TL;DR
This paper introduces a novel methodology to refine Universal Dependencies (UD) annotations for better analysis of word order in noun phrases. By combining UD relation, part-of-speech tags, and language-specific lemma lists, they achieve a more nuanced understanding of how determiners, quantifiers, and numerals are positioned across ten Indo-European languages, revealing previously hidden variability.
What Problem Does This Paper Address?
Analyzing word order variation, especially within noun phrases, is crucial for understanding language typology and evolution. While Universal Dependencies (UD) provides a standardized framework for linguistic annotation, automatically parsed UD data often contains noise (erroneous annotations). Furthermore, UD's categorization of certain word classes, like determiners, quantifiers, and numerals, can be overly broad, lumping together distinct grammatical functions. This heterogeneity makes it challenging to accurately quantify word order entropy (a measure of variability) for these categories, especially when comparing across languages. Existing studies often rely on curated treebanks or don't fully leverage the richness of UD annotations, leading to potential biases and a lack of fine-grained insights.
Methods
The authors propose a three-stage methodology to "tweak" UD annotations for improved word order analysis:
rel(Relation-based): This baseline approach uses only the UD relation (e.g.,det,nummod) to identify the head-dependent relationship. This is similar to many previous studies.rel+upos(Relation + Universal Part-of-Speech): This stage enhances the analysis by incorporating Universal Parts of Speech (UPOS) tags for both the head and dependent. This helps to disambiguate categories that might be conflated at the relation level.rel+pos+lemma(Relation + POS + Lemma): The most sophisticated stage, this adds language-specific lists of lemmata. These lists are manually compiled using descriptive grammars and native speaker input to precisely identify articles, demonstratives, and quantifiers. This allows for fine-grained analysis by distinguishing between these sub-categories.
The core of their method involves extracting word order patterns from automatically parsed texts and quantifying variability using Shannon's entropy.
Datasets
The study utilizes the Parallel Corpus of Indo-European Prose and more (CIEP+). This corpus contains prosaic texts and their translations in 30 languages. For this research, they selected a sample of 10 Indo-European languages: Lithuanian, Polish (Balto-Slavic); Irish (Celtic); Danish, Dutch, German (Germanic); Greek, French, Portuguese, Spanish (Romance). Each language features 18 books (approx. 120K sentences), except for Irish (5 books, approx. 13K sentences). The corpus was parsed using Stanford Stanza with pre-trained UD Models (version 2.8).
Key Insights and Ideas
- UD's broad categories mask variability: The UD framework's tendency to group determiners, demonstratives, and quantifiers under a single
DETUPOS tag anddetrelation obscures significant cross-linguistic differences in word order for these sub-categories. - Lemmas are key for fine-grained analysis: Manually curated, language-specific lemma lists are crucial for disentangling the "Determiners & Quantifiers" macro-category and revealing nuanced word order patterns, especially for closed-class words like articles and demonstratives.
- Numerals benefit most from UPOS: The
nummodrelation is often used for quantifiers as well, leading to a less direct mapping with theNUMUPOS tag. Incorporating UPOS significantly reduces the entropy for numerals, indicating a more consistent placement. - Genre matters: Using a parallel corpus of fiction texts presented a challenge, as parsers are often trained on non-fiction data (like news and Wikipedia). This highlights the need for robust parsing across diverse genres.
Key Results
- The
rel+uposcombination significantly reduced word order entropy compared to therelbaseline, particularly for numerals. In several languages, numeral entropy dropped to near-zero values (e.g., German: 0.02, French: 0.04). - For determiners & quantifiers, the
rel+uposcombination showed less dramatic entropy reduction, partly because therelbaseline already captured low entropy. - The
rel+pos+lemmacombination, by splitting "determiners & quantifiers" into articles, demonstratives, and quantifiers, revealed moderate-to-high entropy for demonstratives (Greek: 0.59) and quantifiers (Irish: 0.71) in specific languages, which was previously hidden. - Articles, being very frequent, showed near-zero entropy across languages, reinforcing the benefit of breaking down the broader category.
Limitations and Future Work
- Manual lemma compilation: The creation of language-specific lemma lists is labor-intensive and relies on descriptive grammars and native speaker input, which can be a bottleneck.
- Limited scope of lemma lists: The lemma lists were primarily used for positive matches for articles/demonstratives/quantifiers and negative matches for numerals.
- Future work: The authors suggest that this methodology can be extended to other categories that are difficult to analyze automatically. Further research could explore automating the lemma list creation or applying this approach to a wider range of languages and text genres.
Summary updated