TL;DR
This paper introduces ParaBLoCC, the first multilingual corpus of parallel Basic Locative Constructions (BLCs). This resource is crucial for advancing research into how humans express and understand spatial relationships across languages, and for improving computational models of spatial language understanding.
What Problem Does This Paper Address?
Understanding how humans express spatial relationships is a fundamental challenge in linguistics and cognitive science. Basic Locative Constructions (BLCs) – sentences that describe the location of one object (figure) relative to another (ground) – are a ubiquitous and essential part of language. While BLCs are found in virtually all languages, their syntactic expression varies significantly. Existing research has largely relied on controlled elicitation or focused on a limited number of languages. This leaves a gap in our understanding of BLCs in natural, uncontrolled language use across a wide range of typologically diverse languages. Furthermore, computational models for extracting spatial information from text often struggle with the nuances of these constructions, especially in a multilingual context.
Methods
The authors introduce the ParaBLoCC corpus, which was constructed through a multi-stage filtration process applied to existing bitext corpora. The process involved:
- Lexical Filtering: Identifying sentence pairs where the English sentence contained a locative spatial expression from a predefined list of fifty common English locative prepositions and twenty-one spatial nominals.
- Syntactic Filtering: Using dependency parsing and subgraph matching to ensure the English sentence's syntactic structure matched one of eight identified BLC syntactic templates.
- Spatial Sense Filtering: Employing a trained neural word sense disambiguation (WSD) model to distinguish between spatial and non-spatial uses of polysemous words, ensuring the identified expressions truly conveyed spatial relationships.
The WSD model, termed "glossbert," was trained on a dataset of sentences exemplifying spatial senses from the Preposition Project.
Datasets
The ParaBLoCC corpus is built upon existing bitext corpora harvested from sources like the Opus Machine Translation Portal, TEDTalks, and others. The corpus contains parallel English sentences paired with sentences from twenty-six typologically diverse languages. The authors also mention using data from "The Preposition Project" and "200 hand-labeled in-domain sentence pairs" for evaluating their spatial sense filter.
Key Insights and Ideas
- Ubiquity of BLCs: BLCs are confirmed to be a cross-linguistically ubiquitous construction, present in all languages studied.
- Diversity in Expression: While BLCs are universal, their syntactic realization varies greatly across languages, from English's "NP Copula PP" to locative suffixes in languages like KwaKwala.
- Importance of Naturalistic Data: Most prior studies on BLCs used controlled elicitation. ParaBLoCC provides a large-scale corpus of BLCs from natural, uncontrolled text, enabling research into real-world usage patterns.
- Challenges in Spatial Sense Disambiguation: Accurately identifying the spatial meaning of polysemous words in context is a significant challenge, even for specialized models.
Key Results
- The ParaBLoCC corpus contains parallel BLCs in English and twenty-six other languages, sourced from various domains.
- The observed BLC occurrence rate in the corpus is estimated at 7.2 per thousand original sentences. However, the authors believe this is an underestimate, with the true rate potentially being around twenty per thousand.
- The spatial sense classifier achieved a recall of .70 and a precision of .69 on a validation set, and .80 recall and .56 precision on an in-domain set, with macro-averaged F1 scores of .66 for both. While not state-of-the-art, this performance was deemed adequate for the large-scale filtration task.
- Error analysis revealed that the spatial sense filter struggled more with longer sentences, especially those with multiple clauses.
Limitations and Future Work
- Language Selection: The corpus is limited to languages with sufficient bitext availability, potentially excluding languages with highly distinct spatial expression systems (e.g., those relying solely on absolute reference frames).
- Annotation Granularity: BLCs are not explicitly annotated with spans in the corpus. Future work could involve adding explicit span annotations and semantic role annotations (figure, ground, spatial relation) for more in-depth analysis.
- Spatial Sense Filter Improvement: The authors acknowledge that the spatial sense filter's performance can be improved, particularly for complex sentence structures. They suggest this as a direction for future research.
- Downstream Applications: The corpus is expected to be useful for multilingual spatial relation extraction, improving geospatial expression resolution models, and enhancing multilingual image captioning models.
Summary updated