Title-Guided Encoding for Keyphrase Generation

• Earlier extraction methods cannot generate absent keyphrases, while generative baselines treat title and body equally and underuse the title’s summary information.Source: Introduction · p. 1
• Keyphrases provide detail on topics also reflected in the title, which the paper describes as a high-level document summary.Source: Introduction · p. 1
• About 33% of absent and around 60% of present keyphrases are TitleRelated, although titles make up only 3%-6% of source text.The paper defines a TitleRelated keyphrase as containing at least one common non-stop-word with the title.Source: Introduction · p. 2
• TG-Net treats the title as a query-like input and gathers relevant title information for each context word before decoding keyphrases.Separate bidirectional GRUs encode the context and title; matching gathers title information, and a further bidirectional GRU merges it into title-guided context representations.Source: Introduction · p. 2
• On five benchmarks, TG-Net significantly outperforms relevant baselines for present and absent keyphrase prediction, especially at very low or high title-length ratios.On the largest dataset, gains reach 9.4% for present-keyphrase F1-measure at 10 and 19.1% for absent-keyphrase Recall at 50 versus the best baseline.Source: Introduction · p. 2

Automatic Keyphrase Extraction

• Extraction approaches identify and rank candidate phrases or label sequences, but cannot generate absent keyphrases.Source: Automatic Keyphrase Extraction · p. 2
• Li et al. (2010) propagates title-phrase importance, whereas Liu et al. (2011) uses the title as target output and does not handle context semantics.TG-Net instead uses contextual title information as an extra query-like input to guide context encoding.Source: Automatic Keyphrase Extraction · p. 2

Automatic Keyphrase Generation

• CopyRNN uses an encoder-decoder with attention and copying; CopyCNN replaces recurrent encoding with convolution and is faster, outperforming CopyRNN on five benchmarks.Source: Automatic Keyphrase Generation · pp. 2–3
• CopyRNN and CopyCNN treat title and body equally; TG-Net uses the title as a query-like input to guide source-context encoding.The title-guided encoding idea is also applicable to CNN-based models, though this paper focuses on an RNN-based model.Source: Automatic Keyphrase Generation · p. 3

• The paper’s notation uses bold lowercase for vectors, bold uppercase for matrices, and calligraphic letters for sets.Θ denotes parameters, and W denotes a parameter matrix.Source: Problem Definition · p. 3
• Keyphrase generation maps a context—the concatenation of title and main body—to a set of keyphrases.The context and each keyphrase are word sequences; the set contains M keyphrases for that context.Source: Problem Definition · p. 3
• Training examples pair each context and target keyphrase with the title as an extra query-like input.The paper splits context-title-keyphrase triplets rather than context-keyphrase pairs.Source: Problem Definition · p. 3

Title-Guided Encoder Module

• TG-Net separately encodes title and context with bidirectional GRUs, matches title words to themselves and body words to the title, then merges title information into context representations.The resulting title-guided context representations form the memory bank for decoding.Source: Sequence Encoding Layer · p. 3, Matching Layer · p. 4, Merging Layer · p. 4

Decoder Module

• An attention-based GRU decoder attends over guided context representations and combines vocabulary generation with copying from source text, producing probabilities over vocabulary plus source words.The copy mechanism provides a soft switch between vocabulary generation and copying.Source: Decoder Module · p. 4

Training

• Training minimizes the negative log likelihood of each target keyphrase word given the preceding words, context, title, and model parameters.The loss is summed across the target keyphrase sequence.Source: Training · p. 4

Experiment Settings

• TG-Net outperforms state-of-the-art models on all five benchmarks for present and absent prediction, with stronger relative gains at very low and very high title-length ratios.Experiments also compare title-length ratios and include ablation and case studies.Source: Experiment Settings · p. 4

Training Dataset

• Generative models train on KP20k's 527,830 training articles, with 20,000 each for validation and testing, then are tested across the listed datasets.KP20k contains 567,830 scientific articles in total.Source: Training Dataset · p. 5

Testing Datasets

• Testing covers Inspec, Krapivin, NUS, and SemEval-2010 alongside KP20k; Table 2 reports dataset statistics and defines training, FFCV, and testing splits.The supervised extractive baseline uses the designated training data or five-fold cross validation, while testing data is used for all models.Source: Testing Datasets · p. 5, Table 2 · p. 5

Implementation Details

• The model uses 100-dimensional embeddings, 256 hidden units, shared embeddings, and Adam with dropout 0.1; reported results average three random seeds.Contexts concatenate titles and abstracts; preprocessing lowercases and tokenizes text, replaces digits, and uses the 50,000 most frequent words.Source: Implementation Details · p. 5

Baseline Models and Evaluation Metric

• Present-keyphrase baselines include TF-IDF, TextRank, Maui, CopyRNN, and CopyCNN; absent-keyphrase baselines are CopyRNN and CopyCNN, evaluated with recall and F1.Keyphrase matches use Porter stemming.Source: Baseline Models and Evaluation Metric · p. 5

Present Keyphrase Predicting

• TG-Net achieves the best present-keyphrase performance on every dataset, including 9.4% higher F1@10 than CopyCNN on KP20k and about 20.2% over CopyRNN.Table 3 reports present-keyphrase results across all test datasets; its “% gain” is over CopyCNN.Source: Present Keyphrase Predicting · pp. 5–6, Table 3 · p. 6

Absent Keyphrase Predicting

• TG-Net consistently outperforms previous sequence-to-sequence models on absent prediction, exceeding CopyCNN by 19.1% in R@50 on KP20k.The authors interpret this as stronger capture of the context’s underlying semantic meaning.Source: Absent Keyphrase Predicting · p. 6, Table 4 · p. 6

Keyphrase Predicting on Various Title Length Ratios

• TG-Net improves across all five title-ratio groups, with larger gains below 3% and above 12%; direct guidance addresses CopyRNN’s long-distance weakness for short titles.Both models perform worse above 12% than at 9%-12%, which the authors attribute to short abstracts that lack enough context for generating all keyphrases.Source: Keyphrase Predicting on Various Title… · pp. 6–7, Figure 3 · p. 6

Ablation Study

• Removing either title-guided context encoding or the copy mechanism lowers present and absent prediction performance on Krapivin.The authors identify both components as effective, with title-guided encoding critical to both tasks and copying helping locate important context.Source: Ablation Study · p. 7, Table 5 · p. 7

Case Study

• In the example, TG-Net predicts the title phrase “nonlinear uncertainties” and generates absent “time-delay systems,” while CopyRNN misses both.The authors describe TG-Net using title information to locate the related abstract content for the absent phrase.Source: Case Study · p. 7, Figure 4 · p. 7

Conclusion

• TG-Net treats the title as an extra query-like input to guide encoding of the document context.The conclusion frames this as explicitly accounting for the title’s leading role in the document.Source: Conclusion · p. 7
• Across five real-world datasets, TG-Net is effective for present and absent keyphrase generation, especially at very low or very high title length ratios.Source: Conclusion · p. 7
• The authors identify evaluation beyond exact-match recall and F-measure as a future direction.Source: Conclusion · p. 7