D 2022

Combining Sparse and Dense Information Retrieval: Soft Vector Space Model and MathBERTa at ARQMath-3 Task 1 (Answer Retrieval)

NOVOTNÝ, Vít and Michal ŠTEFÁNIK

Basic information

Original name

Combining Sparse and Dense Information Retrieval: Soft Vector Space Model and MathBERTa at ARQMath-3 Task 1 (Answer Retrieval)

Authors

NOVOTNÝ, Vít (203 Czech Republic, guarantor, belonging to the institution) and Michal ŠTEFÁNIK (703 Slovakia, belonging to the institution)

Edition

Bologna, Proceedings of the Working Notes of CLEF 2022 - Conference and Labs of the Evaluation Forum, p. 104-118, 15 pp. 2022

Publisher

CEUR-WS

Other information

Language

English

Type of outcome

Stať ve sborníku

Field of Study

10201 Computer sciences, information science, bioinformatics

Country of publisher

Italy

Confidentiality degree

není předmětem státního či obchodního tajemství

Publication form

electronic version available online

References:

RIV identification code

RIV/00216224:14330/22:00126431

Organization unit

Faculty of Informatics

ISSN

Keywords in English

information retrieval; sparse retrieval; dense retrieval; soft vector space model; math representations; word embeddings; constrained positional weighting; decontextualization; word2vec; transformers

Tags

Tags

International impact, Reviewed
Změněno: 6/4/2023 09:35, RNDr. Pavel Šmerk, Ph.D.

Abstract

V originále

Sparse retrieval techniques can detect exact matches, but are inadequate for mathematical texts, where the same information can be expressed as either text or math. The soft vector space model has been shown to improve sparse retrieval on semantic text similarity, text classification, and machine translation evaluation tasks, but it has not yet been properly evaluated on math information retrieval. In our work, we compare the soft vector space model against standard sparse retrieval baselines and state-of-the-art math information retrieval systems from Task 1 (Answer Retrieval) of the ARQMath-3 lab. We evaluate the impact of different math representations, different notions of similarity between key words and math symbols ranging from Levenshtein distances to deep neural language models, and different ways of combining text and math. We show that using the soft vector space model consistently improves effectiveness compared to using standard sparse retrieval techniques. We also show that the Tangent-L math representation achieves better effectiveness than LaTeX, and that modeling text and math separately using two models improves effectiveness compared to jointly modeling text and math using a single model. Lastly, we show that different math representations and different ways of combining text and math benefit from different notions of similarity between tokens. Our best system achieves NDCG' of 0.251 on Task 1 of the ARQMath-3 lab.