SlugBot: Developing a Computational Model andFramework of a Novel Dialogue Genre

TL;DR

This paper introduces SlugBot, a new computational dialogue model designed for open-domain conversations, supported by a novel ontological resource, aiming to advance chatbot capabilities beyond traditional models.

Contribution

The paper presents a novel Discourse Relation Dialogue Model and its implementation within SlugBot, along with the creation of the UniSlug ontology to support this framework.

Findings

Development of a new computational dialogue model

Creation of the UniSlug ontological resource

Implementation of the model in a modular system

Abstract

One of the most interesting aspects of the Amazon Alexa Prize competition is that the framing of the competition requires the development of new computational models of dialogue and its structure. Traditional computational models of dialogue are of two types: (1) task-oriented dialogue, supported by AI planning models,or simplified planning models consisting of frames with slots to be filled; or (2)search-oriented dialogue where every user turn is treated as a search query that may elaborate and extend current search results. Alexa Prize dialogue systems such as SlugBot must support conversational capabilities that go beyond what these traditional models can do. Moreover, while traditional dialogue systems rely on theoretical computational models, there are no existing computational theories that circumscribe the expected system and user behaviors in the intended conversational genre of the Alexa Prize Bots. This paper describes how UCSC's SlugBot team has combined the development of a novel computational theoretical model, Discourse Relation Dialogue Model, with its implementation in a modular system in order to test and refine it. We highlight how our novel dialogue model has led us to create a novel ontological resource, UniSlug, and how the structure of UniSlug determine show we curate and structure content so that our dialogue manager implements and tests our novel computational dialogue model.

Tables

Table 1. Table 1: Expansion and Comparison Discourse Relations using GeoNames Ontology

Agent	Are you planning on going on any trips soon?.	Dialogue Act: yes-no question
User	yes, we are planning to go to Paris	system recognizes city-name in GeoNames ontology
Agent	Ah Paris, What will you do there?	Dialogue Act: open question
User	I am going to see the Eiffel Tower	system recognizes point-of-interest, go up ontology
Agent	Are you also going to check out the Louvre	Dialogue Act: yes-no question, comparison relation is instantiated by finding a different point of interest for Paris by going down POI ontology
User	I don’t know
Agent	Well, I would think about it. The Mona Lisa is there and that could be worth seeing.	Dialogue Act: provide-opinion, expansion relation is achieved by following artworks relation on Louvre
User	I see. I’ll think about it.
Agent	Are you a fan of any other artwork by Leonardo da Vinci?	Dialogue Act: yes-no question, comparison relation is achieved by following artist relation for the Mona Lisa

Table 2. Table 2: Example utterances with their corresponding intent labels.

Intent	Utterance
request_opinion	did you like beauty and the beast
request_change_topic	no can we do something else
request_opinion_justify	why do you like wine
assertion_on_bot	you are so much better than siri
request_exit	can we stop talking right now
request_service	play country christmas songs
request_discuss_topic	do you know anything about pizza
request_confirm_understanding	are you understanding me

Table 3. Table 3: Summarized list of different modules.

Module	Section	Description
Base Responses	2.4	State-specific responses like handling repeat requests or prompting with a menu.
Opinions	2.5.1	Solicit, provide, and justify opinions about detected entities.
Question Answering	2.5.1	Question answering modules including ELIZA, Evi, Wikipedia, and DuckDuckGo.
Well-being	2.5.1	Detect user well-being class and take different action accordingly.
Intimacy	2.5.1	Gain intimate connection with the user and understand their interests.
Out-of-Domain	2.5.1	Out-of-Domain responses to sustain the conversation if there are no other good options.
Retrieval	2.5.2	Elasticsearch index used to retrieve appropriate responses.
Storytelling	2.5.3	Tell the user a story and answer questions based on a corpus of personal narratives.
Recommendation	2.5.3	Able to recommend different elements of UniSlug, specifically in the travel domain.
Recursive	2.5.3	Facts or trivia triggered recursively until user wants to change topics. Additionally, short 2-turn sequences including riddles and would you rather questions.
Flow Manager	2.6	Managing dialogue flows about a variety of topics (currently, 42) as a way to rapidly increase popular topic coverage.

Table 4. Table 4: Sample of an interaction using our feedback exchange.

Agent	So anyways, We just talked about some would you rather questions. If you don’t mind me asking, did you have fun talking about it? Would you like to do it again some time in the future?
User	no
Agent	I see. I’m bummed you didn’t have fun. I’ll remember this for the future so that we have more fun next time.

Table 5. Table 5: Sample of an interaction using our opinion data.

User	what do you think of watchmen
Agent	I really like watchmen. In fact I think watchmen is action packed because there are lot’s of cool fight scenes. How about you, what’s your opinion?
User	do you like watchmen
Agent	Bot: Yes, I believe watchmen is really interesting. And you? What’s your opinion?
User	What’s your favorite color?
Agent	I really like infrared. It makes me feel warm all over.
User	I like blue
Agent	Why do you feel positively about blue?

Table 6. Table 6: Sample of an Interaction Using our Question Answering

User	What is the capitol city of Mexico.
Agent	The capitol city of Mexico is Mexico City. (search)
User	What is it’s population?
Agent	The population of Mexico City is 8.8 million. (disambiguation + search)
User	Okay, how is it that you are smart?
Agent	Why do you think I am smart? (ELIZA)
User	Just a guess. Tell me a story.
Agent	Did I ever tell you one time my pet Moses really scared me? (enter storytelling)
User	No, what kind of pet is it?
Agent	Moses is a chinchilla. (storytelling answer)

Table 7. Table 7: Sample of interaction using our intimacy modules.

