Multi-lingual Mathematical Word Problem Generation using Long Short Term Memory Networks with Enhanced Input Features

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 4709–4716 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 4709 Multi-lingual Mathematical Word Problem Generation using Long Short Term Memory Networks with Enhanced Input Features Vijini Liyanage, Surangika Ranathunga University of Moratuwa Katubedda 10400, Sri Lanka {vijiniliyanage.12, surangika}@cse.mrt.ac.lk Abstract A Mathematical Word Problem (MWP) differs from a general textual representation due to the fact that it is comprised of numerical quantities and units, in addition to text. Therefore, MWP generation should be carefully handled. When it comes to multi-lingual MWP generation, language specific morphological and syntactic features become additional constraints. Standard template-based MWP generation techniques are incapable of identifying these language specific constraints, particularly in morphologically rich yet low resource languages such as Sinhala and Tamil. This paper presents the use of a Long Short Term Memory (LSTM) network that is capable of generating elementary level MWPs, while satisfying the aforementioned constraints. Our approach feeds a combination of character embeddings, word embeddings, and Part of Speech (POS) tag embeddings to the LSTM, in which attention is provided for numerical values and units. We trained our model for three languages, English, Sinhala and Tamil using separate MWP datasets. Irrespective of the language and the type of the MWP, our model could generate accurate single sentenced and multi sentenced problems. Accuracy reported in terms of average BLEU score for English, Sinhala and Tamil languages were 22.97%, 24.49% and 20.74%, respectively. Keywords: Language Generation, Mathematical Word Problem, LSTM, Embeddings, Low- resource Languages 1. Introduction A Mathematical Word Problem (MWP) is ‘a mathematical ex- ercise, where significant background information on the prob- lem is presented as text rather than in mathematical notation’. (Moyer et al., 1984). Solving an MWP requires knowledge in Mathematics as well as in comprehension. There is only a limited amount of research done for multi- -lingual MWP generation. Existing research either focuses mainly on language-dependent Multiple Choice Questions (MCQs) (Chen et al., 2006), or are template based ap- proaches (Koncel-Kedziorski et al., 2016). Despite this dearth, MWP generation is a challenging task. Since MWPs constitute of a combination of texual facts, Mathematical quantities, units and notations, they tend to contain many con- straints. For example consider the MWP: ‘Maria made juice and she used 1 litre of water and 0.25 kg of sugar. How much more water than sugar did Maria use? ’. Here the numerical value that represents the amount of wa- ter should be of a higher value than that of which represents the amount of sugar. Moreover, relevant units (litre for water and kg for sugar) and appropriate combinations of substances or materials (Eg: water and sugar for juice) should be used. Thus, the generation of MWPs should be done while satisfy- ing these constraints. In the recent past, deep learning based techniques for Nat- ural Language Generation (NLG) have become popular among many research domains including spoken dialogue systems (Wen et al., 2015), story generation (Roemmele, 2016), lyric …

Content truncated to 500 words