Intelligent Conversational Agents for Ambient Computing SIGIR 2022 Ruhi Sarikaya Amazon Science.pptx

Editor's Notes

• #2: Ack Gabriella, org committee, SIGIR. Thank you for joining the talk. The summary of the talk is captured in the title itself. We want to enable natural contextual interactions for ambient computing, to do that we need scalable self-learning methods that can handle ambiguity and context for accurately understanding the user’s request and provide the best possible answer. The backdrop for the talk is conversational systems like Alexa. Let us get right to it.
• #4: We celebrated Alexa’s 7th birthday on last Saturday, on Nov 6th. Here you see some of the Alexa Echo family of devices starting launched over the years starting with the first generation. Note that of these devices has specific name, Alexa is agent or AI behind it and it lives in the cloud.
• #5: Alexa AI and devices are not built just for your homes, as Amazon we partnered with other companies to bring Alexa to senior living communities, hospitality business as in Marriott hotels, hospitals , here you see Boston Children’s Hospital. Timothy Driscoll, Director of Technology Strategy at Boston Children’s Hospital, says, “Boston Children's Hospital is using Amazon Echo devices to provide an array of features to patients including entertainment in the form of music and games, hospital and unit-specific frequently-asked questions, and control of the in-room televisions. Our patients will soon be able to express their needs for things like pain management support - "Alexa, tell my nurse I'm in pain," or general comfort - "Alexa, tell my nurse I need a pillow."  Alexa experiences are augmented with end-point specific experiences as call nurses, hospitals related frequently asked questions as in hospitals. Two weeks ago also launched Alexa in Disney resorts,. Alexa is also integrated into infotainment systems of most of the car manufacturers, There is not a single week that goes by that we do not announce something about Alexa or Alexa integration with partners. Just last week also announced Alexa on Range rovers. What is really happening?
• #6: Let us talk about the basics first. How do we sense the world around us? We (as humans) have our senses and we primarily rely on our vision and hearing. These are our primary means of getting input from the outside world. Let us look at computers and how they sense the world? How do they get input from the outside world and particularly from humans? Let’s look at the slightly longer horizon and go to 1946, this is ENIAC the mother of modern day computers. This picture is taken at Upenn. The Apple iPhone is different -- many of the elements of its multi-touch user interface require you to touch multiple points on the screen simultaneously (capacitive screens vs resistive screens) We have been manually pushing our wills into our tools, literally by using our hands. ENIAC programmers Frances Bilas (later Frances Spence) and Betty Jean Jennings (later Jean Bartik) stand at its main control panels.  With ENIAC's 40 panels still under construction, and its 18,000 vacuum tube technology uncertain, the engineers had no time for programming manuals or classes. Bartik and the other women taught themselves ENIAC's operation from its logical and electrical block diagrams, and then figured out how to program it. They created their own flow charts, programming sheets, wrote the program and placed it on the ENIAC using a challenging physical interface, which had hundreds of wires and 3,000 switches. Eniac was impressive: 80 feet long and 8 feet high, weighing 30 tons, with 18,000 vacuum tubes, 70,000 resistors, 10,000 capacitors and 1,500 relays. It had a memory capacity of 20 words and was programmed by setting 6,000 dials and switches, a task that took a crew of workers many hours. The Apple-1 computer, built by hand in 1976 by Steve Wozniak in Apple co-founder Steve Jobs' garage or his sister's bedroom, fetched nearly twice its pre-sale high estimate, Bonhams said. - Nowadays a typical cell-phone has over 40 sensors: accelerometers, gyroscope, thermometers, wifi signals, Bluetooth, RF signal gathering etc.
