• Chatbots: The (Long) Tail Wags the Dog -- Talk by Arun Chaganty, PhD

    Arun Chaganty (Head of AI, eloquent.ai) will give a talk on how recent advances in neural models for sentence similarity and human-in-the-loop annotation can work to power a chatbot in a customer-facing environment. Food and drinks will be provided when the doors open at 6:30PM, doors close at 7:00PM sharp, and the talk begins at 7:00PM. No attendees will be able to enter the building after 7:00PM, so arrive early! Interested in presenting your work at a future Bay Area Research in NLP and ML Meetup? Click here: https://goo.gl/forms/xtn1CrhyhLQk1D5i1 -- Abstract of Arun's talk: Give a customer a chatbot meant to help them track packages, and they’ll ask it to change their delivery address, send them proof of delivery and cook them dinner. In general, customers expect chatbots to handle a large quantity of very specific tasks, and each task requires custom, time consuming, hard coded API integrations. That’s not scalable. In this talk, I’ll go over the novel AI and UX techniques that Eloquent Labs uses to tackle this problem -- and how we’re using our approach to make chatbots work in customer facing environments. First, we will discuss how we apply recent advances in neural models for sentence similarity and human-in-the-loop annotation to robustly support 100s of new intents in a zero-shot learning framework. Second, we will explain how we generate mimic rephrasals to simulate empathetic listening while conveying to the user that we are unable to solve their queries (yet). -- About Arun: Arun is currently the Head of AI at Eloquent Labs, a conversational AI company building chat bots for customer service. Not that long ago, Arun graduated with a PhD in computer science at Stanford University, working with Percy Liang and the Stanford NLP group. -- About this Meetup group: Organized by the AI team at Mya Systems, this Meetup group is looking to bring together people testing and iterating novel approaches for potential applications in Artificial Intelligence. Come network, listen to what folks are working on, and learn from interactive sessions with experts in NLP & ML. -- About Mya Systems: Founded in 2012, Mya Systems brings deep learning and NLP expertise together to disrupt the recruiting operational model. See more at www.mya.com.

  • Sharp Nearby, Fuzzy Far Away: How Neural Language Models Use Context

    Urvashi Khandelwal (NLP Group, Stanford University) will give a talk on the role that linguistic context plays in the behavior of neural language models. Food and drinks will be provided when the doors open at 6:30PM, doors close at 7:00PM sharp, and the talk begins at 7:00PM. No attendees will be able to enter the building after 7:00PM, so arrive early! Interested in presenting your work at a future Bay Area Research in NLP and ML Meetup? Click here: https://goo.gl/forms/xtn1CrhyhLQk1D5i1 -- Abstract of Urvashi's talk: We know very little about how neural language models (LM) use prior linguistic context. In this paper, we investigate the role of context in an LSTM LM, through ablation studies. Specifically, we analyze the increase in perplexity when prior context words are shuffled, replaced, or dropped. On two standard datasets, Penn Treebank and WikiText-2, we find that the model is capable of using about 200 tokens of context on average, but sharply distinguishes nearby context (recent 50 tokens) from the distant history. The model is highly sensitive to the order of words within the most recent sentence, but ignores word order in the long-range context (beyond 50 tokens), suggesting the distant past is modeled only as a rough semantic field or topic. We further find that the neural caching model (Grave et al., 2017b) especially helps the LSTM to copy words from within this distant context. Overall, our analysis not only provides a better understanding of how neural LMs use their context, but also sheds light on recent success from cache-based models. -- About Urvashi: Urvashi Khandelwal is a fourth year PhD student in Computer Science at the Stanford NLP Group, where she is advised by Professor Dan Jurafsky. Her research interests lie at the intersection of natural language processing and machine learning. More specifically, she is interested in analyzing and improving models of language generation to facilitate generalization. -- About this Meetup group: Organized by the AI team at Mya Systems, this Meetup group is looking to bring together people testing and iterating novel approaches for potential applications in Artificial Intelligence. Come network, listen to what folks are working on, and learn from interactive sessions with experts in NLP & ML. -- About Mya Systems: Founded in 2012, Mya Systems brings deep learning and NLP expertise together to disrupt the recruiting operational model. See more at www.mya.com.

