A Mixed Hierarchical Attention based Encoder-Decoder Approach for Standard Table Summarization

• Standard table summarization generates abstractive summaries from tables with a predefined fixed schema, rather than tables with unknown schemas.Source: §1 · p. 1
• Rule- and template-based approaches are not scalable because they require hand-crafted selection and generation rules across many possible choices.Source: §1 · p. 1
• The model computes row representations using static attention over fixed attributes, then applies dynamic attention over rows before decoding.The approach leverages tables' hierarchy of fixed attributes within rows and rows within a table.Source: §1 · p. 1
• Around 18 BLEU (around 30%) improvement over Mei et al. (2016) is reported for the proposed model.Source: §1 · p. 2

2 Tabular Data Summarization

• The task represents each table as records with fixed attributes and record types, paired with a natural-language summary for end-to-end learning.The model takes the table as input and produces the corresponding summary.Source: §2 · p. 2

2.1 Mixed Hierarchical Attention Model (MHAM)

• MHAM uses static attention over each record’s attributes and dynamic attention over records that can vary with each decoder step.Attribute attention reflects the fixed schema; record attention can prioritize different record types while generating a word.Source: §2.1 · pp. 2–3
• Record-type embeddings help calculate attribute-importance scores, which produce static weights for combining each record’s attribute values.The resulting attribute-attended record representations are passed to a GRU record encoder.Source: §2.1 · p. 3
• The GRU record encoder concatenates each hidden state with its embedded record representation; learned static record weights regulate decoder-time attention.At each step, the decoder uses soft attention over encoded records, then maps its GRU state through a softmax to produce an output word.Source: §2.1 · p. 3
• A single pass costs O(TM + TT'), compared with O(TMT') for fully dynamic attention at both levels.The authors also propose a non-hierarchical model with attention only at record level, concatenating each record’s attributes and type.Source: §2.1 · p. 3

3 Experiments

• WEATHERGOV contains 29,528 table-summary pairs from 3,753 U.S. cities, with 12 record types and 36 weather records per table.The train:validation:test split is 25,000:1,000:3,528.Source: §3 · p. 4
• Numbers use binary encodings, while record types use one-hot vectors and categorical attributes use ordinal or one-hot encodings.Mode, direction, and time-interval attributes receive specific ordinal encodings.Source: §3 · p. 4

3.1 Training and hyperparameter tuning

• Models were tuned for validation sBLEU over 500 epochs without regularization, with hyperparameter tuning performed separately for each model.Training used Adam at 0.0001; the reported MHAM configuration used embedding size 100, GRU size 400, and static-record-attention size P of 150.Source: §3.1 · pp. 4–5
• Evaluation reports BLEU and Rouge-L, including customized BLEU, which treats numbers differing by at most five as equivalent.The customized metric considers 20 and 18 the same.Source: §3.1 · p. 5

4 Results and Analyses

• 16.6 cBleu (24%) and 18.3 sBleu (30%) improvements over MBW are reported for MHAM; Table 2 presents the overall model results.MHAM also improves over NHM on all reported metrics.Source: Table 2 · p. 5, §4 · p. 5
• Attention heatmaps show decoder-step record focus and attribute weighting, including temperature focus for “52” and precipitation-potential focus for “mostly cloudy.”Temperature, wind-speed, and gust records receive high weights on their min/max/mean attributes.Source: §4 · p. 5, Figure 3 · p. 5
• Example summaries indicate that hierarchical attention supports more accurate number generation and captures weak signals such as snow accumulation.MHAM also avoids repetition compared with NHM.Source: §4 · p. 5

5 Conclusion and Future Work

• The authors propose variable-schema tabular summarization as future work.The paper’s formulation concerns tables drawn from a predefined schema.Source: §5 · p. 5