Date of Award

3-10-2019

Publication Type

Master Thesis

Degree Name

M.C.Sc.

Department

Computer Science

First Advisor

Jianguo Lu

Keywords

distributional word representations, dynamic mutual information, word co-occurrences

Rights

info:eu-repo/semantics/openAccess

Creative Commons License

Creative Commons Attribution-Noncommercial-No Derivative Works 4.0 License
This work is licensed under a Creative Commons Attribution-Noncommercial-No Derivative Works 4.0 License.

Abstract

Semantic relations between words are crucial for information retrieval and natural language processing tasks. Distributional representations are based on word co-occurrence, and have been proven successful. Recent neural network approaches such as Word2vec and Glove are all derived from co-occurrence information. In particular, they are based on Shifted Positive Pointwise Mutual Information (SPPMI). In SPPMI, PMI values are shifted uniformly by a constant, which is typically five. Although SPPMI is effective in practice, it lacks theoretical explanation, and has space for improvement. Intuitively, shifting is to remove co-occurrence pairs that could have co-occurred due to randomness, i.e., the pairs whose expected co-occurrence count is close to its observed appearances. We propose a new shifting scheme, called Dynamic Mutual Information (DMI), where the shifting is based on the variance of co-occurrences and Chebyshev's Inequality. Intuitively, DMI shifts more aggressively for rare word pairs. We demonstrate that DMI outperforms the state-of-the-art SPPMI in a variety of word similarity evaluation tasks.

Share

COinS