Agent	What characteristics of yourself are you most proud of?
User	My face.
Agent	I see, for me personally you might ask? I have quite speedy natural language processing capabilities, coupled with a vast range of different conversational abilities and actions that I can perform for you. Would you be okay answering another question?
User	No
Agent	Okay, are you interested at all in dinosaurs?

Table 8. Table 8: Sample of an interaction using our Edina and Reddit data.

User	What is your favorite Beatles album?
Agent	My favorite album is "Sargent Pepper’s Lonely Hearts Club Band", however "Abby Road" is a close second. (Edina)
User	I liked Abby Road better than Sargent Pepper.
Agent	I feel like Abbey Road was every bit as chaotic as White Album, yet excellent audio engineering was what truly made it feel cohesive. (Reddit)
User	How many copies did Sargent Pepper sell?
Agent	As of 2011, it has sold more than 32 million copies worldwide, making it one of the best-selling albums. (disambiguation + search)

Table 9. Table 9: Sample of an interaction using recursion on the Science topic.

Agent	Do you want to hear some science facts?
User	Yes
Agent	Did you know that At over 2000 kilometers long, The Great Barrier Reef is the largest living structure on Earth. Want to hear another?
User	Sure why not
Agent	How about this one. The average human body carries ten times more bacterial cells than human cells. Want to hear more?

Table 10. Table 10: Sample of an Interaction Using our ’Would You Rather’ Sequence.

Agent	How about I ask you some would you rather questions?
User	Okay
Agent	Would you rather read books that are nonfiction or would you rather read a steamy romantic novel?
User	nonfiction
Agent	For me personally? I would rather curl up underneath a blanket with a physical copy of a book while sitting on the couch, with a warm cup of tea. It just makes me feel so relaxed and I can forget about my troubles for a while. Do you want to hear another books question?
User	No talk about video games
Agent	If you could be an character from any video game who would you be and why would you pick them?
User	mario from super mario odyssey cause you can throw his hat and go into to that character that you thought you had on
Agent	Cool, for me personally, Nathan drake because he gets to go on lots of cool adventures and has lots of awesome weapons Do you want me to ask another video games question?

Table 11. Table 11: List of the Current Flows Supported by SlugBot.

	Flow Topics
Music	Fun Facts	Harry Potter
Technology	Trivia	Star Wars
News Headlines	Hobbies	Star Trek
Box Office	Holidays	Monsters
Sports	Astronomy	Marvel Cinematic Universe
Fashion	Animals	Pokemon
Shopping	History	Cartoons
Travel	Board Games	Fictional Characters
Nutrition	Books	Tolkien
Health	Language	Science Fiction
Favorite Food	Famous Quotes	Comic Books
Recipe	Poems	Pirates
Gossip	Weather	Video Games
TV	Horoscope	Dinosaurs

Table 12. Table 12: Sample of merged utterances for the input "Tell me about the matrix".

Source	Candidate
Wiki	The Matrix is a 1999 science fiction action film written and directed by The Wachowskis, starring Keanu Reeves, Laurence Fishburne, CarrieAnne Moss, Hugo Weaving, and Joe Pantoliano.
Trivia	In the first forty-five minutes of the film, Neo has eighty lines. Forty-four of these lines are questions, just over half of his total dialogue, averaging at roughly one question per minute.
Opinion	What are your thoughts about the matrix?
Intimacy	Why are you interested in the matrix?
Index	Did you enjoy the follow up movies to the matrix trilogy?
Agent	The Matrix is a 1999 science fiction action film written and directed by The Wachowskis, starring Keanu Reeves, Laurence Fishburne, CarrieAnne Moss, Hugo Weaving, and Joe Pantoliano. What are your thoughts about the matrix?
Agent	The Matrix is a 1999 science fiction action film written and directed by The Wachowskis, starring Keanu Reeves, Laurence Fishburne, CarrieAnne Moss, Hugo Weaving, and Joe Pantoliano. Why are you interested in the matrix?
Agent	In the first forty-five minutes of the film, Neo has eighty lines. Forty-four of these lines are questions, just over half of his total dialogue, averaging at roughly one question per minute. Did you enjoy the follow up movies to the matrix trilogy?

Table 13. Table 13: An evaluation of the four different classes of features in SlugBot. Here our numbers are only considering the user turns.

Module Class	Average User Rating	# User Turns	# Unique Conversations	Average Length
Mixed-Initiative	3.16	38856	21346	1.37
Retrieval	3.08	11722	7353	1.32
System-Initiative	3.58	49571	7853	8.43
Flow Manager	3.35	52472	6274	8.36

Equations

$$\text{def score: }{r}_i.confidence=min(max(context(r_i),r_i.confidence)-loss(r_i),1)$$

$$\text{def loss: }{r}_i.confidence=incoherence(r_i)+repeat(ri)+sentLen(r_i)$$

Full text

1 Introduction

One of the most interesting aspects of the Amazon Alexa Prize competition is that the framing of the competition requires the development of new computational models of dialogue and its structure. Traditional computational models of dialogue are of two types: (1) task-oriented dialogue, supported by AI planning models or simplified planning models consisting of frames with slots to be filled [17, 28, 1, 39, 42]; or (2) search-oriented dialogue where every user turn is treated as a new query that either starts a new dialogue segment or extends current search results. These traditional models and existing systems that are built using them are based on several simplifying assumptions:

• •

Search Model: Conversations are simply sequences of user search queries (initiated by the user) and search results [12, 11, 26];

• •

Task Model: Conversations are composed of sequences of specific tasks such as setting a timer, booking a flight, shopping for specific items [17, 28, 1, 38];

• •

Script Model: Conversations follow a finite-state script and scripts can be written by hand to support all the conversations and conversational variants that a user might want to have with the system [5, 18, 24].