• #8: Low power and cheap sensors are integrated into devices, physical systems and buildings in different industry verticals, from transportation to health care to factories. Likewise on consumer side, more and more sensors are integrated into many devices and appliances we use everyday. It is estimated that there will be about 1 trillion sensors in the world, that is about 120 sensors per person in the world. These sensors measure changes in the environment or in user’s state from temperature, to pressure, touch, camera, light etc. to user’s heart rate to physical movements. More specifically, with GPS data coupled with camera and microphone, coupled with internet connectivity, we are building a digital nervous system for the environment. This paves the way for an intelligent and responsive environment
• #15: The fundamental challenge of Verity is that we need to convert customer feedback, which is oftentimes binary or low dimensional (‘yes/no’, ‘defect/non-defect’), into high dimensional predicted labels such as transcription, NER annotation. The key strategy to do so is through the “hypotheses” → “exploration” → “confirmation” process. i.e. Verity system generates or selects potential alternative hypotheses for the labels we want customer’s feedback on, injects those hypotheses into Alexa runtime either through implicit exploration (e.g. query rewrite, or nbest ranking) or through explicit exploration (e.g. ask customer “do you mean frozen the movie, right?”), then collects customer feedbacks for these hypotheses to generate high confidence synthetic labels Verity are composed of two major components: 1) Exploration Module and 2) Feedback Collection and Label Generation Module. The responsibility of the exploration module includes: 1) Confidence prediction. Predict confidence of the current turn in order to decide whether or not to trigger exploration; 2) Hypothesis generation. Generate alternative hypotheses for the labels that Verity want to collect customer feedback on; 3) Exploration. Inject the hypotheses into Alexa runtime to explore, either through: a) implicit exploration i.e. directly replacing existing 1-best hypothesis with alternative hypotheses through systems, or b) explicit exploration i.e. presenting multiple hypotheses for customers to choose from through mechanisms such as voice confirmation and on-screen dynamic feedback (e.g. presenting a list of candidate transcriptions and ask “which of the following transcriptions matches what you said? ”) The responsibility of the feedback collection and label generation module is to collect/interpret customer feedback, and convert those information into high quality synthetic labels that could be used for training respective models. The feedback collection & understanding module collects user feedbacks such as implicit feedbacks (barge-in, paraphrases), explicit feedbacks (e.g. “did I answer your question?”, “yes”). and unsolicited feedback (e.g. positive feedback: ”alexa, thank you”, negative feedbacks “alexa, that is not what I said”). The Multi-task label generation model is a multi-task deep-learning based model that labels data leveraging contexts derived from customer feedbacks.
• #16: Verity Label Generation Model architecture: Turn level encoder + Dialogue/session level transformer Turn level textual encoder: RoBERTa MLP for categorical/numerical features Input features include: Textual features (user utterance, alexa response, previous turns, next turns) Categorical features (domain, intent, dialog status) Numerical features (num of tokens, asr/nlu confidence scores) Raw audio data Training Multi-task learning heads. Using explicit/implicit user feedback Manual annotation Self-supervision: Synthetic contrastive learning 
• #17: Left panel: Results for Goal Segmentation and Evaluation. We use single-pass human annotation as baseline, and compares two variants of our models: 1) fine-tuned with single task separately; 2) fine-tuned in a Multi-task setting. Note that we use three-pass “Gold” annotation as ground-truth. DA profile: professionally trained human annotators who has passed annotation quality bar; Results are broken down to two subsets: 1) “Single turn” whereas each goal contains only one turn, 2) “Multi-turn” whereas each goal contains a dialogue of multiple turns. We show three metrics respectively: 1) Segmentation accuracy, 2) Combined (segmentation and evaluation) accuracy; 3) Combined F-1 score (weighted average of F-1 scores for three goal classes “success”, “failure”, “not actionable”). “Data support” provide the size of test dataset. Key takeaways: Human is more accurate on goal segmentation task but models outperforms human on goal evaluation task. Combining goal segmentation and evaluation, model is slightly better than human in terms of Accuracy but slightly worse than human on weighted F1 score (due to class imbalance, human has higher precision than model, which contributes to human’s higher F-1 but lower accuracy). Multi-task generally outperforms single-task. This is evaluated on full dataset, model outperforms human on high confidence subset (50% highest confidence data). Right panel shows results for Intent Classification and Slot tagging tasks. Here we uses a RoBERTa model without dialogue context as baseline, and compare our model (RoBERTa with dialogue context) against the baseline. For Slot tagging task, we use RoBERTa + CRF as baseline Key takeaways: Context based model outperforms baseline on all intents and slot types. Context based model brought significant improvements on specific intents and slot types, e.g. 12% relative improvement for “CheckOrderStatusIntent” and 7% relative improvement for “ShoppingListType” Data is only single-pass so not compared to human
