A new semantically annotated corpus with syntactic-semantic and cross-lingual senses

Technical Summary: A New Semantically Annotated Corpus with Syntactic-Semantic and Cross-Lingual Senses

Problem Statement
The paper addresses the challenge of defining an appropriate and effective sense inventory for Word Sense Disambiguation (WSD), particularly within the context of Machine Translation (MT) applications. The authors identify a limitation in existing approaches:

• Translational Inventories: Common in multilingual WSD (e.g., SemEval tasks), these define senses based on actual translations in parallel corpora. While application-oriented and attested, they rely on cross-lingual evidence rather than the distributional (syntactic/semantic) evidence typically used by WSD systems. This mismatch can lead to heterogeneous sense descriptions where contexts with different linguistic features are assigned the same label, increasing task difficulty.
• Lexicon-Style Inventories: Approaches using computational lexicons (e.g., VerbNet, Lexicon-Grammar) define senses based on syntactic and semantic characteristics of the source language. While computationally convenient for feature extraction, these distinctions may be too coarse-grained for multilingual contexts, failing to capture nuances that result in different translations in the target language.

The authors propose that neither approach individually provides an optimal granularity for WSD and suggest combining evidence from both translations and computational lexicon entries to determine polysemous word senses.

Methodology
The authors constructed a new semantically annotated corpus focusing on 20 French polysemous verbs (e.g., arrêter, comprendre, conduire, venir). The methodology involved four distinct steps:

1. Corpus Construction:

   • Data Source: The EuroParl-Intersection corpus (Koehn, 2005), a multilingual corpus aligned across six languages.
   • Selection: 20 verbs were selected based on their polysemy, originally identified during the ARCADE campaign (2000).
   • Alignment: French (Source Language) and English (Target Language) versions were aligned at the word level using GIZA++ (Och and Ney, 2003). To ensure alignment validity, the authors performed bidirectional alignment (fr-en and en-fr) and filtered out instances where the target word in one direction did not appear in the reverse alignment.
   • Extraction: Contexts containing the 20 target verbs and their validated English translations were extracted.

2. Sense Inventory Definition:
   The authors generated four distinct sense inventories for every verb instance:

   • Translational (TRSL): Labels based on the actual English translation of the verb instance in the parallel corpus.
   • Lexicon-Grammar (LG): Labels based on the verb's entry in the Lexicon-Grammar tables (a computational dictionary for French). These entries are defined by syntactic and semantic features (e.g., subcategorization frames). An entry is denoted as {V,C}_table_entry (e.g., V_12_17).
   • Fine-Grained (LG#TRSL): A concatenation of the LG entry and the translation (e.g., V_12_17#understand). This creates a highly granular label combining syntactic-semantic evidence with cross-lingual evidence.
   • Clustered Fine-Grained: A clustering of the fine-grained labels where all instances sharing the same LG entry are grouped together, regardless of their specific translation.

3. Data Annotation:

   • LG Annotation: Performed manually by mapping verb instances to their corresponding entries in the Lexicon-Grammar tables.
   • TRSL Annotation: Performed automatically via word alignment and validated manually.
   • Combined Annotation: The LG#TRSL and Clustered inventories were generated automatically by concatenating or clustering the previously assigned labels.

Key Contributions

• New Corpus: The creation of a sense-tagged corpus containing 20 French polysemous verbs with a total of approximately 76,740 instances (based on the sum of samples in Table 2).
• Four-Label Annotation Scheme: The corpus is unique in that every instance is simultaneously tagged with four different types of sense labels, allowing for comparative analysis of different granularity levels and evidence sources.
• Hybrid Inventory Construction: The proposal and implementation of a method to combine cross-lingual translations with syntactic-semantic lexicon entries to create a "fine-grained" sense inventory that addresses the granularity issues of using either source alone.

Results and Statistics
The paper provides statistical descriptions of the corpus (Table 2), detailing the sample size and the number of unique senses for each of the three primary inventories (LG, TRSL, and LG#TRSL) across the 20 verbs.

• Sample Size: The corpus contains between 1,851 and 8,240 instances per verb, with an overall mean of 3,837 instances per verb.
• Sense Granularity:
  • The LG inventory shows a moderate number of senses per verb (mean of 11), reflecting syntactic-semantic distinctions.
  • The TRSL inventory shows a higher number of senses (mean of 129), reflecting the variety of translations.
  • The LG#TRSL inventory yields the highest granularity (mean of 222), representing the intersection of syntactic structure and specific translation.
• Example: For the verb comprendre, the corpus contains 8,240 instances, mapped to 8 LG entries, 183 unique translations, and 308 unique LG#TRSL combinations.

Significance and Claims
The authors modestly frame the paper as a resource-building effort rather than a report on WSD system performance. The significance lies in:

1. Addressing Granularity: Providing a solution to the problem that translational inventories may be too coarse for SL distributional analysis, while lexicon-style inventories may be too coarse for TL differentiation.
2. Resource Availability: Offering a dataset where senses are defined by both application-oriented (translation) and linguistic (syntactic-semantic) evidence, facilitating the development of WSD strategies that can leverage both types of information.
3. Future Work: The authors state their intention to extend this methodology to create similar corpora for 20 French polysemous nouns and 20 adjectives.

The paper does not present experimental results of a WSD classifier trained on this data, but rather establishes the corpus and the multi-layered annotation framework as a foundation for future research in cross-lingual WSD and sense inventory definition.