogue systems have potential to inform and significantly improve the capabilities of systems that currently aim to converse with people. Although the computational linguistic methods that have recently enabled great progress in many areas of speech and natural-language processing differ significantly from those used even a short while ago, the dialogue principles uncovered in earlier research remain relevant. The next two sections of this paper give brief summaries of the principal findings and foundations in models of dialogue and collaboration established in the 1970s through the 1990s. The following section examines principles for dialogue systems derived from these models that could be used with current speech and language processing methods to improve the conversational abilities of personal assistants, customer service chatbots, and other dialogue systems. The paper concludes with scientific and ethical challenges raised by the development and deployment of dialogue-capable computer systems.My first papers on dialogue structure (Grosz [Deutsch], 1974, 1975; Grosz, 1977) were based on studies of task-oriented dialogues I collected to inform the design of the discourse component of a speech understanding system. In these dialogues, an expert (who was eventually to be replaced by the system) instructed an apprentice in equipment assembly. The expert and apprentice were in separate rooms, they communicated through a teletype interface (with a camera available for still shots on demand), and the apprentice was led to believe the expert was a computer system. These dialogues were collected as part of what were, to my knowledge, the first “Wizard of Oz” experiments.1Figure 2 is one example from this collection. It illustrates the influence of dialogue structure on definite noun phrase interpretation. In Utterance (3), the expert (E) directs the apprentice (A) to loosen two setscrews. In Utterance (8), the expert instructs the apprentice to loosen “the screw in the center”. The only intermediate explicit mention of screw-like objects are subsequent references to the setscrews in Utterance (5) and the elided subdialogue about them between Utterances (3) and (5). With only two objects, one cannot be in the center. Indeed, the screw to which the expert refers in Utterance (8) is in the center of the wheelpuller. This dialogue fragment has a structure that parallels the task of removing the flywheel. When the first step of that task (loosening the setscrews) is completed, as indicated by the apprentice in Utterance (5), the focus of attention of expert and apprentice move to the step of pulling off the wheel with the wheelpuller. At the point of mention of “the screw in the center”, the setscrews are no longer relevant or in context; the wheelpuller is.A more striking example is provided by the dialogue sample in Figure 3. Each of the successful assemblies of the air compressor ended in one of the ways shown in this figure. Despite admonitions of grammar teachers and editors that pronouns be used to refer to the last object mentioned that matches in number and gender, actual use of pronouns can vary markedly. In this case, the pronoun “it” is used to refer to the air compressor after half an hour of dialogue in which there was no explicit mention of it.Again, context and the focus of attention has shifted over the course of the assembly and the expert-apprentice dialogue related to it. At the point at which the last utterance occurs, the air compressor is assembled and is the focus of attention. Not only did the participants in these dialogues have no trouble understanding what was to be plugged in or turned on (even though they were in a room full of equipment that might have provided other options), but also readers of the transcripts of these dialogues are not confused.The narrative fragment in Figure 4, which is from the work of Polanyi and Scha (1983), shows that this kind of discourse structure influence on referring expressions is not restricted to task dialogues. The linearly nearest referent for the pronoun “them” in the last utterance is the children mentioned in the previous utterance. It is the groceries that were put away though, not the children. This example also illustrates the role of prosody in signaling discourse structure. Intonation is crucial to getting the interpretation of this story right. The second utterance is said as an aside, with strong intonational markers of this separation. Prosody, then, is integral to dialogue processing, and spoken dialogue cannot be handled as a pipeline from speech recognition to pragmatic plan recognition.These fragments illustrate a primary finding of my early dialogue research: dialogues are not linear sequences of utterances nor simply question–answer pairs. This work established that task-oriented dialogues are structured, with multiple utterances grouping into a dialogue segment, and their structure mirrors the structure of the task. Subsequently, Candy Sidner and I generalized from task dialogues to discourse more generally, defining a computational model of discourse structure (Grosz and Sidner, 1986). This model defines discourse structure in terms of three components as shown in the schematic in Figure 5. The linguistic structure comprises the utterances themselves, including their various linguistic features, which cluster into dialogue segments. Just as written language is broken into paragraphs, the utterances of spoken language naturally form groups. The attentional state tracks the changes in focus of attention of the dialogue participants as their conversation evolves. The intentional structure comprises the purposes underlying each dialogue segment and their relationships to one another. Relationships between dialogue segment purposes (DSPs) determine embedding relationships between the corresponding discourse segments (DS), and thus they are foundational to determining changes in attentional state. These three components of discourse structure are interdependent. The form of an utterance and its constituent phrases are affected by attentional state and intentional structure, and they in turn may engender changes in these components of