A Generative-Discriminative Framework for Title Generation in the E-commerce Domain
MetadataShow full item record
Multi-Engine (system combination) Machine Translation (MEMT) systems produce several outputs which will then be refined to produce a higher quality translation output. In the case where source sentences are absent in the pipeline, the task of post-editing will be even more challenging. This is then analogous to finding a consensus token sequence of user-input source sentences, in order to remove any potential errors. In a similar setting, major e-Commerce sites contain millions of user-created titles for a large number of products. These titles are similar in nature to translation outputs from different machine translation systems. Listed in these sites are products that have individually numerous source titles created by different sellers. Target product titles could serve as an effective way with which collections of products can be clustered and displayed. At the same time, the continuous upload of user-created titles calls for the need for new target product titles on a regular basis. Since the creation of effective target titles poses a challenge, many e-Commerce sites resort to manual curation of titles. However, the process of creating target titles would amount to great costs since manual construction of these titles is both a costly process and one that is never-ending. On the other hand, selection of the target product title from the existing seller-created source titles would subject the title to potential errors in the title. Therefore, the idea of using a source titles as a replacement of the ideal title would be futile. In light of the above scenarios, we aim to investigate a robust system that takes as input multiple user-input source strings with potential errors, and generate a single high-quality target string. More concretely, our research is performed on collected source title sets coming from the domain of e-Commerce, in order to mimic the generation of high-quality human-created target product title. Thus, this thesis focuses on the idea of creating target titles given existing source titles. We propose a Generative-Discriminative framework which consists of two main components: (i) a generative feature-rich decoder that transforms the initial source title search space into one with better quality generated titles, and (ii) a discriminative attention-based rescorer that selects the best title among the generated ones to be the predicted target title. For the practical need of the automatic generation, we propose a robust approach that could generalize the supervised models to large unseen datasets. Moreover, we tested our systems in a bi-lingual setting and found that it performed competitively in both German and English.
- Linguistics