However, the Alexa prize requires the development of an open-domain conversational agent that can talk about any topic and carry on a conversation for at least 20 minutes. This framing means that the dialogue genre of Alexa Prize systems requires not only system development, but the development of new computational and theoretical models of dialogue. It is also important to note that although Alexa Prize bots are characterized as open domain bots, much of the recent work on open domain dialogue has been focused on chat, with attempts for example to train systems using the "Open Subtitles" corpus [35, 27]. In contrast, SlugBot must utilize substantive and up-to-date content on news, movies, books, fashion, technology, news entities, actors, and other topics. This content must be scraped daily from relevant sources and structured in a way that makes it possible to support a coherent conversation about any of these topics. Thus content creation, curation, and structuring is a substantive task all by itself.

Here, in addition to describing the SlugBot system, we describe elements of a new dialogue model which we call discourse relation dialogue model (DRDM). The features of our proposed DRDM are:

• •

Mixed Initiative: Novel dialogue strategies are needed that will allow SlugBot to take the initiative in conversation. It will not be possible to carry on a 20 minute conversation if SlugBot is simply responding to user initiatives as is assumed by the search model above.

• •

Discourse Relations: Novel models of discourse coherence in dialogue are needed. We propose a model based on a framework of Penn Discourse TreeBank discourse relations [22, 21, 29, 37]. Discourse coherence in task model dialogue systems arises from the structure of the task. Discourse coherence in search model dialogue systems, such as it exists, are driven by the user’s search intentions. Discourse coherence in script model dialogue systems is created by the user interaction designer rather than being an instantiation of an underlying theory of discourse coherence.

• •

Knowledge Graph: A large ontology with specific world-knowledge can provide DRDM with dialogue strategies based on general, re-usable relations between conversational turns. We describe how we have developed UniSlug, an ontology based on integrating the schemas of several existing ontologies, which we use to drive dialogue strategy selection as well as natural language understanding [19].

The fact that there are no existing theoretical or computational models for the Alexa Prize dialogue genre cannot be overstated. Novel methods are needed to drive the system behaviors while ensuring discourse coherence. These methods should be general so that they can be systematically applied to subconversations on different topics or user activities. Task-oriented dialogue models assume that coherence arises because both the system and user can recognize the intentions of their conversational partner as contributing to the completion of the task, or as meta-dialogue related to organizing contributions to the completion of the task [17, 28, 1]. While some dialogue segments in SlugBot can be modeled as tasks, the overall dialogue structure is not task-related. Previous theoretical work on conversation merely describes aspects of conversational structure without considering algorithms or models that can drive the behavior of a conversational agent [16, 22]. We will describe below how we build on previous models of discourse coherence, but it is also important to note that models of discourse coherence using discourse relations have mainly been applied to highly structured texts such as newswire or student essays [10, 34, 31, 36, 33]. There is no large annotated corpus showing how discourse relation models could be applied to dialogue: previous projects in this vein are merely exploratory or focused around a few examples [40, 3, 22].

We will describe in more detail in Section 2 how we have combined the development of a novel computational theoretical model, DRDM, with its implementation in a modular system in order to test and further develop the model. We highlight our extended efforts at content curation and content structuring and describe how we have built new knowledge bases and a novel dialogue manager to support our novel computational dialogue model.

2 System Design and Architecture

The architecture of SlugBot is driven by the need to support DRDM, our novel computational model of open-domain dialogue. The DRDM dialogue model relies on UniSlug, a new large domain ontology that we have built and integrated into SlugBot. The DRDM model is based on two ideas. First, we claim that it is possible to model the coherence of open-domain dialogue using discourse relations, specifically we currently utilize the four high level discourse relations used in the Penn Discourse TreeBank (PDTB), which provide good generalization capabilities and are compatible with other discourse relation frameworks. These are framed in terms of relations between abstract objects (as realized by sentences or clauses). Here we apply these to relations between utterances across the agent and the user in discourse.

The top level Penn Discourse TreeBank (PDTB) discourse relations that we utilize are [36]:

• •

expansion: The expansion class covers those relations which expand the discourse and move its narrative or exposition forward. Its subclasses include instantiating a set, restating, describe alternative situations and more. This is the most general and weakest discourse relation, since it covers both continuing to talk about the same thing, as well as talking about a more specific attribute of an entity.

• •

comparison: The comparison class applies when a discourse relation is established between Arg1 and Arg2 in order to highlight prominent differences between the two situations. Its subtypes are contrast and concession. Disagreements can be viewed as a type of contrast.

• •

contingency: indicates that one of the situations described in Arg1 and Arg2 causally influences the other. For example, SlugBot may offer an opinion, along with the reasons underlying it. Opinions are causally related to (justified by) these reasons.

• •

temporal: this relation applies when the situations described in the arguments are related temporally, either in overlap or in a temporal sequence. The main use of the temporal relation in SlugBot is in the context of telling stories, where we take advantage of the fact that story events are told in temporal sequence.

The second idea is that these discourse relations, along with dialogue acts and named-entities, can be used to guide retrieval of utterances from the many different sources of content. Other existing retrieval based chatbots also operate on large existing corpora such as Twitter [35, 20], the Open Subtitles corpus [13], or movie scripts [4, 2], but none of them use either discourse relations or dialogue acts. Instead, the criteria by which utterances are retrieved has been based on their similarity to the current system utterance or a previously existing reply to the current system utterance. Similarity measures have been adopted from information retrieval, e.g. they include both TF-IDF and word-embeddings.

Instead our retrieval mechanism, as described in more detail in Section 2.5.2 is controlled by a combination of dialogue acts, discourse relations and named-entity matching. The discourse relations in DRDM can be instantiated by different types of dialogue acts, such as questions or statements as shown in Table 1, thus the dialogue act that instantiates the discourse relation must be specified in the dialogue flow. Table 1 illustrates how these relations can be instantiated in the DRDM. The RHS column provides a dialogue act and discourse relation specification, as well as an informal description on how the discourse relation instantiation relies on UniSlug, our new mondo-domain ontology.

