PaperlayerView exact revision on arXiv

Source-linked AI summary

A Graph Representation of Semi-structured Data for Web Question Answering

Xingyao Zhang, Linjun Shou, Jian Pei, Ming Gong, Lijie Wen, Daxin Jiang

arXiv:2010.06801v1cs.CLcs.AI

TL;DR

Web tables and lists carry structural relations useful for question answering, but many prior methods flatten them and overlook those relations. GraSSLM models their components and relations as a graph, with pre-training and graph reasoning; it outperforms state-of-the-art baselines by up to 1.77 in F1 score on average.

  • Problem

    Prior Web QA methods often flatten tables and lists into text, leaving their inherent structural relationships unused despite their relevance to answering queries.

    Source: §1 · pp. 1–2

  • Method

    GraSSLM represents table and list components and their relations in a graph, using pre-training objectives and graph reasoning to model textual and structural information.

    Source: §1 · p. 2, §3 · p. 4, §3.4 · p. 6

  • Results

    Up to 1.77 in F1 score on average over the best state-of-the-art baselines, GraSSLM improves performance on real query/table-list pairs.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Experiments on real-world data support the effectiveness of modeling textual and structural information together for Web QA.

    Source: §1 · p. 2, §6 · p. 10

    The method covers relational, entity, and matrix tables and enumerate and group lists without distinguishing among these types.

    Source: §2 · p. 2

Abstract

from arXiv · show

The abundant semi-structured data on the Web, such as HTML-based tables and lists, provide commercial search engines a rich information source for question answering (QA). Different from plain text passages in Web documents, Web tables and lists have inherent structures, which carry semantic correlations among various elements in tables and lists. Many existing studies treat tables and lists as flat documents with pieces of text and do not make good use of semantic information hidden in structures. In this paper, we propose a novel graph representation of Web tables and lists based on a systematic categorization of the components in semi-structured data as well as their relations. We also develop pre-training and reasoning techniques on the graph model for the QA task. Extensive experiments on several real datasets collected from a commercial engine verify the effectiveness of our approach. Our method improves F1 score by 3.90 points over the state-of-the-art baselines.

1 Introduction

Web tables and lists contain relational information useful for QA, but prior approaches often flatten them and miss those relations. The paper introduces GraSSLM, a graph model with pre-training and reasoning, and reports improved F1 on real query-data pairs.

1 Introduction

  • Semi-structured data can answer 30% of user-query intents, and tables and lists provide relational information useful for QA.Examples include questions answered from a Web table or a Web list.Source: §1 · p. 1
  • Prior approaches commonly rearrange table and list text into documents without using their inherent structural relationships.The paper identifies vertical relations among cells in the same column as one example.Source: §1 · pp. 1–2
  • GraSSLM represents semi-structured data components as graph nodes and their relations as edges, jointly modeling text and structure.The model is intended to reveal semantic correlations across components.Source: §1 · p. 2
  • Its node prediction objective predicts masked-node content from unmasked neighbors, learning contextual attention used in graph reasoning.The reasoning stage updates each node representation with aggregated information from neighboring nodes.Source: §1 · p. 2
  • Up to 1.77 in F1 score on average over the best state-of-the-art baselines, GraSSLM improves Web QA performance.The evaluation uses real-world query/table-list pairs labeled for relevance by crowdsourcing annotators.Source: §1 · p. 2

2 Problem Statement

The section defines QA over a query and a Web table or list, then describes the components and semantic relations represented in the data. It treats several table and list formats within the same framework.

2 Problem Statement

  • The framework applies to relational, entity, and matrix tables and enumerate and group lists without distinguishing among these types.Source: §2 · p. 2

2.1 Components of Web Semi-structured Data

  • Tables include captions, headers, cells, subject columns, and attribute columns; a list is treated as a single-column table without a header.Subject columns contain key entities, while attribute columns contain information about them.Source: §2.1 · p. 3, Figure 2 · p. 3

2.2 Relations among Components in Tables and Lists

  • The four relation types are caption-content, header-cell, subject-attribute, and cell-cell relations among components.A caption describes the content body; header-cell links are typically class-instance relations.Source: §2.2 · p. 3
  • Subject-attribute relations connect an entity’s subject-column name to its attribute values, while cell-cell relations link cells in the same row or column.Same-row cells often describe one entity’s attributes, and same-column cells are often instances of the same class.Source: §2.2 · pp. 3–4
  • Lists have only caption-content and cell-cell relations in this framework.Source: §2.2 · p. 4

2 Problem Statement

  • QA over semi-structured data assigns a match score to a query and table or list, predicting the likelihood that the data answers the query.Source: §2.2 · p. 4

3 Method

GraSSLM represents a query and semi-structured Web data as a graph, then uses graph reasoning to predict their match. Its pretraining combines whole-cell masking with neighbor-based token prediction.

