TL;DR
This paper introduces a novel deep learning approach to automatically extract spatio-temporal information (location, time, cultural type) from unstructured Chinese archaeological texts. By developing a dedicated corpus and employing BiLSTM-CRF for entity recognition and BiGRU-Dual Attention for relationship extraction, the authors demonstrate a significant improvement in information extraction accuracy, paving the way for more efficient knowledge discovery in archaeology.
What Problem Does This Paper Address?
Archaeological site texts are a treasure trove of information about ancient human activities, but this data is largely unstructured and scattered across various reports, dictionaries, and online entries. Extracting key spatio-temporal details like site names, locations, dynasties, and cultural types from these texts manually is a time-consuming and inefficient process. Existing automated methods often struggle with the nuances of Chinese language and the specific domain of archaeology, leading to poor portability and high implementation costs. This paper aims to bridge this gap by developing an automated system capable of accurately extracting this crucial information, thereby facilitating the creation of archaeological knowledge graphs and enabling better reuse of archaeological data.
Methods
The core of this research lies in a two-stage deep learning pipeline:
- Named Entity Recognition (NER): The authors employ a Bidirectional Long Short-Term Memory with Conditional Random Fields (BiLSTM-CRF) model. This architecture is well-suited for sequence labeling tasks like NER. BiLSTM captures contextual information from both forward and backward directions in the text, while CRF helps model the dependencies between adjacent tags, leading to more accurate entity identification.
- Relationship Extraction (RE): For identifying relationships between the recognized entities, a Bidirectional Gated Recurrent Units with Dual Attention (BiGRU-Dual Attention) model is utilized. This model builds upon BiGRU to process sequential data and incorporates a dual attention mechanism. The word-level attention helps focus on the most informative words for relation classification, while the sentence-level attention refines the representation by down-weighting less relevant parts of the sentence. This dual attention approach aims to improve accuracy by better capturing the semantic relationships.
Datasets
The study introduces a novel corpus for Chinese archaeological site texts. This corpus was constructed by gathering and organizing data from:
- 625 Baidu Baike entries of sites.
- 300 archaeological excavation reports from CNKI.
- 2325 entries from the Dictionary of Chinese Archaeology.
This multi-source dataset was manually annotated using the BIO strategy for entity recognition. For relationship extraction, 8120 corpora were selected from the results of entity recognition.
Key Insights and Ideas
- Domain-Specific Corpus is Crucial: The lack of a publicly available, annotated corpus for Chinese archaeological texts was a significant hurdle. The creation of this new dataset is a foundational contribution.
- Deep Learning Outperforms Traditional Methods: The experimental results clearly show that deep learning models (BiLSTM-CRF and BiGRU-Dual Attention) significantly outperform traditional statistical methods like Hidden Markov Models (HMM) in both entity recognition and relationship extraction for this domain.
- Dual Attention Enhances Relationship Extraction: The BiGRU-Dual Attention model's effectiveness highlights the power of attention mechanisms in capturing complex semantic relationships within text, especially in a domain with potentially ambiguous phrasing.
- Knowledge Graphs as a Natural Output: The paper demonstrates how the extracted spatio-temporal information, represented as structured triples, can be directly used to build knowledge graphs, offering a powerful tool for visualization, querying, and deeper analysis of archaeological data.
Key Results
- Entity Recognition: The BiLSTM-CRF model achieved an F1 score of 87.87% on the test set for named entity recognition. It outperformed both HMM (81.16% F1) and a standalone BiLSTM (86.47% F1).
- Relationship Extraction: The BiGRU-Dual Attention model achieved an F1 score of 88.05% for relationship extraction, outperforming a BiLSTM-Attention baseline (85.76% F1).
- Entity-Specific Performance: For entity recognition, "cultural type" and "geographical location" entities showed the highest F1 scores (94.33% and 90.50% respectively), likely due to clearer linguistic markers. "Historical dynasty" had a lower F1 (82.28%), suggesting more complexity in its textual representation.
- Relationship-Specific Performance: The "Location of site" relationship achieved the highest F1 score (90.01%) in relationship extraction, correlating with its higher prevalence in the dataset.
Limitations and Future Work
- Low Recall for Certain Entities: The authors acknowledge that the BiLSTM-CRF model exhibits low recall for some entities, particularly "historical dynasty," due to the variability of sentence patterns and the potential for new entities to emerge.
- Annotation Corpus Limitations: The performance of relationship extraction is noted to be influenced by the lack of a comprehensive annotated corpus for relationships.
- Future Work:
- Expand the annotation corpus to include more entity types (e.g., excavated artifacts, site areas) and relational annotations.
- Improve data augmentation techniques to balance entity relation distributions.
- Develop more advanced semantic search, intelligent Q&A, and other upper-level applications based on the constructed archaeological knowledge graph.
Summary updated