    3
  • The Natural Language Decathlon: Multitask Learning as Question Answering

    Bryan McCann and Nitish Keskar (Salesforce Research) will give a talk on decaNLP and their new Multitask Question Answering Network (MQAN), which jointly learns all tasks in decaNLP without any task-specific modules or parameters. Food and drinks will be provided when the doors open at 6:30PM, doors close at 7:00PM sharp, and the talk begins at 7:00PM. No attendees will be able to enter the building after 7:00PM, so arrive early! Interested in presenting your work at a future Bay Area Research in NLP and ML Meetup? Click here: https://goo.gl/forms/xtn1CrhyhLQk1D5i1 -- Abstract of Bryan and Nitish's talk: Deep learning has improved performance on many natural language processing (NLP) tasks individually. However, general NLP models cannot emerge within a paradigm that focuses on the particularities of a single metric, dataset, and task. We introduce the Natural Language Decathlon (decaNLP), a challenge that spans ten tasks: question answering, machine translation, summarization, natural language inference, sentiment analysis, semantic role labeling, zero-shot relation extraction, goal-oriented dialogue, semantic parsing, and commonsense pronoun resolution. We cast all tasks as question answering over a context. Furthermore, we present a new Multitask Question Answering Network (MQAN) jointly learns all tasks in decaNLP without any task-specific modules or parameters in the multitask setting. MQAN shows improvements in transfer learning for machine translation and named entity recognition, domain adaptation for sentiment analysis and natural language inference, and zero-shot capabilities for text classification. We demonstrate that the MQAN's multi-pointer-generator decoder is key to this success and performance further improves with an anti-curriculum training strategy. Though designed for decaNLP, MQAN also achieves state of the art results on the WikiSQL semantic parsing task in the single-task setting. We release code for procuring and processing data, training and evaluating models, and reproducing all experiments for decaNLP. -- About Bryan McCann and Nitish Keskar: Bryan McCann is a Research Scientist at Salesforce Research in Palo Alto where he works on Deep Learning and its applications to Natural Language Processing. He is particularly interested in the connections between natural language tasks, how knowledge can be transferred between them, and how to develop more general models for NLP. Nitish Keskar is a Senior Research Scientist at Salesforce Research in Palo Alto where he works on Deep Learning and its applications to Natural Language Processing and Computer Vision. He is particularly interested in efficient training methods and issues pertaining to generalization and scalability. -- About this Meetup group: Organized by the AI team at Mya Systems, this Meetup group is looking to bring together people testing and iterating novel approaches for potential applications in Artificial Intelligence. Come network, listen to what folks are working on, and learn from interactive sessions with experts in NLP & ML. -- About Mya Systems: Founded in 2012, Mya Systems brings deep learning and NLP expertise together to disrupt the recruiting operational model. See more at www.mya.com.

    1
  • Building chatbots that scale using machine learning and open source

    Justina Petraitytė (Rasa, http://www.rasa.com/) will give a talk on an open-source AI framework Rasa has created for developers looking to build Conversational AI systems from the ground up. Food and drinks will be provided when the doors open at 6:30PM, doors close at 7:00PM sharp, and the talk begins at 7:00PM. No attendees will be able to enter the building after 7:00PM, so arrive early! Interested in presenting your work at a future Bay Area Research in NLP and ML Meetup? Click here: https://goo.gl/forms/xtn1CrhyhLQk1D5i1 -- Abstract of Justina's talk: AI assistants are getting a great deal of attention from the industry as well as the research. However, the majority of assistants built to this day are still developed using a state machine and a set of rules. That doesn’t scale in production. In this talk, Justina will introduce you to two open source frameworks developed by Rasa - Rasa NLU and Rasa Core, which enable developers to create AI assistants solely based on real conversational data and machine learning. During this talk, you will learn about the machine learning behind these libraries and what is the process of building clever assistants using Rasa frameworks. -- About Justina: Justina is a Developer Advocate at Rasa where she focuses on developer education, growing the Rasa open source community and making sure that developers have the resources and tools to build great AI assistants with Rasa. She has a background in Econometrics and has been working as a Data Scientist in the video games industry for a few years before joining Rasa. -- About this Meetup group: Organized by the AI team at Mya Systems, this Meetup group is looking to bring together people testing and iterating novel approaches for potential applications in Artificial Intelligence. Come network, listen to what folks are working on, and learn from interactive sessions with experts in NLP & ML. -- About Mya Systems: Founded in 2012, Mya Systems brings deep learning and NLP expertise together to disrupt the recruiting operational model. See more at www.hiremya.com.

