SemSeq: A Regime for Training Widely-Applicable Word-Sequence Encoders

A sentence encoder that can be readily employed in many applications or effectively fine-tuned to a specific task/domain is highly demanded. Such a sentence encoding technique would achieve a broader range of applications if it can deal with almost arbitrary word-sequences. This paper proposes a training regime for enabling encoders that can effectively deal with word-sequences of various kinds, including complete sentences, as well as incomplete sentences and phrases. The proposed training regime can be distinguished from existing methods in that it first extracts word-sequences of an arbitrary length from an unlabeled corpus of ordered or unordered sentences. An encoding model is then trained to predict the adjacency between these word-sequences. Herein an unordered sentence indicates an individual sentence without neighboring contextual sentences. In some NLP tasks, such as sentence classification, the semantic contents of an isolated sentence have to be properly encoded. Further, by employing rather unconstrained word-sequences extracted from a large corpus, without heavily relying on complete sentences, it is expected that linguistic expressions of various kinds are employed in the training. This property contributes to enhancing the applicability of the resulting word-sequence/sentence encoders. The experimental results obtained from supervised evaluation tasks demonstrated that the trained encoder achieved performance comparable to existing encoders while exhibiting superior performance in unsupervised evaluation tasks that involve incomplete sentences and phrases.