This work addresses disambiguating Web tables - annotating content cells with named entities and table columns with semantic type information. Contrary to state-of-the-art that builds features based on the entire table content, this work uses a method that starts by annotating table columns using automatically selected partial data (i.e., a sample), then using the type information to guide content cell disambiguation. Different sample selection methods are introduced and tested to show that they contribute to higher accuracy in cell disambiguation, comparable accuracy in column type annotation with reduced computation.
Enabling machines to effectively and efficiently access the increasing amount of
tabular data on the Web remains a major challenge to the Semantic Web, as the classic
indexing, search and NLP techniques fail to address the underlying semantics carried
by tabular structures [
Sample driven Table Interpretation opens up several challenges. The first is defining
a sample with respect to each task. The second is determining the optimal size of the
sample with respect to varying sizes of tables. The third is choosing the optimal sample
entries, since a skewed sample may damage accuracy. Our previous work in [
been carried out in semantic Table
Interpretation, such as Venetis et
al. [
TableMiner is previously described in [
Modified TableMiner For the purpose of this study, TableMiner is modified (T Mmod)
to contain only the first phase and part I of the second phase. In other words, we do not
revise the column classification results obtained from sample data. Therefore T Mmod
may only use a fraction of a column’s data to classify the column, which reduces
computation overhead compared to classic ‘exhaustive’ methods.
Sample selection The choice of the sam- Algorithm 1 Sample based classification
ple can affect learning in T Mmod in two
ways. While the size of the sample is 1: Input: Tj; Cj ← ∅
dealt with by the convergence measure 32:: forparlelvcCeljl T←i;jCinj Tj do
described in [
One-sense-per-discourse (ospd) First and foremost, we make the hypothesis of ‘one-sense-per-discourse’ in table context, that if an NE-column is not the subject column of the table (e.g., the first column in Figure 1 is a subject column), then cells containing the same text content are extremely likely to express the same meaning1. Thus to apply ospd we firstly re-arrange cells in a column by putting those containing duplicate text content adjacent to each other. Next, when disambiguating a content cell, the feature representation of the cell concatenates the row context of the cell, and that of any adjacent cells with the same text content (e.g., in Table 1 we assume ‘AetoliaAcarnania’ on the three rows to have the same meaning, and build a single feature representation by concatenating all the three rows). Effectively this creates a richer feature representation for cells whose content re-occur across a table.
Feature size (fs) With fs, we firstly apply ospd, then rank cells in a column by the size of their feature representation as determined by the number of tokens in a bag-ofwords representation. This would allow T Mmod to start with cells that potentially have the largest - hence ‘richest’ - feature representation in Algorithm 1.
Name length (nl) With nl, we count the number of words in the cell text content to be disambiguated and rank cells by this number - name length (e.g., in Table 1 ‘AetoliaAcarnania’ has two words and will be disambiguated before ‘Boeotia’). nl merely relies on the name length of a cell content and does not apply ospd. The idea is that longer names are less likely to be ambiguous. 4
We evaluate the proposed methods of sample selection using two datasets: LimayeAll and Limaye2002. LimayeAll contains over 6000 tables and is used for evaluating content cell disambiguation. Limaye200 contains a subset 200 tables from LimayeAll with
Cell disambiguation (LimayeAll) Column classification (Limaye200) T Mmod T M mosopdd T M mfsod T Mmnlod T Mmod T M mosopdd T M mfsod T Mmnlod 0.809 0.812 0.812 0.813 0.723 0.719 0.721 0.723
Table 1. Cell disambiguation and column classification accuracy in F1.
columns manually annotated with Freebase concepts, and used for evaluating column
classification. As a baseline, T Mmod without any sample selection techniques is used.
It simply chooses cells from a column in their original order in Algorithm 1. This is
compared against T M mosopdd, which applies ospd to non-subject NE-columns, preserves
the original order but disambiguates groups of cells containing the same text content;
T M mfsod that applies ospd to non-subject NE-columns then prioritizes cells that
potentially have richer feature representation; and T Mmnlod that prioritizes cells containing
longer text content. Results on both datasets are shown in Table 1. It suggests that,
compared against T Mmod, the sample selection techniques can enhance the accuracy of cell
disambiguation marginally. In the column classification task however, they do not add
benefits in terms of accuracy. By analyzing the computation overhead in terms of the
automatically determined sample size in each table, it shows that the sample selection
techniques have reducing the amount of data to be processed in column classification.
As an example, T Mmod converges on average after processing 58% of cells in a table
column, i.e., it manages to classify a table column using a sample size of 58% of the
total number of cells in that column. T M mosopdd reduces this to 53%, for T M mfsod 52%
and for T Mmnlod 58% (unchanged). This may contribute to noticeable reduction in CPU
time since the construction of feature space (including querying knowledge bases) for
each data unit consumes over 90% of computation time [
Acknowledgement: Part of this work is carried out in the LODIE project (Linked Open Data Information Extraction), funded by EPSRC (EP/J019488/1).