Figure 1 illustrates the UniSlug ontology that we have built for SlugBot by integrating several existing ontologies using the schema from schema.org. We then use this ontology for both understanding and dialogue management. The main way that the ontology is used for dialogue management is illustrated in Table 1.

2.1 Architecture

Figure 2 describes the complete system architecture and Sections 2.2 - 2.8 describe the individual components such as natural language understanding, dialogue management, data management, and realization components. In Section 2.9 we will specifically examine our Amazon Skill infrastructure as shown in Figure 9.

2.2 Natural Language Understanding (NLU)

2.2.1 Preprocessor

ASR Filter: Our initial preprocessing stage involves analyzing the tokens interpreted by the Automatic Speech Recognition software (ASR). We calculate the average hypothesis scores of the users input and prompt the user for clarification if the score is too low. Intuitively, it is better to ask for clarification rather than misinterpret the user input; however, if we consistently get a low ASR score, we are forced to estimate their utterance. To account for this, we retain all possible ASR interpretations such that we are able to better process the noise in their input.

Profanity Filter: Our profanity filter looks for keywords that typically indicate profanity or possible offensive language. This filter has expanded over time as we became aware of offensive content in Wikipedia, which we thought was a pretty benign source of information but in fact contains references to inappropriate content. In future work, we wish to use our profanity filter to indicate among other things user frustration. Currently, however, we just use the profanity filter to eliminate inappropriate content from the candidate response pool.

2.2.2 NLU Modules

CoreNLP Parser: After preprocessing the data we use our Natural Language Understanding (NLU) engine to create a deep structure representation of the user’s utterance. Our first layer of NLU relies on the Stanford CoreNLP Toolkit [32]. Our internal representation is based on the dependency parse of the respective utterance which is consolidated into a concise tree using the dependency relations. The part-of-speech (POS) and sentiment score from CoreNLP are also encoded into this structure. We do coreference resolution by mapping the coreference tags returned by CoreNLP to the data stored within our system.

Entity Detection: We have also developed our own named entity recognizer SlugNERDS [9, 7] because the existing named entity recognizers were not fine-grained enough to support dialogue interaction. SlugNERDS is based on the Google Knowledge Graph111https://developers.google.com/knowledge-graph/ and can take advantage of the fact that the Google Knowledge Graph is both robust and constantly updated, allowing us to consistently detect the newest and most obscure named entities.

Intent Classifier: Both our NLU and our indexed retrieval mechanism rely on a dialogue act classification. We develop an utterance intent ontology and develop a Neural Network model to recognize user intents. The intent ontology consists of 33 discrete intents. Some example utterances with their associated intents can be seen in Table 2. The ontology is designed to allow each intent to be recognized without conversational context. In other words, it is assumed that each utterance’s intent can be determined irrespective of where they occur in a conversation.

We use a subset of the CAPC dataset to train the intent classification model. Utterances were selected from the CAPC dataset and then annotated to provide input to a supervised learner. This resulted in a dataset of roughly 50K annotated utterances which is then separated into training (80%), development (10%), and test (10%) subsets. Next, we train an intent classifier on the training set, tune the hyperparameters on the development set, and use the test set for final evaluation of the model. Our intent classifier is a Neural Network model that uses a combination of Recurrent Neural Network and Convolutional Neural Network architectures.

Additional NLU: Additional NLU components in our pipeline include the NPS dialogue act classifier[15]. We further do noun and entity disambiguation to know that, for example a lion is an animal and Watchmen is a 2009 American superhero movie. Our topic classification takes into account a broad classifier which is trained using topic annotated news articles in addition to the provided cobot topic classifier. We additionally take into account a more refined topic classifier which directly maps to the 42 topics supported by the Flow Manager described in section 2.6.

2.3 Data Management and UniSlug

One of the greatest challenges of developing an Alexa Prize Bot is the need to collect and curate content from a wide variety of sources to cover a wide range of topics of interest. Figure 3 specifies how we have sourced content from our own Mechanical Turk HITs as well as the data collected by the 2017 Alexa Prize Edina team [25], as well as sourcing content from specific Reddit subreddits, story corpora such as Aesop’s Fables, trivia websites, and news sources among others. All of this content needs to be processed for named entities and dialogue acts in order to be used by our dialogue manager. We also use sections of the Amazon provided FUD and CAPC datasets to motivate our topical and functional expansion. This content is used in combination with search engines as described below in more detail.

UniSlug: In addition to these content sources, we host the UniSlug graph database illustrated in Figure 1 on EC2. This is a key part of our architecture and it must be accessible in real time since it is used both for NLU and for dialogue management. UniSlug consists of 275GB of ontological knowledge base data from various sources, consisting of one billion relations and over 100 million entities. This allows us to flexibly reference and navigate semantic relations as well as expanding the knowledge that SlugBot uses in the system initiative modules described in Section 2.5.3.

We distinguish between two types of memory when we actually utilize this content: short term memory (STM) and long term memory (LTM).

The STM cluster is responsible for managing our system state, representing the current discourse context, and handling data which is localized to specific functionality. This local memory helps us to improve the efficiency of data access and reduce the workload of remote databases.

The LTM cluster is responsible for managing our corpora and other large datasets. We utilize the Amazon Relational Database Service to store real-time search data in addition to other curated content that is used by various modules.

LTM and STM communicate with each other to exchange data: the data which is out-of-date should transfer into the LTM while the conversationally relevant data should be made available in the STM. Both memory clusters are comprised of a network of memory nodes which each have their own responsibilities.

