Abstract
We study automatic title generation for a given block of text and present a method called DTATG to generate titles. DTATG first extracts a small number of central sentences that convey the main meanings of the text and are in a suitable structure for conversion into a title. DTATG then constructs a dependency tree for each of these sentences and removes certain branches using a Dependency Tree Compression Model we devise. We also devise a title test to determine if a sentence can be used as a title. If a trimmed sentence passes the title test, then it becomes a title candidate. DTATG selects the title candidate with the highest ranking score as the final title. Our experiments showed that DTATG can generate adequate titles. We also showed that DTATG-generated titles have higher F1 scores than those generated by the previous methods.
Introduction
Automatic Title Generation (ATG) is a specialized form of text summarization aimed at producing succinct, informative labels for documents. Traditional methods, such as TF-IDF or Naive Bayesian models, often result in a "bag of keywords" that lacks grammatical structure. More advanced approaches using Context-Free Grammar (CFG) trees are computationally expensive and often struggle with syntactic ambiguity.
DTATG (Dependency-Tree Automatic Title Generator) addresses these issues by utilizing dependency trees to prune "central sentences" into grammatically sound titles. By focusing on the syntactic skeleton of a sentence rather than just keyword frequency, DTATG produces titles that are both readable and highly relevant to the document's core message.
Methodology
The DTATG framework follows a multi-stage pipeline to transform raw document content into a refined title.
Phase 1: Central Sentence Extraction
The process begins with document segmentation and keyword extraction using the RAKE (Rapid Automatic Keyword Extraction) algorithm. To identify the most representative sentences, the authors employ a novel scoring metric: $$Rank_2 = \sum_{i=1}^{n} 2^{s_i}$$ where $s_i$ represents the score of keyword $i$. This exponential weighting prioritizes sentences containing high-impact keywords, ensuring the "central sentence" captures the document's primary focus.
Phase 2: Dependency Tree Compression
Once a central sentence is selected, a dependency tree is constructed using the Stanford Dependency Parser. Unlike CFG, dependency trees map the functional relationships between words (e.g., subject-verb-object). The system applies a Dependency Tree Compression Model (DTCM) to prune non-essential branches. For instance, the root (main verb), nsubj (nominal subject), and dobj (direct object) are typically preserved, while determiners (det) and auxiliary verbs (aux) are removed unless they are part of a identified keyword.
Phase 3: The Title Test
Candidate titles must pass a final "Title Test" to ensure quality. The criteria include:
- Conciseness: A maximum of 15 words and a "Subject + Verb + Object" structure.
- Fluency: Adherence to grammatical rules.
- Topic-relevance: Inclusion of at least one major document theme.
Evaluation and Results
The system was evaluated using a corpus of 2,225 BBC news articles across five categories: Business, Entertainment, Politics, Sport, and Tech.
Quantitative Performance
DTATG was compared against a standard TF-IDF baseline. As shown in the figure below, DTATG achieved significantly higher F1 scores in four out of the five categories. The performance advantage was particularly pronounced in the "Sport" category, where the structural approach of DTATG far outperformed simple keyword matching.
Human Assessment
Five human evaluators rated the generated titles on a 5-point scale. DTATG consistently scored above 4.0 for both fluency and topic relevance. Interestingly, in categories like "Entertainment" and "Sport," generated titles were sometimes preferred over original human-written headlines for being more descriptive.
Discussion and Future Work
A key insight from this research is the effectiveness of dependency trees as "pruning shears." By maintaining the syntactic skeleton of a sentence, the system preserves readability while reducing length. Furthermore, the $Rank_2$ formula proves that prioritizing high-quality keywords is more effective than simply counting them.
However, the system currently relies on extracting a single "central sentence." This approach can struggle with complex documents that lack a clear thesis statement (referred to as "type-2 documents"). Future work will focus on multi-sentence synthesis and more advanced compression techniques, such as word substitution and reordering, to further enhance title naturalness.
Summary updated