discourse structure.Sidner and I examined two levels of attentional state. The global level, which is portrayed in Figure 5, is modeled by a focus space stack. Focus spaces contain representations of the objects, properties, and relations salient within a discourse segment as well as the discourse segment purpose. When a new dialogue segment is started, a space is pushed onto the stack, and when the segment completes, the corresponding focus space is popped from the stack. This level of attentional state influences definite noun phrase interpretation and generation and intention recognition processes. Access to objects, properties, and relations at lower levels of the stack may not be available at higher levels. The local level of attentional state, which Sidner and I referred to as “immediate focus” (“local focus” being a tongue twister), models smaller shifts of attention within a discourse segment. Sidner (1979, 1981, 1983a) deployed immediate focusing as a means of controlling the inferences required for interpreting anaphoric expressions like pronouns and demonstratives by limiting the number of entities considered as possible referents. Webber’s contemporaneous work on the range of anaphors a single seemingly simple definite noun phrase could yield made clear the importance of being able to limit candidates (Webber, 1979, 1986). It is important to that has been used for various other in the of computational to to discourse and is used to refer to speech in the prosody and to in some short while and ways to a by using about which was in an utterance. defined a of of a in a which strong to immediate used this to on how in the of inferences required to a of the of an utterance into a of the of the discourse of which it was a part and after their paper I was to be by to be a at the of that and I their with discourse into what as which the of of and considered on generation as well as interpretation (Grosz, and Grosz, and of the of on has including and of in and many other in ways are in and development of a range of and have developed a model that with driven and of linguistic markers of discourse and research in this with their work on the intonational of that and with global and local defined in the model Sidner and I developed and and and that and between discourse and of such phrases as example from a talk in other words I will have to the full of the of the now what about state illustrates the importance of determining which use a and I ways in which use to discourse structure (Grosz and with us on determining ways in which with form of referring and attentional and Grosz, and for these of intonational markers of discourse was to from multiple people of speech with the In one of the earlier attempts to develop a of naturally we designed a set of which led multiple to give around the with discourse structure, the was by multiple for using the and and as well as for discourse structure using a based on the model of discourse structure. The design of that can speech of a from multiple remains a one that to be for the of capabilities of personal assistant and component of discourse structure, the intentional structure, has in work on language as in and subsequent work on language as in artificial intelligence 1975; and and Sidner, The example in Figure from an example of the need to the an utterance to In this dialogue is to to of the on the not completely I have used this example it is task-oriented than and so illustrates the general need for this This research on language as several algorithms for and utterances that communicated This though, was on utterances or of and the methods it used were based on research on single and I to these methods to the dialogue but eventually that intentional structure could not be by of single This led us to a computational of collaborative which we (Grosz and Sidner, and I generalized this model to one able to handle an number of and more for (Grosz and in his the to enable of and to (Grosz and research how could be used as the of the intentional structure of dialogue In for the importance of collaboration to language processing has from research in and that has established the importance of to development and language of the of the of collaboration. The in refers to a to an action. In the each of the for is in terms of and those for in terms of can be no intention without and of intention that an be able to it an intention to As a the in various of the carry such a The on that each is required for to by ways in which could fail if the on or that it are not if some not as in those could away from the without their or to inform other if they were to their if they not to each as in (5), they could fail to to each In such the could The the and in this are an of the need for the capabilities for collaboration to be designed into systems from the (Grosz, computational of and differ in the but have of which if not to the to of the or in the might engender dialogue related to it. The dialogue in Figure 2 can be used to the used to model intentional structure and to illustrate the kind of about that is to dialogue structure and utterances in a In the intentional structure is a of some and relationships between these (e.g., that one is as part of determine relationships between the the structure. As the schematic in Figure the expert and apprentice have a to the which as a removing the so is part of the in the is part of the for in the in Figure it that the apprentice believe is able to the This leads the expert and apprentice to have a for the apprentice to how to the have the for Utterances (3) and in the dialogue to this This clarification subdialogue to one of for the of the for an to be able to that action. of clarification dialogue can from the that an to be able to the objects that in an the elided subdialogue about setscrews in Figure 2 is such a clarification With Utterance (5), the apprentice off these and a new one related to pulling off the flywheel. The focus spaces representations of the setscrews are popped from the stack so that only the after Utterance the wheelpuller are in focus when Utterance (8) are the objects relevant to determining the referent of “the