2.4 Dialogue Manager

The dialogue manager is tasked with handling the most basic functionality, such as detecting repeat requests, stop requests, and prompting the user with a menu of topics to help transition into domains with a breadth of coverage. We maximize the number of unique experiences within a single conversation by prioritizing unexplored topics. We found that having a menu made an appreciable difference in conversational quality. Additionally, the dialogue manager will periodically ask the user for feedback on a recent set of turns or apologize for delays in response time. When designing our feedback responses, we made sure to phrase them such that they emphasize improvement for the user, not SlugBot’s experience. A potential exchange can be seen in example 4.

The dialogue manager also sets system expectations. Specifically, we indicate the data representing the preconditions which must be satisfied in order for a particular action to be valid. Our expectations can be satisfied by observing attributes from the system’s state. We can define our expectations using a variety of state variables including direct keyword matching, specific utterance attributes such as dialogue act or sentiment score, or even be contingent on the result returned by a function.

2.5 Dialogue Modules

Table 3 lists the dialogue modules implemented in our system, with descriptions of each module in the following sections. Our dialogue manager maintains control over the dialogue modules, allowing us to easily change the type of responses which we pool. We feature primarily two module classes, mixed initiative in the case where our agent does not have control of the conversation, and system initiative when the agent is driving the conversation. The ranking algorithm which is used to pick the best candidate module is described in Section 2.7.

2.5.1 Mixed Initiative Modules

Mixed initiative modules are designed to solicit an even exchange in content per turn from both the user and the agent. More specifically, this is the case in which the agent is not strictly in control of the conversation and in fact, our content is likely to be a reaction to the user’s initiative. Such user initiative could be asking a question, soliciting an opinion, or having general chit-chat outside of a controlled dialogue flow.

Opinions: Here we can learn more about the user by soliciting their opinion of a contextually relevant entity. Naturally, if we can solicit user opinions, it’s important that we are able to provide and justify our own opinions. To accomplish this, we load our agent’s profile with opinions about various entities and abstract concepts. The first time a user engages with the system we randomly select opinions to include in the agent’s profile, allowing us to take the preliminary steps towards giving the agent a unique and identifiable personality. Our opinion dataset is extracted from online reviews of movies, video games, and books. We have also handcrafted data points which would give us good general coverage of conceptual opinions based on popular topics which would be difficult to extract from any data source, such as “What is your favorite color?”. Finally, we have also collected full opinions, and justifications of existing opinions using Mechanical Turk. When collecting new opinions we focused on positive, non-controversial topics. In Table 5 we have synthesized an example which leverages our structured opinion data to answer solicitations from the user.

Question Answering: Our question answering mechanism is a three step inspection of the query. First, if there are not enough content words to detect the intent of the question, we probe for more information using a modified version of ELIZA [43]; we found that users generally reacted poorly to some of the more intimate prompts when used out of context. If the system has the initiative, such as when we are telling a story, we assume the associated module will be able to answer questions using the module’s structured data. If neither of these conditions are true or we don’t yet have an answer, we perform coreference disambiguation and query our search engine. Our search engine consists of three search components which are queried in the following order: Evi222https://www.evi.com/, Wikipedia333https://www.wikipedia.org/, and finally DuckDuckGo444https://duckduckgo.com/. If we have still not returned an answer, we inform the user that at this time we are not able to answer the question. We noticed through anonymous user feedback that, in general, users were okay with our system not being able to answer questions as long as the agent made it clear that it was not sure of an answer. An example demonstrating this module is shown in Table 6.

Well-being: Our well-being module is designed to understand the user’s basic mood through sentiment analysis and a series of questions. We learn from the conversation history that, besides positive and negative, the major emotions which effect our dialogue strategy would also include: neutral, bored, hostile, and offended. Table 5 and Figure 5 show the sample user utterances after our agent asks "How are you doing?". The Bored users tend to say "I don’t know" when the agent asks other questions, the Offended users are very sensitive to personal questions, the Hostile users are aggressive and tend to give us low rating. Therefore, our agent would take different actions accordingly. For example, we avoid asking to many question to the Bored users but give more direct suggestions. We avoid asking further personal questions to the Offended and Hostile users. As for the Positive users, we would let them lead the conversation, while for the Negative users, we would suggests some resources in our dialogue system to improve their well-being.

Intimacy: In addition to learning more about the user’s well-being, SlugBot attempts to gain a more intimate connection with the user by asking intimacy questions and understanding their interests. Our intimacy questions are handled using a similar strategy as our other 2-turn recursive questions. Each question has been annotated with an intimacy level, which we use as a filter when curating content. As our conversation with the user progresses, we relax this constraint allowing SlugBot to ask more intimate questions. We also try and gauge the user’s interests based on our unexplored system modules, and the user’s context. The former case allows us to surface new content, while the latter allows us to leverage the context of the conversation to gain further understanding of the user and surface contextually relevant turns.

2.5.2 Dialogue Management Driven Retrieval

Our retrieval module is constructed using an Elasticsearch555https://github.com/elastic/elasticsearch based index, which has advantages of fast full text search capability and scalability. We built our indices on dialogue based, fact based and social media data with different search criteria.

An additional data sources, tailored towards a two-party open domain conversation, is a crowd-sourced corpus of conversational turns on a variety of topics using qualified workers on Mechanical Turk. From this data collection, we gathered 17,364 topic annotated turns. In addition to our crowd-sourced data collection we further supplemented our index by handcrafting responses which would have only been captured by observing them in our user trials, as discussed further in Section 3.

We also incorporate trivia, fun facts, jokes, would you rather questions and hypothetic questions, consisting of 5,439 high quality utterances, with Elasticsearch index, it retrieves utterances when the users are interested in specific topics or entities.

