NLP in the wild and Neural Summarization

Hi Everyone,

After a long break we are back with our 11th meeting, celebrating three years of Israeli NLP meetups! We will meet at Gong's HQ in Herzliya where we will enjoy two talks: the first by Eilon Reshef who will tell us about building and managing complex NLP pipelines in production, and the second by Tal Baumel who will tell us about his recent research on neural text summarization. See speaker bio's and talk abstracts below.

Looking forward to see you there,
Roee

Talk 1: Building Real-World NLP in Successful Business Applications

At Gong, we have over a dozen different live models in the natural language understanding domain (NLP, audio/speech, and ML), used across hundreds of enterprise customers. We will review how those models are built, evaluated, deployed and used in real-life scenarios. We will outline different types of models (domain-agnostic vs. domain specific, supervised vs. unsupervised, deep learning vs. classical NLP/ML), and cover how data is obtained, how the models are evaluated (pre and post production), and how they’re used for new predictions. We will show case studies of data flows and pipelines, and discuss where humans (general labelers and domain experts) come into the loop.

Speaker Bio:

Eilon Reshef is the CTO and Co-Founder of Gong, the technology leader in the conversation intelligence space, helping sales teams understand what works and what doesn’t. Prior to Gong, Eilon was the VP Product and Co-Founder of Webcollage, a SaaS provider in the channel marketing space, and held several other leadership positions in hi-profile technology companies. Eilon holds an M.Sc. in Computer Science (summa cum laude) from the Weizmann Institute of Science and a B.Sc. (summa cum laude) from the Technion.

Talk 2: Query Focused Summarization Using Sequence-to-Sequence Models

Automatic summarization is one of the many tasks in the interdisciplinary field of natural language processing (NLP). Ever since the introduction of the field by Luhn in the 1950s, state-of-the-art methods of automatic summarization have included extracting sentences from input texts. Such extractive methods capture salient information, but fail to produce fluent and coherent summaries.

Recent advancements in neural methods in NLP have achieved state-of-the-art results in most artificial intelligence benchmarks, including automatic summarization. In this talk, we will survey the field of automatic summarization in general, and focus on neural network methods for the task of automatic summarization. In particular, we critically review the datasets that have been used to enable supervised methods in automatic summarization.

We will then cover the variant tasks of summarization - generic vs. query-focused summarization, single document vs. multi-document and extractive vs. abstractive methods. Neural methods have recently been shown to apply well to single-document generic abstractive summarization under supervised training. We extend this initial step towards abstractive techniques by developing and assessing neural models for multi-document generic summarization and abstractive query-focused summarization.

Our methods combine supervised and unsupervised steps, with new forms of attention-based sequence to sequence neural models and established models of relevance assessment developed in extractive summarization.

We will conclude with an assessment of evaluation methods for summarization -- with proposed tools to assess the ``abstractiveness'' of a summary given the source documents and the fluency of documents generated using neural encoder-decoder techniques.

Speaker Bio:

Tal Baumel is an NLP researcher @ Oath (previously Yahoo Research).
He completed his PhD with Prof. Michael Elhadad at the NLP lab in Ben-Gurion University. Tal's work focused on automatic text summarization and was published in highly refereed venues like ACL and AAAI.