Contains 140 documents where each document is a fully reconstructed transcript from CallHome
Each document averages about 1,800 words
The dataset contains a total of about 257,000 words across the dataset

Intended Use Case

This dataset is intended to be used as part of a Continued Pretraining (CPT) corpus for exposing a base model to highly informal and conversational Spanish.

Dataset Map

The dataset is a parquet file consisting of columns for the main Continued Pretraining documents, original CallHome utterances, and cleaning history for each document.

file_path: Contains the name of the .cha file from the spa corpus the transcript is associated with.
reconstructed_document: The reconstructed document that can be fed directly into a CPT pipeline. Documents are reconstructed by appending the speaker label (for instance A:) to the start of the cleaned utterance, and seperating cleaned utterances by a single newline.
cleaned_utterances: The list of utterances consisting of the document after they went through cleaning.
original_utterances: Utterances as they originally appeared after being parsed by pylangacq chats.utterances.
audit_trail: Shows what was cleaned in the document. For example, consider Doc Line 34 (Raw ID 33) - B1: ["ANNOTATION ARTIFACT REMOVED: 'background noise'"]. Line 34 tells us this applies to the 34th line of the clean utterances, and the 33rd line of the uncleaned utterances. The action tells us the string "background noise" was removed from the utterance.

Dataset Creation

Downloading & Loading

Spa.zip containing the transcripts for 140 recorded calls in Spanish was downloaded from TalkBank here. Spa.zip was then loaded with pylangacq for inspection.

Analysis

Next the dataset was analyzed for English tokens to ensure the data was uniformally Spanish. This Spanish-English Language ID model was used to create a list of potential English leakage into the dataset. The model returned hits for English and "other" tokens across the corpus. Notes are provided about the results of this process below.

### Discoveries About the Contamination

1198.cha contains 200 hits for "noise" on ctrl + f inspection
has is flagging on benign Spanish words like "hasta", "muchas", and "has" is a legitmate form of haber in reference to tu (you informal)
1255.cha yields 52 hits for "crying" on ctrl + f inspection indicating it is a major source of the contamination for this word
NOTE: leaked annotations like "noise" and "crying" appear to almost exclusively come at the beginning or end of an utterance
oye is a conjugation of "oír" and legitmate Spanish
clear is an artifact of pylangacq not parsing &=throat clear correctly, can be dropped in entirety without causing linguistic damage
vocalization appears to be an artifact of parsing from annotations like &=child vocalization as demonstrated throughout 0731.cha. Can confidently drop with regex filter
NOTE: After removing the offending annotation artifact some lines may be completely blank, these turns should be completely omitted from the final corpus
Keep once instances as these are either extracted from "entonces" or refer to the number "once" (11)
Uses of dime seem to be legitimate and most of the time appears alone as "dime" (tell me)
Ellipses ... are a part of standard speech representing silence, the ellipses can stay in the corpus without modification
+/. seems to represent almost "half baked" thoughts, we can remove them from the dataset with regex filter (meaning of text won't change)
Additonal: instances of underscored English not being split correctly, for instance "British_Airways", 124 instances of English phrases seperated by 1 or more "_" across the dataset

Response Plan

Keep upstream data "pure" until after this cell
Ensure Removal Phase tracks the diff between original utterance and newly cleaned utterance to ensure proper Spanish is not being broken by our cleaning strategy (audit logging)
Replace all instances of "background noise" in text with "", then replace all instances of "noise" with "" ONLY if it is surrounded by whitespace
Replace all instances of "crying" in text with ""
Replace all instances of "clear" with ""
Replace all instances of "vocalization" with ""
Rerun LID post cleaning and analyze results

Cleaning

All cleaning actions are noted in the audit_trail column of the dataset (see explanation in Dataset Map Section above).

Code Switching: Somtimes legitimate code switching between Spanish and English occured in the dataset. These code switching phrases would appear with underlines instead of spaces seperating words. For instance, "green card" would appear as green_card in the raw corpus. During this step the raw English phrases were retained, but we swaped _'s for ' '.
Both the strings "background noise" and "noise" were found to be contaminating the dataset. First we replaced all "background noise" instances with an empty string ("").
A few hundred instances of stray annotation marks were observed including: "crying", "clear", "vocalization", and "noise". These instances were only stripped and removed if they were "alone" (surrounded by whitespace characters) or at the beginning or end of an utterance. Surrounding whitespace was consumed to avoid leaving stray double spaces upon removal.
Next, the two most agressive symbols from the "other" category were cleaned. First we replaced instances of +... with a standard ellipses (...). Then, replaced all instances of + remaining with nothing.
Finally, whitespace normalization was conducted on the utterances replacing all blocks of 1 or more whitespace with a single space ( ).

GitHub

The full code and methodology of the cleaning pipeline can be found here

Acknowledgements

TalkBank Citation:

Linguistic Data Consortium (2008). CABank Spanish CallHome Corpus. TalkBank. doi:10.21415/T51K54

Canavan, A., & Zipperlen, G. (1996). CALLHOME Spanish Speech LDC96S35. Philadelphia: Linguistic Data Consortium.

Thank you to sargosarker for providing the Spanish/English LangID model used for analysis.

Downloads last month: 50

Total file size:

2.4 MB