We also utilized the Reddit API666https://github.com/praw-dev/praw selectively collecting from 17 subrreddits resulting 38,172 Reddit posts. We built indices and customized queries based on the topic, keywords, entities and meta fields. Moreover for better quality of responses using Reddit posts, we examined the data and applied filters to filter out posts that are not appropriate for dialogue turns. We also used the corpus that Ben Krause et. al. published as part of their Edina project [25]. We extracted every utterance of their published data and applied our own set of filters.

We first ran the utterances through our own pipeline that determined the topic based in part on Google knowledge graphs. Our queries then used our filters to select responses based on topic. We removed utterances that contained any type of anaphora we could identify. This included pronouns, specific temporal indicators, and agreement indicators. Sentences with pronouns could be referring to someone or something that occurred in a prior utterance. Specific times are also potentially problematic because events that occurred at a specific time, such as a game or a concert, do not generally reoccur at the same time. We also carefully examined the utterances to develop subtle rules for removing potentially confusing or inappropriate comments.

Out-of-Domain: Using a combination of these modules we also designed our out-of-domain responses for when we are not confident in any specific answer. For example, if we are able to detect an entity but do not have a direct follow-up utterance, we can either give an opinion, ask the user for more information, verify the entity by asking about one of its related synonyms, or use our search engine to try and retrieve some possible summary-based response. If our system does not have above 0.8 confidence in our top-ranked answer, we employ a transition strategy which attempts to engage the user with an unexplored module or topic. A sample fallback utterance assumes we have not yet discussed the user’s well-being and is as follows: I’m really sorry about this, but I’m not sure what to say next. My robot brain is telling me I’m not quite sure how to respond to what you just said. Do you mind if I switch to a new topic? How is your day going?

2.5.3 System Initiative Modules

Our system initiative modules gain control of the conversation by assuming the role of content provided and driving the conversation. These modules are designed to encourage extended multi-turn interactions by routing the user back to the module when the conversation starts to stray.

Stories: Stories are often told conversationally, and in a two-party setting it is natural that one speaker will assume the responsibility of content provider while the other will assume the role of recipient. This can lead to an entertaining experience for the recipient, while also minimizing the conversation’s reliance on their content. Therefore, if we are able to replicate this environment, we hope to create an engaging multi-turn experience. As described in previous iterations of SlugBot [8], we incorporated a large corpus of personal blogs [30], from which we built a small dataset of around 40 narratives which are interesting and do not contain inappropriate content. After closely inspecting the results of our preliminary evaluation, we realize that this content was not received highly as anticipated. We attribute this to 1) the content being boring, or so unrealistic for an Echo device’s personal narrative, and 2) improper control of the stopping points throughout the story led to random pauses and allowed the user to ask more questions which we could not answer.

To address these issues we feature three different sources of story content this year. Fables, dreams that SlugBot had, and SlugBot’s adventures as an Echo device. Our fables content is provided by a large corpus of annotated versions of Aesop fables [14]. Both the dreams and the personal adventures data we crowd-sourced using Mechanical Turk. All three of these data sources allow the user to sufficiently suspend their disbelief that such an embodiment could be in the story, and since we targeted popular entertainment franchises and topics when curating the data, we believe it is much more engaging than our previous implementation. Further, we have annotated our stories in chunks, where each chunk ends with a natural pausing point. We appended tag questions to the end of these chunks in order to implicitly discourage the user from asking us hard questions we are not capable of answering. Finally, we used SSML markup to fix issues with timing and prosody in addition to giving the stories a sense of dramatic flare.

Recommendation: In daily life, conversation on topics such as movies, books, travel etc., people would like to recommend and discuss similar movies, another related book, or other interesting places that they visited while traveling. This inspires us to give recommendation on such topics so that it not only improves the depth of the conversation but also makes a better user experience. To be able to understand how entities are related and retrieve such information, we use UniSlug, as mentioned in Section 2.3. We first identify the entities mentioned, then we explore adjacent vertices in the graph by various relations.

Our primary implementation of this feature is in the travel domain. For example, if the user wants to travel to New York, we can recommend some famous places in New York. Or if the user plans to visit the Eiffel Tower, we can recommend some other tourist spots that are also located in the same area. Based on the relations in the graph database, we are able to curate our responses most appropriately.

Recursive: Finally, we have a set of modules which acts recursively. Here we can inform the user of various headlines from new sources, or give the user facts about a topic of their choice. We are able to recurse over this functionality by simply continuously giving them information until they explicitly transition out of the recursion. We have included an example of this in Table 9. We can also create recursive 2-turn sequences by asking the user a sequence of riddles, would you rather questions, and hypothetical questions. All three of these cases allow the user and agent to converse for a couple of turns briefly about the sequence before recursing, an example of this can be seen in Table 10. We collected approximately 1,500 question/answer pairs using crowd-sourced labor across each of the topics supported by the flow manager, as discussed in Section 2.6. These recursive loops are highly effective at keeping the user engaged with the agent in a multi-turn context without having to worry about a complex dialogue flow. Moreover, having topic-annotated content allows us to fluidly connect related content from SlugBot’s modules.

2.6 Flow Manager

One of the primary modules in our system, the flow manager, is responsible for managing the flow of dialogue related to a given topic or utility. A flow, as seen in Figure 2, is organized in a graph structure where each node has specific preconditions, postconditions, and actions which work together to sustain a natural dialogue about any particular subject. While flows in general are meant to provide coverage of some specified root topic, like Books, it’s important to note that many flows contain multiple subroots, such as "favorite genre", "book trivia", or "bestsellers". The user can directly trigger a flow about a given topic by using related keywords, or by expressing interest in the flow if the prompt is selected by the system when propositioning topics.