• #19: This slide describes how the rewrite system (FLARE) fits into the spoken dialog system and coordinates with other parts to achieve defect reduction. We further introduced two pipelines to deliver rewrites within the “Reformulation Engine” (FLARE): (1) Precompute (offline modeling) (2) Online. The text description covers the concept of each pipeline, and their cons and pros:Precompute: Pros: (1) no latency constrains and complex model can be utilized; (2) rich offline signals that is not available during online mode, e.g. user follow up turns, user rephrase, Alexa response, metrics such as CPDR, Music30sPrecision, Video-Click-Through-Rate Cons: (1) can’t utilize online contextual information, e.g. previous turns, screen-text, etc. (2) can not correct long-tail queries Online: Pros: (1) better utilization of online contextual signal to improve the rewrite quality, e.g. previous turns, screen-text, ASR-nBest, etc; (2) capture rewrite opportunities for long tail queries; Cons: (1) latency constrains; (2) absent of offline information, e.g. NLU hypothesis, CPDR metrics, Alexa’s response to user query, etc. Guardrails and metrics (CPDR) is not covered in this slide Existing publications from Alexa that can be cited (which is already in production) Precompute: “Self-aware feedback-based self-learning in large-scale conversational AI”, Ponnusamy et al., to appear in NAACL 2022 “Contextual Rephrase Detection for Reducing Friction in Dialogue Systems”, Wang et al., EMNLP 2021 “Feedback-based self-learning in large-scale conversational AI agents”, Ponnusamy et al., IAAI 2020 Runtime “Personalized Search-based Query Rewrite System for Conversational AI”, Cho et al., NLP4ConvAI 2021 “Search based self-learning query rewrite system in conversational AI”, Fan et al., De-MaL 2021
• #20: The slides provides a neural-based approach to extract rephrase from session data. The left side provides an example session where rephrase can be extracted. The right side provides the model architecture about how it will encode the session information and output a span detection that indicates the rephrase. Example description (left side): The first turn (“play tyler hero explicit”) and the second turn (“play tyler hero explicit by jack harlow”) is defect (marked as ‘RED’ ) The first turn is defect because user is rephrasing; the second turn is defect because Alexa responds “Sorry, I can’t find that” The third turn (‘play tyler hero by jack harlow’) is non-defect (marked as ‘Green’). And the first turn and the third turn construct a rephrase pairs that can be used for rewriting purpose The fourth turn (“play record year by eric church”) is non-defect (marked as “Green”). We added this turn to illustrate that the session contains rephrase and non-rephrase (e.g. user swtich topics), and the model needs to learn to differentiate it Model Architecture description (right side) Description of the task: We consider a dataset D of M multi-turn dialogue sessions, such that D = {Si}, i=1..M, and every session S is an ordered set of N turns: S = {(Qi,Ri)}, i=1..N. Here i indicates the index of turn, and each turn i consists of a pair (Qi,Ri), where Qi is the user’s query and Ri is the agent’s response to query Qi. Any two successive turns have a time gap of less than a minute. Given a dialogue session S and a source turn, i.e., input pair of query and response (Qi,Ri), the goal of our model is to predict whether Qi is rephrased in any of the following turns (Qj,Rj)| i < j ≤ N. If so, the model should predict the span of Qj and return null otherwise. Input: input is truncated to a maximum of 512 tokens and a maximum of 10 turns Time-bin is used to encode time intervals between turns. Consider a source turn (including a request and a response) t_src = (Q_src,R_src), for which we want to detect a rephrase in the session. We refer to its timestamp as ω_src. We calculate the time difference ∆i = ωi − ω_src, where ωi is the timestamp of a turn ti, for all the turns in the session. ∆i ∀i ∈ [1,n] are then mapped to their respective time-bin tokens. These time-bin tokens represent equal sized intervals in ∆’s range of [-60, 60] seconds. The model we used are listed as BERT, but it can also be RoBERTa Output: We cast rephrase detection as a span prediction problem where we predict the probability of start and end span locations on each token’s position, using the embedding output of the f inal BERT layer. We introduce a start vector W_S and an end vector W_E (both W_s and W_E will be trainable parameters) Assuming the final hidden vector for the ith input token as Ti, the score of a candidate span from position i to position j is defined as: s_ij = WS · Ti + WE ·Tj, where i < j. We use snone = W_S ·T_CLS +W_E · T_CLS to represent the score of no-rephrase span. We set threshold τ to decide whether to predict no-rephrase or not. If maxj>i s_ij > s_none + τ, then we regard the maximum score span as the rephrase span and null otherwise. We use some heuristic post filtering to generate the final rewrite pairs launched in production
• #21: Markov based query rewriting model that learns from recurring customer rephrase patterns.
