GDWE

Source code and datasets for the paper "Graph-based Dynamic Word Embeddings" accepted by IJCAI 2022

Installation

Environment:

gcc 4.4.7 or higher is required.

automake 1.1.16

autoscan # generate `autoscan.log` and `configure.scan`
mv configure.scan configure.ac

change the content of configure.ac as follows:

AC_PREREQ([2.69])
AC_INIT([FULL-PACKAGE-NAME], [VERSION], [BUG-REPORT-ADDRESS])
AC_CONFIG_SRCDIR([config.h.in])
AC_CONFIG_HEADERS([config.h])

# Checks for programs.
AC_PROG_CXX
AC_PROG_AWK
AC_PROG_CC
AC_PROG_CPP
AC_PROG_INSTALL
AC_PROG_LN_S
AC_PROG_MAKE_SET
AC_PROG_RANLIB

# Checks for libraries.

# Checks for header files.
AC_CHECK_HEADERS([stdlib.h sys/time.h])

AC_CHECK_LIB([thread/pthread/pthread_create])

# Checks for typedefs, structures, and compiler characteristics.
AC_CHECK_HEADER_STDBOOL
AC_C_INLINE
AM_INIT_AUTOMAKE
AC_TYPE_SIZE_T
AC_TYPE_UINT32_T
AC_TYPE_UINT64_T

# Checks for library functions.
AC_FUNC_STRTOD
AC_CHECK_FUNCS([floor gettimeofday pow sqrt strtol])

CXXFLAGS="-std=c++0x -O3 -funroll-loops -pthread"

AC_CONFIG_FILES([Makefile
                 src/Makefile])
AC_OUTPUT

aclocal # generate `aclocal.m4`
autoconf # generate `configure`
autoheader # generate `config.h` and `config.h.in`
automake --add-missing # generate `Makefile.in` etc.
bash ./configure
make

Basic Usage

Data Preparation

corpus/ provides dataset NYT for experiments in the GDWE paper.

Training

The following command trains the GDWE model on the NYT corpus:

% bash scripts/train_NYT_graph.sh

For more model settings, you can use -h to show the arguments

% ./src/yskip -h

Cross-time Alignment Evaluation

The following command uses pre-trained word embeddings to proceed space alignment evaluation test:

% cd eval
% python eval_align_online2_test.py --embdir [emb_path] --result [result_file_name] 

Example:

% cd eval
% python eval_align_online2_test.py --embdir ../model/e0.2_c5_N4_i1_b1000_E0.2_J0.05_A1.25/ --result gdwe-nyt-test

Text Stream Classification

The following command uses pre-trained word embeddings to proceed text stream classification:

% bash scripts/run_classification.sh