Semanlink - Ville AND J'y étais

- Chit-chat (casual, non goal oriented) open-domain. Must be relevant to the context and diverse (informative) to be engaging.
- why creating a chat? to prove an AI can speak like a human, commercial reasons, link to services.

Task oriented vs non task oriented: this tutorial is about the second one.

Retrieval based vs generation based.

Basic knowledge of DL for chatbots:

- word embeddings
- sentence embeddings (CNN, RNN)
- dialogue modeling: seq-to-seq with attention

Response selection for retrieval based chatbots:

- single turn response selection (slides 37-57)
    - framework 1: matching with seq embeddings
    - framework 2: matching with message-response interaction (46)
    - extension of 1: KnowledgeMatching with External Knowledge (53)
    - extension of 2: RepresentationsMatching with Multiple Levels of Representations (54)
    - insights from comparison between 1 and 2 (57)
- multi turn response selection (62)
    - context is now: mess + history
    - again, 2 frameworks

Emerging directions (79):

- matching with better representations
    - Self-Attention (82)
    - fusing multiple types of repr. But how to fuse matters (83)
    - pre-training

Learning a matching model for response selection (84)

Generation based models for chatbots:

- single turn generarion (89)
    - Basic generation model
        - seq2seq
        - Attention
        - Bi-directional modeling
- multi turn generation
    - Contexts are important
    - Context sensitive models
    - Hierarchical context modeling
    - Latent variable modeling
    - Hierarchical memory networks

Diversity in conversations (99)

Content introducing (106)

Additional elements (113)

- Topics in cnversation
- Emotions

Persona in chat:

- Persona
- ...
- Knowledge
- Common sense

RL and Adversarial learning in conversations (125)

Evaluation (132)

Future trends:

- Reasoning in dialogues
- X-grounded dialogues

2018-11-06 About

Joint Models in NLP - Slides - Tutorial (EMNLP 2018) - Yue Zhang

Tags:

**Joint models: solve 2 tasks at once.**

Related tasks: POS tagging, NER, chuncking. Pipeline tasks

Motivations:

- reduce error propagation
- information exchange between tasks

Challenges:

- Joint learning
- Search

History: statistical models. 2 kinds:

- Graph-Based Methods
    - Traditional solution:
        - Score each candidate, select the highest-scored output
        - Search-space typically exponential
- Transition-Based Methods
    - Transition-Based systems: Automata
        - State: partial result during decoding, Action: operations that can be applied for state transition
        - Output constructed incrementally

- Deep learning based model
    - Neural transition based models
    - Neural graph-based models
        - Cross task
            - Seminal work: Collobert, Ronan, et al. "Natural language processing (almost) from scratch."
            - Not all tasks are mutually beneficial
            - Ramachandran, et al.  “Unsupervised pretraining for sequence to sequence learning.”
            - Peters, Matthew E., et al. "Deep contextualized word representations." (ELMo)
            - "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding."
            - ULMFIT
            - Correlation between multi-task learning and pretraining
        - Cross lingual
        - Cross domain
        - Cross standard

2018-11-06 About

PROCEEDINGS of the BlackboxNLP Workshop

Tags:

BlackboxNLP (2018 workshop)

2018-11-06 About

Analyzing and interpreting neural networks for NLP (Workshop's Home page)

Tags:

BlackboxNLP (2018 workshop)

2018-11-06 About

Writing code for Natural language processing Research

Tags:

2018-11-05 About

Transfer learning with language models

Tags:

2018-11-05 About

Multi-Task Identification of Entities, Relations, and Coreference for Scientific Knowledge Graph Construction

Tags:

2018-11-04 About

Conference Schedule - EMNLP 2018

Tags:

EMNLP 2018

2018-11-04 About

Self-Governing Neural Networks for On-Device Short Text Classification - Sujith Ravi | Zornitsa Kozareva (2018)

Tags:

2018-11-02 About

EMNLP (2018) - ACL Anthology - Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing

Tags:

EMNLP 2018

2018-11-02 About

Deep Latent-Variable Models for Natural Language - Tutorial - harvardnlp

Tags:

2018-11-01 About

Facebook Research at EMNLP – Facebook Research

Tags:

2018-11-01 About

Trying to Understand Recurrent Neural Networks for Language Processing (tweets)

Tags:

2018-11-01 About

Writing Code for NLP Research, AllenNLP's tutorial at #emnlp2018

Tags:

2018-10-31 About

Tutorials - EMNLP 2018

Tags:

2018-10-31 About

[1809.00782] Open Domain Question Answering Using Early Fusion of Knowledge Bases and Text

Tags:

2018-09-06 About

[1601.03764] Linear Algebraic Structure of Word Senses, with Applications to Polysemy

Tags:

> Here it is shown that multiple word senses reside
in linear superposition within the word
embedding and simple sparse coding can recover
vectors that approximately capture the
senses

> Each extracted word sense is accompanied by one of about  2000 “discourse atoms” that gives a succinct description of which other words co-occur with that word sense.

> The success of the approach is mathematically explained using a variant of
the random walk on discourses model

("random walk": a generative model for language). Under the assumptions of this model,  there
exists a linear relationship between the vector of a
word w and the vectors of the words in its contexts (It is not the average of the words in w's context, but in a given corpus the matrix of the linear relationship does not depend on w. It can be estimated, and so we can compute the embedding of a word from the contexts it belongs to)

[Related blog post](/doc/?uri=https%3A%2F%2Fwww.offconvex.org%2F2016%2F07%2F10%2Fembeddingspolysemy%2F)

2018-08-28 About

2018 Conference on Empirical Methods in Natural Language Processing - EMNLP 2018

Tags: