mars-wei/table2vec-lideng

Table2Vec: Neural Word and Entity Embeddings for Table Population and Retrieval

Table2Vec: Neural Word and Entity Embeddings for Table Population and Retrieval.

This repository contains resources developed within the master thesis:

Table2Vec: Neural Word and Entity Embeddings for Table Population and Retrieval.

The programming language is Python. Table embeddings are built using TensorFlow.

Table2Vec

There are hundreds of millions of tables in web pages. These tables are much richer sources of structured knowledge than free-format text.

Table2Vec is a novel approach that employs neural language modeling to embed different table elements into semantic vector spaces, which can benefit table-related retrieval tasks.

Dataset

• The table corpus Wikipedia Tables, which consists of 1.6M high-quality relational tables in total. The statistics is as follows:

Core column	Tables in total	Tables* in total
existing entities	726,913	212,923
60% entities	556,644	139,572
80% entities	483,665	119,166
100% entities	425,236	78,611
100% unique entities	376,213	53,354

Table* represents the tables that have more than 5 rows and 3 columns. Core column refers to the left most column.

• The data/queries.txt file contains the search queries.

Table embeddings

Different table embeddings and their training parameters

Embedding	Total terms	Unique terms	Negative samples	Window size
Table2VecW	200,157,990	1,829,874	25	5
Table2VecH	7,962,443	339,433	25	20
Table2VecE	24,863,683	2,159,467	25	50
Table2VecE*	5,367,837	1,285,708	25	50

Functionality

Table2Vec currently supports three table-related tasks:

• Table retrieval
• Row population
• Column population

Methods and results:

The evaluation is undertaken by trec_eval. For row population, the run files are too big to added on Github.

1. Row population :

Methods	1	2	3	4	5
BL1	0.4360	0.4706	0.4788	0.4786	0.4711
BL2	0.2612	0.2778	0.2845	0.2846	0.2817
BL3	0.2912	0.3024	0.3028	0.2987	0.2910
Table2VecE*	0.4982	0.5522	0.5598	0.5543	0.5476
BL1 + Table2VecE*	0.5581	0.6147	0.6400	0.6524	0.6533
BL2 + Table2VecE*	0.5461	0.6027	0.6187	0.6217	0.6223
BL3 + Table2VecE*	0.5487	0.6049	0.6218	0.6249	0.6251

2. Column population ( the runfile/CP/ folder contains the run files by various methods ):

Methods	Runfile	1	2	3
Baseline	baseline_R1(2,3).txt	0.2507	0.2845	0.2852
Baseline + Table2VecH	combined_R1(2,3).txt	0.2551	0.3322	0.4000

3. Table retrieval ( the runfile/TR/ folder contains the run files by various methods):

Method	Runfile	NDCG@5	NDCG@10	NDCG@15	NDCG@20
Baseline	gt.txt	0.5527	0.5456	0.5738	0.6031
Baseline + Word2Vec	w2v.txt	0.5954	0.6006	0.6315	0.6588
Baseline + Graph2Vec	g2v.txt	0.5844	0.5764	0.6128	0.6340
Baseline + Table2VecW	t2vW.txt	0.5974	0.6096	0.6312	0.6505
Baseline + Table2VecE	t2vE.txt	0.5602	0.5569	0.5760	0.6161