screw in the in this era of machine and efforts to build dialogue systems, work on dialogue can for systems design and in a has argued for the of for design of systems in the of dialogue structure and collaboration I have in current dialogue systems, though, is that such systems will not have the of and that work on and dialogue systems have developed ways to limit the of I will briefly to an to the principles these in dialogue systems. I need to the we have used the of collaboration in a to yield design principles for in the context of our work on are multiple ways one can the kind of these can to system or of as by research on and Grosz, and Grosz, can as they have in research on collaborative for and can also an for understanding design or for work deployed in these last two In work with we are methods to improve the in the of children with with of such and several of in this that not only challenges for but also make for a with methods on these we this kind of the has a structure one is in the activities are have and are being and in a not As a have of the of each required by to determining what to Figure shows a between by and the of capabilities they a system to a in of the of about in This led in research to develop an to support for the Grosz, and systems to to people and in a that in an it will be crucial to the structure and right. so will for linguistic form and for Figure some possible ways to use principles to in the of current dialogue systems that are based on machine of various The shows some ways in which dialogue structure can from and from their The second shows ways in which and might give to dialogue structure. of a dialogue underlying into the design of these methods some of dialogue structure and collaboration could improve the as of structure or relations to algorithms and and and some of the of dialogue systems based on of work on into models for in processing also support for this research have my dialogue research from its people and language The of dialogue that I have challenges for current systems. are the that (e.g., chatbots, are to handle than assistants, customer and that short dialogues are than human dialogues, though, are and people frequently of as their shift from those by system build systems of sensibly with people and of ways people talk with one in and the full of language use than systems that people to their language The progress in computational recently for being more in our scientific and challenges I will on can we the fundamental principles of dialogue and collaboration models with can we build models that from one to the required by based systems and the of dialogue machine methods would (e.g., for dialogue in their can we the of dialogue the full range of and can that research are into deployed systems, so that they not make ethical and natural-language processing systems also some of the ethical challenges at the current In to the potential of dialogue system these challenges include the of chatbots on ways people with one and of systems like customer service people in and that is from the included in the is that cannot to an that they they are of This is one dialogue system but current systems The dialogue fragment in Figure 1 is or on of made by current dialogue systems, though, are not or but ethical in the course I mentioned earlier have errors in their of personal assistant systems and in that to be dialogue based on of with each people naturally from of language we would that a system able to the is the nearest would also be able to to can I a and can I go to a there are systems that fail on such seemingly related a of that how to treat a as one system students may not be so for a but a could be were the about a or other as capabilities for collaboration cannot be on but be designed in from the (Grosz, so be into from the of system I will end by readers to about the they the systems they and the they make about those systems from an ethical perspective, and to so from the of their the design of these the people and for this and the I am to the of earlier a that many people work I have from my in computational and in a great about language structure and and their work and my research on my move from areas of computer to by as a topic the of a system that would a story and it from one point of remains so far as I the I that full as they are of and were more to than and to the speech understanding The they to was to build the discourse and components of the speech system. in the early 1970s speech systems efforts that their systems to context into but no one ways to context in a natural-language system. and argued that task-oriented dialogue was than and at at the that speech was than it more along with and a natural-language processing on which I could explore various to dialogue and they a great about and of which I by the speech was an research one that my to my research and also to computer efforts at research has been by with Candy Sidner, a in the model of discourse structure and the model of collaborative with work the for dialogue models with speech and and with a in and of The and with I have have enabled the research advances in this paper, and my academic I have been these to work with and and also with students based in other in dialogue or in collaboration led them to to their and even to I them all for their and for the they to our the talk on which this paper is I also some lessons I learned as my which I have to on to my I include brief mention of them in they will subsequent of students and other early to important and to in investigating though they may at is not to be by the of may have their and be When he about my that it was an but I would dialogue could not be an understanding of was to and I to that it did if one was in language second is to from about the importance of a When Sidner and I first presented that models of for would not for modeling several and we of them developed their we were right. said new questions, new to from a new and advances in and With the of computer systems in and being used to human language for all of features, the importance of the of language and has even than it was when I first the of ways to