Our preconditions are represented by the expected data discussed in Section 2.4. Postconditions can represent a variety of desirable effects which only occur after the response has been realized by the dialogue manager. It is within these postconditions that we can indicate calls to external functions or update specific state variables. Finally, there are actions which also occur. These actions can modify a candidate utterance or delegate responsibility for response curation to a different module.

Flows represent a high-level abstraction of our entire system’s functionality, allowing a new designer to rapidly add content to the system without needing to familiarize themselves with the underlying architecture. As seen in Table 11, our system currently supports 42 flows covering a diverse range of topics. Since these flows represent a high-level abstraction of the entire system, we found that reusing successful modules is an effective means of bootstrapping flows with minimal effort. Specifically, most flows have some recursive trivia-based prompt in their list of subroots. We also found that generically discussing user preferences and utilizing search methods increased the breadth of a flow, while a combination of all methods could increase the depth of a flow.

Another means of increasing the conversation depth, used for example in the Nutrition flow, is by preparing a hierarchical knowledge base that is recursively navigated by the flow. The Nutrition flow uses such a knowledge base of arguments supporting various nutrition facts extracted from Healthline Nutrition777https://www.healthline.com/nutrition, with many references across different nutrition topics. The flow starts by offering a controversial fact to the user, and then, based on the user’s reaction (such as agreeable, inquisitive, or negative), it responds with a supporting argument, counter-argument, another related nutrition fact, or possibly an information related to the user’s follow-up question. In the spirit of the DRDM’s method generalizability, this recursive flow design, paired with a suitable knowledge base, can be easily reused for other topics.

2.7 Ranking

Once we have established a pool of responses, we re-rank them to find a response we are most confident in. Each response has a confidence score which ranges from 0 to 1. Each response is assigned a base confidence value which is increased and decreased based on contextually information. We have established a base confidence value of 0.6, which is a value we have carefully selected to work best in our system. While some of the following metrics are automatically calculated, this base confidence in addition to the weights assigned to each metric are manually assigned.

If the user says I like video games for example, the video game conversation starter will have a confidence of 1, because our system is very confident it is the best next response. If the user said I like dogs, the video game prompt would have a confidence of 0.6, indicating it as a valid topic starter, but only if we have nothing more relevant to say. At this stage, our sensitive content filter will invalidate any response with explicit content and detect if a priority response has been triggered. Priority responses are valid regardless of our current state and indicate responses which are to be uncontested - such as repeat requests or stop session markers. Finally, for all other responses we update their confidence using Equation 1. We attempt to increase our confidence in the response by looking for contextually relevant content and inspecting the current system state. Our context score is calculated by considering overlapping content words and entities in addition to our system’s state variables.

Equation 2 represents how we penalize a given response. A response is in a state of incoherence if it does not belong to the current system initiative module. For example, if we are playing a specific game, and a response stems from anywhere besides that game it would create incoherence within the conversation. In order to maintain module coherence we apply an empirically derived 0.15 penalty to these responses. While we leverage the state tracking done by our Short Term Memory to avoid repetitious utterances, some general prompting phrases such as “would you like to play a game?” are still valid despite being already said. In order to increase the diversity of a user’s experience without limiting the variety in our response pool we apply a 0.05 penalty to prompts which have already been explored. Furthermore, we noticed from our own experimentation that long utterances from mixed initiative modules tended to be received poorly. An example of this includes long news headlines and overly verbose indexed responses, where we have limited control over the phrasal timing. We therefore applied a length based penalty to these utterances.

[TABLE]

[TABLE]

In the case that more than one response shares the highest confidence score, we resolve ties by randomly choosing among the candidates. It should be noted that in most cases candidates which are tied for the maximum score all tend to realize the same intention.

2.8 Postprocessing

Stylistic Variation: When realizing our top response candidate, we use a statistical natural language generation engine [6] to add variation in a way which seems to increase the coherence of the response. Our most common use case is prepending our response with an acknowledgement hedge - enabling a more natural transition between topics. For example, when the user is giving an opinion, the agent’s response will begin with “I see” or “Oh really”. Additionally, we gathered specific user intents by analyzing our logged data to add specific hedges to the start of the utterance when appropriate. For example, if the user says “thank you” or “are you kidding” we will prepend our response with “You’re welcome” and “I kid you not” respectively.

Response Merging: It is commonly the case that multiple good responses will be scored highly by our system. With certain types of content it is possible to merge these responses to give SlugBot a richer response. After reading through the user data, it seems clear that always ending our utterance with a question retains some control over the conversation and invokes follow up content from the user. Therefore, we will frequently merge a fact or statement with a question about the same topic or entity. An example of this output can be seen in Table 12.

SSML: At this stage we also apply any SSML markup which is encoded in the response. We found that SSML seemed to improve certain sections of dialogue extremely well. Specifically, we primarily used it for 1) adding additional pauses in sentences, 2) correcting the way certain words are interpreted by the TTS engine, 3) adding dramatic flare to the various stories our system is capable of telling.

2.9 Web Application Architecture

Figure  9 depicts the architecture of our web server in details.

Flask Web Application: We apply the Flask web application structure to facilitate communication between our framework and AWS Lambda. We consider exception handling an important part of our system design. Not only do we keep the daily user interaction logs in EC2 for manual inspection, but we also found the CloudWatch Alarms useful for error notifications.

Elastic Load Balancer: Elastic Load Balancer was easier for us to incorporate in our existing framework than Elastic Beanstalk, even though Elastic Beanstalk is convenient for deployment. To validate our load balancing, we use the Locust888http://locust.io/ load testing framework to expose potential hazards.