    8
  • CoQA: A Conversational Question Answering Challenge by Siva Reddy, Stanford NLP

    Siva Reddy (Computer Science Postdoc, Stanford University) will give a talk on CoQA, a novel dataset released last month for building conversational question-answer systems. Food and drinks will be provided when the doors open at 6:30PM, doors close at 7:00PM sharp, and the talk begins at 7:00PM. Interested in presenting your work at a future Bay Area Research in NLP and ML Meetup? Click here: https://goo.gl/forms/xtn1CrhyhLQk1D5i1 -- Abstract of Siva's talk: Humans gather information by engaging in conversations involving a series of interconnected questions and answers. For machines to assist in information gathering, it is therefore essential to enable them to answer conversational questions. In this talk, I will present our work on CoQA, a novel dataset for building Conversational Question Answering systems. CoQA contains 127k questions with answers, obtained from 8k conversations about text passages from seven diverse domains. The questions are conversational, and the answers are free-form text with their corresponding evidence highlighted in the passage. We analyze CoQA in depth and show that conversational questions have challenging phenomena not present in existing reading comprehension datasets, e.g., coreference and pragmatic reasoning. We evaluate strong conversational and reading comprehension models on CoQA. The best system obtains an F1 score of 65.1%, which is 23.7 points behind human performance (88.8%), indicating there is ample room for improvement. This is a joint work with Danqi Chen. We launch CoQA as a challenge to the community at https://stanfordnlp.github.io/coqa/ -- About Siva Reddy: Siva Reddy is a postdoc in Computer Science at Stanford University working with Prof. Christopher Manning. His research focuses on enabling natural communication between humans and machines. Prior to the postdoc, he was a Google PhD Fellow at the University of Edinburgh under the supervision of Prof. Mirella Lapata and Prof. Mark Steedman. -- About this Meetup group: Organized by the AI team at Mya Systems, this Meetup group is looking to bring together people testing and iterating novel approaches for potential applications in Artificial Intelligence. Come network, listen to what folks are working on, and learn from interactive sessions with experts in NLP & ML. -- About Mya Systems: Founded in 2012, Mya Systems brings deep learning and NLP expertise together to disrupt the recruiting operational model. See more at www.hiremya.com.

    9
  • Talk: A deep dive into natural language inference by Amir Kargar, ML Engineer

    Amir Kargar (Machine Learning Engineer, Mya Systems) will give a talk that picks apart the inner workings of state-of-the-art ML models trained for natural language inference. Food and drinks will be provided when the doors open at 6:30PM, doors close at 7:00PM sharp, and the talk begins at 7:00PM. Interested in presenting your work at a future Bay Area Research in NLP and ML Meetup? Click here: https://goo.gl/forms/xtn1CrhyhLQk1D5i1 -- Abstract of Amir's talk: Natural Language Inference (NLI) is a subtask of Natural Language Understanding (NLU) which gets two inputs, hypothesis and premise, in the form of text and determines the relationship between the two, in terms of entailment, contradiction, and neutral. Most of the state-of-the-art literature on NLI shares the same paradigm of encoding, attention, inference, and classification. In this talk, we will go over how each of these steps are done in detail and how each contribute to the performance of the end-to-end network. We will also talk about how the same paradigm could be applied to other NLP tasks such as paraphrase detection or semantic similarity analysis. --- About this Meetup group: Organized by the AI team at Mya Systems, this Meetup group is looking to bring together people testing and iterating novel approaches for potential applications in Artificial Intelligence. Come network, listen to what folks are working on, and learn from interactive sessions with experts in NLP & ML. -- About Mya Systems: Founded in 2012, Mya Systems brings deep learning and NLP expertise together to disrupt the recruiting operational model. See more at www.hiremya.com.

