ACL-16 @ ProbModels

  • Tutorial
  • LSTM has been a standard (and strong) baseline in the NLP toolbox
  • Question generation
    • Generating Factoid Questions With Recurrent Neural Networks
    • Generating Natural Questions About an Image
  • Reading comprehension
    •  matching -> inference
    • A Thorough Examination of the CNN/Daily Mail Reading Comprehension Task
    • WIKIREADING: A Novel Large-scale Language Understanding Task over Wikipedia
  • Involve human during learning
    • On-line Active Reward Learning for Policy Optimisation in Spoken Dialogue Systems
    • Learning Language Games through Interaction
  • Using natural language as an representation layer
    • Using Sentence-Level LSTM Language Models for Script Inference
    • Deep Reinforcement Learning with a Natural Language Action Space
    • Natural Logic

How to Build SyntaxNet

The missing zlib.BUILD file can be found in https://github.com/tensorflow/serving/blob/master/zlib.BUILD.

Follow the steps in https://github.com/bazelbuild/bazel/issues/623:

Summary:

  • download jpeg.BUILD.txt, png.BUILD.txt, and WORKSPACE.txt from @srsaharoy ‘s message #623 (comment) . (the 2nd post with attachments) and place these files in the tensorflow lib without the .txt extension instead of the existing files (in my case ~/git/tensorflow/tensorflow).
  • create folder with external source files: ~/git/tensorflow/fix/files/re2 ~/git/tensorflow/fix/files/jpeg-9a/jpeg-9a ~/git/tensorflow/fix/files/gemmlowp ~/git/tensorflow/fix/files/libpng-1.2.53/libpng-1.2.53 ~/git/tensorflow/fix/files/six-1.10.0 Note the dir-in-dir for jpeg-9a and libpng-1.2.53. This is necessary.
  • change paths in WORKSPACE file to match the location of the aux source files
  • bazel shutdown
  • bazel build -c opt //tensorflow/cc:tutorials_example_trainer

EDIT: the path for ‘six’ in WORKSPACE needs to point to the folder that contains six-1.10.0, e.g.:
path = “/home/…/git/tensorflow/fix/files”
This is because six.BUILD searches the sources in “six-1.10.0/six.py”.

EDIT: (for commit a108da2 16-DEC-2015) download & extract https://bitbucket.org/eigen/eigen/get/3.3-beta1.tar.gz ; change eigen_archive to new_local_repository with path = “/home/…/git/tensorflow/fix/files”

This may be also true for other modules.

https://github.com/bazelbuild/bazel/issues/623

openafs on Ubuntu 16.04

Error message:

aklog: a pioctl failed while obtaining tokens for cell

Solution:

  • rollback your kernel to 4.2
    • check your current kernel version: uname -r
    • remove it: sudo apt-get purge linux-image-x.x.x.x-generic
    • reboot
$ sudo -i
# apt-get remove openafs-client openafs-krb5 openafs-modules-dkms krb5-config krb5-user ntp
# apt-get install krb5-config krb5-user ntp openafs-client openafs-krb5 openafs-modules-dkms
# echo "inf.ed.ac.uk" > /etc/openafs/ThisCell
# /etc/init.d/openafs-client start
# exit
$ kinit [DICE username]
$ aklog
$ /etc/init.d/openafs-client stop
$ /etc/init.d/openafs-client start