• #22: This slides describe about how the DFS system work (using retrieval/rank models to generate rewrite instead of relying on customer’s own rephrasing). The ‘routine’ work was reflected in the “Customer routine phrase” injection, which will be used to build personalized index. Motivation of the work: Enable long tail traffic rewrite We generate rewrite based on customer habitual usage patterns with the agent. The global layer is added to (1) avoid over index to personalized cases (e.g. a user who has strong affinity of ‘we don’t talk any more’ might have interest to explore a recent new song ‘we don’t about Bruno any more’; (2) to correct rewrite that has not shown in the user’s interaction history but are popular queries among global users The example on the top right showed that: (1) user gave a query that is ambiguous and as ASR errors, the system find two possible rewrite from both global and personalized layer, and chose the personalized layer in the end The Retrieval model uses a Dense Passage Retrieval (DPR) models that extract embedding for index and for query respectively and use some similarity measurement to decide the rewrite score The Ranking model combines fuzzy match (e.g. through a single encoder structure) with various meta data (e.g. impression, CPDR, etc.) to make a reranking decision 
• #24: Key Takeaways: Customers can instantly personalize Alexa in the moment by teaching her (as opposed to waiting for Alexa’s ML models becoming smarter over time)
• #26: Modern dialog assistants are complex systems that process user requests in multiple stages(see Figure 1). First, a voice trigger (or wake word) model determines whether the user is speaking to the assistant. Following the trigger component, an ASR module converts user audio stream into text. This text is sent to NLU component which determines the user request. ER system recognizes and resolves entities, and the system generates the best possible response (Result stage) using several sub-systems that are specific to each dialog assistant. Finally, the response is rendered into a human-like speech using a Text-to-Speech (TTS) system. In order to keep improving such assistants it is important to identify defects at scale. The manual analysis to identify defects and root causes of defect is infeasible large traffic volumes. In our system are are not only detecting system defects, but we are identifying which component of the dialog assistant is responsible for this defect. It is important to note an error in an upstream component (e.g. ASR) can propagate through the system to the final response to the user. In such cases it is likely that multiple components fail. Thus, we focus on the first component that fails in a way that is irrecoverable, which we call the “failure point”. In this work we recognize five error points as well as a “correct” class (meaning no component failed). The possible Failure points are: FW (errors in voice trigger), ASR Error (errors in use speech transcription), NLU Errors (IC+DC errors, for example incorrectly routing play Harry Potter to Video or music), ERR errors (Entity Resolution and Recognition), Result Errors (incorrect result, for example playing wrong Harry Potter movie).
• #27: To be better illustrate Failure Point Problem, let's examine this multi-turn dialog. In the first turn, the user is traying to open a garage door, however the conversational assistant didn't recognize user's speech correctly and it thought that user wanted to open a "garbage door". Entity Resolution system didn't recover from this error and also failed. Finally, the dialog assistant fails to perform the correct action. In this turn, ASR is the failure point. We don't mark Entity resolution as the failure point even though it also failed, because the ERR might have worked in case of correct ASR. On the second turn user repeats the request. ASR makes a small error by not recognizing article "the" in the speech, Dialog Assistant Makes a correct action, hence we would mark this turn as correct, as the ASR Error didn't lead to the system failure. The last turn highlights one of the limitation of our method. The user is asking the dialog assistant to make a sandwich, which is an action dialog assistants can not do today. All systems have worked correct, however the user is not satisfied. In our work, we do not consider such turns as defective, however the sentiment-based approaches would mark this as unsatisfactory. By the way, in theory, the system expectation might have been to invoke a “pleasantry/joke style response”, in which case this would have indeed been a defect even from system perspective.
• #28: Our best Failure Point Isolation model achieves close to human performance on average across different categories (>92% vs human). This model uses extended dialog context, features derived from logs of the assistants (e.g., ASR confidence), and traces of decision making components (e.g., NLU intent). The weakest performance is in FalseWake (FW) class at 70%. We outperform humans in Result and Correct class detection. ASR, ER, NLU is in the 90-95% range.