    6
  • Boosting NLP with reinforcement learning: tuning & grounded language learning

    Andy Mullenix (Machine Learning Engineer, Mya Systems) will give a talk on new approaches in NLP that use reinforcement learning to push state-of-the-art performance across a number of language tasks. Food and drinks will be provided. Interested in presenting your work at a future Bay Area Research in NLP and ML Meetup? Click here: https://goo.gl/forms/xtn1CrhyhLQk1D5i1 -- Abstract of Andy's talk: Many tasks in NLP have seen new state-of-the-art results by incorporating reinforcement learning (RL) into the training process. Summarization, machine translation, and dialogue learning have all benefited from its inclusion. Simultaneously, RL is being used in new ways to rethink language learning itself in multi-agent systems and grounded learning environments. We will explore these cases to see how RL can be used today to improve model performance, as well as consider open questions in research. --- About this Meetup group: Organized by the AI team at Mya Systems, this Meetup group is looking to bring together people testing and iterating novel approaches for potential applications in Artificial Intelligence. Come network, listen to what folks are working on, and learn from interactive sessions with experts in NLP & ML. -- About Mya Systems: Founded in 2012, Mya Systems brings deep learning and NLP expertise together to disrupt the recruiting operational model. See more at www.hiremya.com.

    5
  • A New Approach to Natural Language Transfer Learning by Sanjana Ramprasad

    Sanjana Ramprasad (Machine Learning Engineer, Mya Systems) will give a talk discussing her new approach to natural language transfer learning, which uses structural ontological information and thesaurus knowledge to add information to existing word representations. Food and drinks will be provided. Interested in presenting your work at a future Bay Area Research in NLP and ML Meetup? Click here: https://goo.gl/forms/xtn1CrhyhLQk1D5i1 -- Abstract of Sanjana's talk: Recent trends in NLP tend towards transfer learning. Improvements have been shown in classification tasks at the word level by using transfer learning from machine translation (Learned in translation: Contextualized Word Vectors; Mcann et al ., 2017) and in using language models to learn syntax, usage and semantics of words (Deep contextualized word representations; Peters et al). While the above two use language models and machine translation, I will introduce a new transfer learning approach that uses more structural ontological information and thesaurus knowledge to add information to existing word representations. Specifically, we obtain multi word prototypes. We first show how this improves performance on several word similarity and relatedness tasks as compared to single prototype words. We also show this approach outperforms SOTA multi-word representations and ontology grounding techniques and also compare our approach with the above mentioned transfer learning approaches in extrinsic tasks. Thus this talk will introduce existing trends of transfer learning in NLP and a new approach to the same. --- About this Meetup group: Organized by the AI team at Mya Systems, this Meetup group is looking to bring together people testing and iterating novel approaches for potential applications in Artificial Intelligence. Come network, listen to what folks are working on, and learn from interactive sessions with experts in NLP & ML. -- About Mya Systems: Founded in 2012, Mya Systems brings deep learning and NLP expertise together to disrupt the recruiting operational model. See more at www.hiremya.com.

    2
  • Talk: Bridging NLP and CV together, combining representations for diverse tasks

