Geoffrey Zweig and Christopher J.C. Burges
Work on modeling semantics in text is progressing quickly, yet there are few existing public datasets which authors can use to measure and compare their systems. This work takes a step towards addressing this issue. We present the MSR Sentence Completion Challenge Data, which consists of 1,040 sentences, each of which has four impostor sentences, in which a single (fixed) word in the original sentence has been replaced by an impostor word with similar occurrence statistics. For each sentence the task is then to determine which of the five choices for that word is the correct one. This data was constructed from Project Gutenberg data. Seed sentences were selected from Sherlock Holmes novels, and then imposter words were suggested with the aid of a language model trained on over 500 19th century novels. The language model was used to compute 30 alternative words for a given low frequency word in a sentence, and human judges then picked the 4 best impostor words, based on a set of provided guidelines. Although the data presented here will not be changed, this is still a work in progress, and we plan to add similar datasets based on other sources. This technical report is a living document and will be updated appropriately as new datasets are constructed and new results on existing datasets (for example, using human subjects) are reported.