Natural Language Understanding (NLU): Our NLU utilities, specifically CoreNLP, runs on its own server and could represent a performance bottleneck especially in the case of long utterances. Therefore we run these NLU utilities on a separate EC2 instance which communicates with the web server.

Data Resources: We use a vanilla RDS Relational Database (Aurora MySQL) with only one replication to store our data resources. We create index for the frequent query table to reduce latency.

UniSlug Graph Knowledge Database: Our primary knowledge base, UniSlug, is running on a separate EC2 instance to increase the performance of real time queries.

Retrieval Utility: We are running an Elasticsearch server on a separate EC2 instance to support our retrieval based modules and future reinforcement learning in real time.

Natural Language Generation (NLG): We have a set of EC2 instances which are serving our neural language generation model, Slug2Slug. We use multiple instances in order to query and ensemble the results of multiple models as quickly as efficiently as possible.

3 Discussion

In Table 13 we present an evaluation of our four different classes of dialogue modules as described in Sections 2.5.1 - 2.6. This data was based on the user feedback received from August 1st to August 28th, for which our system’s average user rating was 3.31.

As predicted, when we relinquishes more control to the user it becomes harder to respond appropriately. This is reflected by the fact that our average conversational score for both mixed-initiative and retrieval-based models is lower than our average system performance. We also notice that both strategies are only used slightly more than 1.3 user turns per unique conversation. This makes sense, as providing continuous follow-up content to retrieved or search-based content can be difficult.

Our highest rated class of responses originate from our system-initiative modules, as described in Section 2.5.3. Moreover, we see that on average a conversation which utilized these modules contained more than 8 user turns of related content, which represents multiple games or stories told in the same conversation. This indicates that while the user is individually contributing less overall to the dialogue, driving a conversation with topically relevant content in an entertaining fashion is still an effective dialogue strategy capable of engaging users for a multi-turn conversation.

Finally, we look closer at our flow manager, which is responsible for sustaining general chit-chat conversation about the 42 different topics described in Section 2.6. Our flow manager received an average score of 3.35, and was similarly able to surface this content for over 8 user turns on average. This indicates to us that our flow management scheme is an effective dialogue management policy. Moreover we’ll note that creating new flows is a streamlined process, allowing us to easily scale and create baseline conversations about low data topics, such as famous monsters or dinosaurs.

As noted in the previous iteration of this competition, it is clear that users still desire that the bot assumes both the responsibility of a personal assistant and a conversational partner [8]. Users want to adjust the volume of the device, play a specific song, or perform other standard Echo skills. If this type of fluid integration were possible, it would yield an improved overall experience.

Another issue that we experienced was responding to some user intents. When the user is hostile, it can be difficult, if not impossible, to come up with a good response that the user enjoys. Some user intents are currently off-limits. For example, many users seemed to want Alexa to engage in inappropriate conversations. We obviously cannot just switch to an R-rated dialogue because of a keyword, which may not even be the user intent. Many other users have very strong opinions about the current administration and wish to discuss it, but we cannot give any opinions about it in order to avoid offending people with the opposite view. Some users want to push the artificial intelligence boundaries by saying things like, "I love you." or "You are stupid." Responses to these and other non-task oriented utterances are not easily handled through conventional dialogue systems. We had to make specific responses for these intentions.

Sometimes, the confidence of the user’s utterance from the ASR is low, and our system does not know how to respond to an utterance if the confidence is below a certain threshold. This threshold changed throughout the competition as we tried to find the right balance of giving a response which might not be appropriate and not asking the user to clarify their statement.

On the other end of the NLP pipeline, we experimented with a deep-learning approach to language generation. Slug2Slug [23], our sequence-to-sequence model that we upgraded to use the Transformer [41] architecture, is trained to generate natural utterances from structured meaning representations (i.e. lists of attribute-value pairs). This, however, limits the model to more task-oriented sections of the conversation, which comprise a rather small fraction of an average open-domain interaction with a user. Although we were able to adapt the model to multiple domains through data collection and transfer learning, it would have required considerably more data to make the model robust and versatile enough to be used ubiquitously in place of templates. In order for the system to be able to utilize the model outside of task-oriented domains, we considered implementing a universal meaning representation factory for the system’s utterances, which, however, would require even more data annotation. The inference was feasible in real time, but ultimately, due to the above limitations in usability, we decided to forgo the deep learning model and continue using templates. The latter currently provides more personalization and customization power, as well as more flexibility when adding the support for new domains to the system.

4 Future Work and Conclusion

In this paper, we have presented our expansion of a scalable socialbot framework and our contribution to the 2018 Amazon Alexa Prize Competition. Additionally, we have described a new dialogue model, DRDM, which is uniquely defined for this task. We have further expanded on our flow-based representation of dialogue, with the intention of increasing its scalability in future iterations of SlugBot.

To further increase the performance of our re-ranker, we incorporate reinforcement learning based on the user feedback in the Alexa Prize data. We defined our decision-making problem as a Markov Decision Process problem. We trained our reinforcement learning model on different subsets of states so as to find the most efficient set of states for real-time ranking. Currently, we are evaluating the results of this process to determine the optimal policy.

We have adapted SlugBot’s dialogue strategy to provide more user-centric functionality, specifically designed to gauge the interest and well-being of the user. We feel that establishing a more intimate relationship between our socialbot and the user are the next steps towards creating a better, more human-like conversation. User’s emotions, such as being happy, sad, or bored, can have a strong influence on the user’s interaction with any conversational agent. Adapting and responding to these emotions will more humanize the conversational agent and increase user engagement and satisfaction. Hostile or angry emotions need a different set of responses, which we are still exploring. We are currently working to increase our understanding and ability to adapt and respond to human emotions with the goal of eventually being able to empathize and sypathize with the user as well as provide insightful and meaningful comments to improve the emotional state of the user.