    Emmanuel Ameisen, AI Lead at Insight Data Science, will give a talk at our June Meetup titled "Bridging NLP and CV together, combining representations for diverse tasks." Food and drinks will be provided. Interested in presenting your work at a future Bay Area Research in NLP and ML Meetup? Click here: https://goo.gl/forms/xtn1CrhyhLQk1D5i1 --- Summary of this talk: Many problems combine Natural Language Processing and Computer Vision. Common ones include image and video captioning, image search, and reading text from videos. The approaches combining CV and NLP developed to solve these problems extend naturally to other domains, such as selecting the best thumbnail from videos and generating code from sketches of websites. Having led over 70 AI projects (including the ones above) at Insight, Emmanuel will go over some common approaches and pitfalls, diving deeper into some of the challenges that need to be overcome to get these models to work. --- About Emmanuel Ameisen: Emmanuel Ameisen is the AI Lead at Insight. Emmanuel has years of experience going from product ideation to effective implementations. At Insight, he has led over 70 AI projects in a variety of domains, with Fellows that have gone on to join top ML teams around the country. Previously, he implemented and scaled out predictive analytics and machine learning solutions for Local Motion and Zipcar. Emmanuel holds master’s degrees in artificial intelligence, computer engineering, and management from three of France’s top schools. --- About this Meetup group: Organized by the AI team at Mya Systems, this Meetup group is looking to bring together people testing and iterating novel approaches for potential applications in Artificial Intelligence. Come network, listen to what folks are working on, and learn from interactive sessions with experts in NLP & ML.

    2
  • NIPS and EMNLP De-brief & Presentation: Multimodal Learning with Modal Attention

    Done submitting your papers to NIPS and EMNLP? Come hear a presentation by Daniel Chun (Machine Learning Engineer, Mya Systems) titled "Multimodal Learning with Modal Attention". Daniel is working on solving GuessWhat?!, which is a cooperative game designed by a team led by Harm de Vries at University of Montreal in order to test and push the limits of AI in solving problems that require understanding of both vision and language. Food and drinks will be provided. Interested in presenting your work at a future Meetup? Click here: https://goo.gl/forms/xtn1CrhyhLQk1D5i1 -- Overview of Daniel's talk: In each instance of this game, there are 3 players, the Questioner, the Oracle, and the Guesser, and there is an image, which all 3 players can see. At the beginning, the Oracle chooses an object in the image but does not reveal the choice to other two players. The Questioner then asks a series of questions to the Oracle in an attempt to get information about the object. The Oracle answers each question with either yes, no or N/A. When the Questioner feels like there has been enough question/answer pairs to guess what Oracle has in his mind, a list of candidate objects in the image is given to the Guesser, who then guesses which one of them the Oracle had in mind all along. The game is a success if the Guesser chooses correctly. As one might think, solving this game will require coming up with and training models that can not only understand language and visual cues separately, but also understand them together. There has been quite a few papers attempting to solve this game. Some use traditional machine learning approaches like supervised learning with maximum likelihood estimation, while others use newer approaches like reinforcement learning. However, in all of these papers, the researchers do not feed in the image information to the Guesser and the Oracle models. Instead, they only feed information about the category and the location of the object(s). In fact, they all find that adding the image information actually degrades performance. This is most likely the fault of the models themselves and not the fault of too much information being available. This is where my research comes in. I'm looking for ways to incorporate the image information using both spatial attention and modal attention techniques in order to correct this common fault among the approaches tried by others so far. Modal attention will allow my models to be able to decide which modalities to pay attention to given a question. For example, a question like "Is it in the left side?" can be answered using only the location of the object, and in fact paying attention to other modalities such as the image or the category label of the object will hamper the model by introducing noise. Spatial attention refers to being able to decide "where" to look in the image, which is important since not every image-related question will require looking at every part of the image. For example, a question like "Is it on the counter?" will require only looking at the object and the counter and not the rest of the image, and in fact, paying attention to rest of the image will hamper the model by introducing noise. Although spatial attention has been looked at in a variety of different contexts within the machine learning research community, modal attention is a concept that has not been explored much in the community, so I'm excited to see what it can bring to the table. --- About this Meetup group: Organized by the AI team at Mya Systems, this Meetup group is looking to bring together people testing and iterating novel approaches for potential applications in Artificial Intelligence. Come network, listen to what folks are working on, and learn from interactive sessions with experts in NLP & ML. -- About Mya Systems: Founded in 2012, Mya Systems brings deep learning and NLP expertise together to disrupt the recruiting operational model. See more at www.hiremya.com.

    5