3 Method

  • GraSSLM combines contextual language embeddings, graph construction, and graph reasoning with classification to predict query–data matches.Source: §3 · p. 4

3.1 Graph Construction

  • Four structural edge types represent caption-content, header-cell, subject-attribute, and neighboring-cell relations; query edges connect the query to data nodes, while list graphs omit two table-edge types.Source: §3.1 · pp. 4–5

3.2 Graph Initialization

  • The model concatenates the query and semi-structured data text, then uses a pretrained language model and BiLSTM-based span representations to initialize graph nodes.Source: §3.2 · p. 5

3.3 Graph Reasoning and Prediction

  • A two-layer GCN updates node representations by aggregating neighbors; mean pooling and an MLP produce the query–data match score.Source: §3.3 · p. 5

3.4 Pre-training Strategy

  • Whole Cell Masking masks 15% of cells and replaces masked tokens with mask tokens, random tokens, or originals in an 80/10/10 split.Source: §3.4 · p. 6
  • Neighbor Prediction Objective predicts masked-node tokens from neighbor representations, using positional embeddings; it reuses WCM-masked input and is summed with WCM loss.Source: §3.4 · p. 6

4 Experiments

Experiments compare GraSSLM with strong baselines on table and list query-matching datasets, then assess graph components and pre-training objectives. GraSSLM leads across the datasets, while ablations and objective comparisons quantify contributions from explicit graph reasoning and pre-training.

4 Experiments

  • Table-QM has about 34,000 labeled cases, List-QM about 62,000, and DTable-QM 2,457 cases selected for having over 50 cells and 10 columns.DTable-QM is a subset of Table-DM focused on more complicated tables with larger amounts of information.Source: §4 · pp. 6–7
  • Pre-training uses 7 million Web tables and 100 million Web lists extracted from large-scale unlabeled datasets.The table corpus comes from 10 million sampled Web pages; the list corpus is derived similarly.Source: §4 · p. 7
  • Baselines include document-retrieval methods, STR, BERT, and TAPAS, which encodes table rows and columns.The earlier baselines described treat semi-structured examples as documents or do not consider table/list structure.Source: §4 · p. 7
  • The study measures average F1 for binary matching, uses 80/10/10 train-validation-test splits, and reports means and standard deviations over three runs.The best validation-set model is tested on the held-out test set.Source: §4 · pp. 7–8

4.1 Overall Performance

  • GraSSLM consistently leads all baselines, beating BERT by up to 5.44%, STR by up to 7.68%, and TAPAS by up to 3.90% on List-QM.The performance comparison covers Table-QM, List-QM, and DTable-QM.Source: Table 2 · p. 8, §4.1 · p. 8
  • On DTable-QM, GraSSLM scores 71.19% despite severe drops for all baselines; explicit graph modeling guides attention over noisy contexts.The authors connect this to semantic reasoning on complicated tables.Source: §4.1 · p. 8

4.2 Ablation Studies

  • Removing GCN reduces performance by 3.36% on average and 5.12% on lists; removing Cell-Cell relations costs 1.74% on average.Header-Cell and Subject-Attribute removal costs 1.33% and 0.57% on average, respectively, while replacing Bi-LSTM reduces performance by 0.79%.Source: Table 3 · p. 9, §4.2 · p. 9

4.3 Pre-training Strategy Analysis

  • WCM alone improves performance by up to 1.97% and NPO by up to 1.84%; their combination achieves the largest increase, 1.97%.WCM guides token-level embeddings, while NPO propagates semantic correlations among neighboring nodes through graph connections.Source: Table 4 · p. 9, §4.3 · pp. 9–10

5 Related Work

Prior work ranges from document-retrieval methods that emphasize textual content to models that encode some table structure. The paper situates its approach against these methods by addressing relations beyond simple row and column structure.

5 Related Work

6 Conclusion

Semi-structured Web data are a rich source for Web QA, but prior methods do not fully use their structural information. The paper models textual and structural information together and reports experimental evidence of its effectiveness.

6 Conclusion

  • The proposed approach models both textual and structural information in semi-structured Web data for QA.Source: §6 · p. 10
  • Previous methods do not take full advantage of structural information in semi-structured Web data.Source: §6 · p. 10
  • Extensive experimental results verify the effectiveness of the approach.Source: §6 · p. 10

Related papers

Loading 2010.06801v1…