Can Spanish Be Simpler? LexSiS: Lexical Simplification for Spanish

Stefan Bott, Luz Rello, Biljana Drndarevic, Horacio SaggionCOLING 2012

TL;DR

This paper introduces LexSiS, the first system for automatically simplifying Spanish text by replacing difficult words with easier synonyms. It leverages readily available resources like online dictionaries and the web, and employs a novel approach combining word vector models, word frequency, and word length to determine simplicity. LexSiS demonstrates promising results, outperforming baseline methods in preserving meaning and offering simpler synonyms.

What Problem Does This Paper Address?

Automatic text simplification aims to make texts easier to understand by transforming them into equivalent versions while preserving the original content. This can involve various sub-tasks like syntactic simplification, content reduction, and lexical simplification. Lexical simplification, the focus of this paper, specifically targets replacing complex words with simpler synonyms. While significant progress has been made in English, with many systems relying on large parallel corpora, Spanish has lagged behind due to a lack of such resources. This paper addresses this gap by developing and evaluating LexSiS, a lexical simplification system for Spanish that utilizes freely available resources and does not require parallel corpora. This work is particularly relevant for improving text accessibility for individuals with reading comprehension difficulties.

Methods

LexSiS employs a two-step approach:

  1. Identifying the Correct Synonym: For a given word, LexSiS first checks if it has alternatives in the Spanish OpenThesaurus. It then uses a word vector model, trained on an 8 million word Spanish corpus, to represent the context of the target word and potential synonyms. By comparing the context vector with vectors representing different senses of the target word (derived from OpenThesaurus entries), LexSiS selects the word sense that best matches the context.
  2. Determining the Simplest Candidate: Within the chosen word sense, LexSiS identifies the simplest synonym. This simplicity score is calculated using a combination of:
  3. Word Length: A formula is used where shorter words are generally considered simpler, with a more pronounced simplicity gain for longer words.
  4. Word Frequency: The logarithm of the word's frequency count is used, assuming that more frequent words are simpler.
  5. Combined Simplicity Score: A weighted linear combination of the length and frequency scores is used, with weights determined through experimentation.

    Additionally, LexSiS incorporates a filter to discard substitutes that are not significantly simpler (a difference in simplicity score less than 0.5) or that have a low cosine distance to the context vector (indicating a poor semantic fit).

Datasets

  • Simplext Corpus: A corpus of 200 Spanish news articles, with 40 manually simplified. This corpus was used for corpus analysis (word length and frequency distributions) and to create a gold standard for evaluation. It was annotated with part-of-speech tagging, named entity recognition, and parsing using Freeling.
  • 8M word Spanish corpus from the Web: Used to train the word vector models.
  • Spanish OpenThesaurus (version 2): A freely available thesaurus used to find synonyms.
  • Evaluation Datasets:
    • T-S/B (Target vs. System and Baselines): 50 sentences with automatically generated substitutions by LexSiS and two baselines (Random and Frequency).
  • T-G/S/B (Target vs. Gold, System and Baselines): 12 sentences from the manually simplified gold standard, also processed by LexSiS and baselines.
  • G-T (Gold vs. Target): 14 instances from the gold standard where LexSiS did not propose a substitution.

Key Insights and Ideas

  • Leveraging Freely Available Resources: The paper demonstrates that it's possible to build a functional lexical simplification system for Spanish using readily available resources like online dictionaries and the web, without the need for expensive and time-consuming parallel corpus creation.
  • Combined Simplicity Metric: The approach of combining word length and word frequency, with specific formulas derived from corpus analysis, provides a more nuanced measure of simplicity than relying on a single factor.
  • Contextual Appropriateness is Crucial: The use of word vector models to ensure semantic fit (word sense disambiguation) is vital for producing meaningful simplifications, as demonstrated by the example of "hogar" (fireplace vs. house).
  • Spanish Simplification Challenges: The paper highlights the unique challenges of simplifying Spanish, particularly the lack of large-scale parallel resources compared to English.

Key Results

  • Meaning Preservation: LexSiS achieved a higher meaning preservation percentage (72.49%) compared to the Random baseline (65.03%) and the Frequency baseline (66.12%) across all datasets.
  • Simpler Synonyms: LexSiS offered simpler synonyms (30.16% simpler globally) more frequently than the Frequency baseline (28.42% simpler globally).
  • Performance on Gold Standard: When compared against the manually created gold standard, LexSiS showed a high synonym percentage (76.00%) and a significant proportion of simpler synonyms (35.96%).
  • Limitations of Gold Standard: The paper notes that the manual gold standard was small (26 substitutions) and that LexSiS could not simplify a substantial portion of these, often due to words being too frequent or not present in its dictionary.

Limitations and Future Work

  • Out-of-Dictionary Words: A significant limitation is the system's inability to handle words not present in its dictionary. Future work could involve inducing substitutes for these words using methods similar to the word vector model or larger datasets.
  • Frequency Computation without Word Sense Disambiguation: The current frequency calculation disregards word senses, which could be improved by incorporating word sense disambiguation before counting frequencies.
  • Single Lexical Unit Focus: LexSiS currently only handles single-word substitutions. Future work will address the identification and simplification of multi-word expressions and collocations.
  • Language Independence Potential: The approach, relying on small-scale, findable resources, suggests potential for language independence, which could be explored further.

Summary updated

BibTeX
@inproceedings{bott2012can,
  author = {Stefan Bott and Luz Rello and Biljana Drndarevic and Horacio Saggion},
  title = {{Can Spanish Be Simpler? LexSiS: Lexical Simplification for Spanish}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2012},
}
Figure