TL;DR
This paper introduces two novel German corpora, "Tschick" and "Auferstehung der Toten" (AdT), annotated with fine-grained linguistic features for conversational narratives. These corpora address the lack of diverse, naturalistic language data in NLP by focusing on informal registers and providing detailed annotations of referring expressions (REs), including their grammatical and thematic roles, and referential chains.
What Problem Does This Paper Address?
The field of Natural Language Processing (NLP) heavily relies on large, high-quality datasets. However, most available corpora are in English and focus on formal written genres like Wikipedia articles or news. This creates a significant gap when it comes to understanding and generating more naturalistic, colloquial, and spoken language. Spoken language data is particularly challenging to handle, so researchers often turn to written texts that mimic conversational style. This paper tackles this challenge by creating and annotating two German corpora derived from novels with a conversation-like narrative style. This allows for the study of how informal language influences the choice of referring expressions, which are crucial for coreference resolution and natural language generation.
Methods
The authors developed a multi-layered annotation scheme and applied it to excerpts from two German novels. The annotation process was carried out using the web-based software WebAnno. Key methodological aspects include:
- Multi-layer Annotation: Annotations were performed at multiple levels, including:
- Sentence Segments: Sentences were broken down into smaller, comparable units to account for the variable length and fragmented nature of colloquial speech.
- Animate Referring Expressions (REs): All REs referring to animate entities were annotated.
- RE Features: For each RE, detailed features were specified:
- RE Type: Categorization based on a hierarchy (e.g., personal pronoun, proper name, definite DP, indefinite DP).
- Grammatical Role: Subject, direct object, indirect object, or oblique.
- Thematic Role: Proto-Agent, Proto-Patient, or Proto-Recipient, based on the proto-role approach.
- Referential Chains: Annotation of coreferential relationships between REs.
- Character Names: Mapping of referential chains to character names for better analysis of recurring entities.
- Perspective Information: Ongoing annotation of the character whose perspective is represented by the RE.
- Feature Importance Analysis: The authors employed XGBoost models and SHAP analysis to determine which annotated features (e.g., segment distance, grammatical role, thematic role, antecedent RE type) are most influential in predicting the form of a referring expression.
Datasets
The paper introduces two new corpora:
- Tschick Corpus: Derived from 9 chapters of the German novel "Tschick" by Wolfgang Herrndorf (2010). This novel is described as a road novel or coming-of-age novel with a naturalistic, conversation-like narration style and authentic youth language.
- Auferstehung der Toten (AdT) Corpus: Derived from the first four chapters of the Austrian crime novel "Auferstehung der Toten" by Wolf Haas (1996). This novel features an omnipresent narrator and a style reminiscent of oral language.
Both corpora are publicly available on the Open Science Framework.
Key Insights and Ideas
- Informal Language Matters: The study strongly suggests that the register of a text significantly influences the choice of referring expressions. Informal language, common in the chosen novels, leads to different RE patterns compared to formal texts.
- Pronouns Dominate Informal Narratives: Personal pronouns constitute a remarkably high proportion of REs in these conversational narratives (nearly 70% of all REs), a stark contrast to formal texts where their proportion is much lower.
- Null REs are Significant: The corpora include "zero" REs (null anaphora), which are typically absent in formal German texts but are important in colloquial speech.
- Segment Distance is Key: Feature importance analysis revealed that the distance between an RE and its antecedent (measured in segments) is a crucial factor in determining the RE's form, particularly favoring non-pronominal forms (names and DPs) for longer distances.
- Thematic Roles Play a Role: Contrary to initial expectations from variable importance graphs, SHAP analysis showed that thematic roles (like Proto-Patient) significantly contribute to the choice of RE classes.
- Rich Annotation is Necessary: The multi-layered annotation scheme provides a comprehensive characterization of RE usage, going beyond simple coreference to include prominence-lending features and referential chains, which are essential for deeper linguistic analysis and NLP applications.
Key Results
- The combined corpora contain 3264 annotated REs referring to animate entities.
- Personal pronouns are the most frequent RE type, accounting for 42.59% of all annotated REs.
- The AdT corpus shows a higher proportion of third-person singular pronouns (53.64%) compared to the Tschick corpus, which has a higher proportion of first-person singular pronouns (44.86%).
- Feature importance analysis identified segment distance and the antecedent's RE type as the most significant predictors of an RE's form.
- SHAP analysis indicated that longer segment distances promote non-pronominal REs, while thematic roles also contribute significantly to RE class prediction.
Limitations and Future Work
- Corpus Size: The current corpora are relatively small, being based on excerpts from novels. The authors plan to expand them by including more chapters.
- Animate Referents Only: The current annotations are limited to animate discourse referents. Future work aims to annotate inanimate entities with the same features.
- Partial Perspective Annotation: Perspective information is currently only annotated for a subset of REs. This will be expanded to other referential forms.
- Unbalanced RE Distribution: The corpora have an unbalanced distribution of RE forms, with a strong tendency towards pronouns. This could bias modeling attempts. Future work will involve annotating more REs and leveraging state-of-the-art models like German BERT (GBERT) to improve prediction reliability.
- WebNLG-like Corpora: The authors envision using these annotations to build WebNLG-like corpora for end-to-end neural modeling of RE generation.
Summary updated