csi-corpus

This repository contains annotated screenplays [1] for 39 CSI:Crime Scene Investigation episodes from seasons 1-5. Each episode contains

1. word-level gold standard entity annotations using three types (perpetrator, suspect, other)
2. human behavioral perpetrator guesses (sentence-level)
3. case disambiguation information for episodes featuring two cases (sentence-level).

If you would like to obtain the visual and auditory features used in this work please email l.frermann@ed.ac.uk. Please note that in order for us to make these available to you, you must purchase the CSI:Crime Scene Investigation DVDs (seasons 1-5) which are available on sites like Amazon and eBay.

If you use this data in your research please cite the following paper:

Lea Frermann, Shay B. Cohen, Mirella Lapata (2017) Whodunnit? Crime Drama as a Case for Natural Language Understanding. Transactions of the Association for Computational Linguistics (TACL).

Annotation Format

Each file contains the screenplay corresponding to one CSI episode (identified in the title), one word per line. It contains the following annotations in tab-separated columns

• CaseID: indicates which case the sentence is relevant to (some episodes cover two cases)

• sentID: sentence counter (case specific)

• speaker: speaker of the utterance (None for scene descriptions)

• word: the actual word of the utterance

• killer_gold: binary indicator whether word refers to perpetrator (word-level label)

• other_gold: binary indicator whether word refers to any other participant in the plot (word-level label)

• suspect_gold: binary indicator whether word refers to suspect (word-level label)

• human_guess: human behavioral perpetrator guesses when watching the episode for the first time (sentence-level label)

• medion_time: mid-point of video segment corresponding to current sentence

• start_time: start time of video segment corresponding to current sentence

• end_time: end time of video segment corrqesponding to current sentence

• i1_time-i5_time: five equally spaced time points covering video segment corresponding to current sentence. Audio feature samples were taken at these points.

Note that except for the columns 'word', 'killer_gold', 'suspect_gold', and 'other_gold' all labels are provided on the sentence-level (i.e., their values are invariant for words in the same sentence).

---

[1] screenplays were downloaded from http://transcripts.foreverdreaming.org/