Salt Lake City
- Tweets 2,344
- Following 598
- Followers 6,011
- Likes 9,071
Some good news! I'm joining the University of Utah @UUtah @UtahSoC as an Assistant Professor this summer. I'm excited I'll be part of the U's NLP group and work alongside @viveksrikumar & @EllenRiloff!
75
27
496
If you publicly shared your statement of purpose to help prospective PhD students, please can you reply with a link to it 🙏
20
89
405
No one:
Me: here is another collection of explainable AI papers
github.com/amarasovic/interp…
4
70
362
This Monday I gave the last lecture in my course on explainability 🫡
Readings/slides are available at utah-explainability.github.i…
+ lecture recordings at piped.video/playlist?list=PL… 🎬
3
50
244
28,374
I made a Jupyter notebook with examples of how I use matplotlib to visualize results in papers and posters :)
github.com/amarasovic/matplo…
1
53
226
This point is never mentioned: There are *so many* international grad students "stuck" in the US because they might be denied re-entrance / renewing their travel visa would take forever. Their only chance of presenting in person is that occasionally conferences are in the US.
5
13
226
45,074
My first lecture was my first remote lecture. Conclusion: I need human feedback 😅
I talked about probing and examining attention. Slides: github.com/amarasovic/presen…
2
17
117
How do you include inexperienced undergrads in NLP research? How were you included as an undergrad?
I'm talking about students who haven't yet completed all of the relevant coursework (to me DL, ML, NLP), or completed any research project. 1/
17
8
105
41,206
Returning to experiments I ran on Friday nitter.app/Skeeter696969/status/1…
2
11
102
#ACL2023 ended 10+ days ago, time to replace "coming soon" with actual code in your github repos
3
6
103
7,696
Maybe this is also a good time to announce that I'm on the faculty job market ‼️
Reach out if I’m a good fit!
My keynote on contrastive explanations at BlackboxNLP will be streamed on Nov 11 (Thur) at 1:15 PM (Seattle time*). There is a QA session after, looking forward to it!
* Other times:
- Punta Cana: 5:15 PM
- London: 9:15PM
- Beijing: 5:15AM
1
33
97
Risking turning my twitter into instagram, but c'mon
ALT A street covered in fresh snow. Pink colored sunset in the backround.
2
1
95
6,199
#ACL2023 #ACL2023NLP track "Interpretability and Analysis of Models for NLP" is missing 20% of assigned reviews 🥲
Please please please, if you can review even 1 paper until March 16, fill out this form: forms.gle/rzuTaspRyNpnS5KH9 🙏
11
79
97
40,043
.@huggingface appreciation at UtahNLP baking party 🤗🎄❤️☃️
ALT A group of people around a table in a living room. On the table are cookie cutters, rolling pins, dough, trays, etc.
ALT A tray with a big cookie in the huggingface emoji shape.
9
72
49,885
Watch out for our new arxiv preprint where we delve into a question: are the longstanding robustness issues in NLP resolved? I'll give you a glimpse of our findings. 1/
ALT Title: Whispers of Doubt Amidst Echoes of Triumph in NLP Robustness Authors: Ashim Gupta, Rishanth Rajendhran, Nathan Stringham, Vivek Srikumar, Ana Marasović Affiliation: Kahlert School of Computing, University of Utah
1
10
74
14,239
This work has been a huge help in recent discussions about explainable AI / NLP that I had, as well as shaping my thinking about the right future directions in this line of work
"We want to increase the user's trust in the model," or "we want a more trustworthy model" - you probably saw this sentiment in many papers. But what exactly does this mean?
New paper! --> arxiv.org/abs/2010.07487 @trustworthy_ml
With @anmarasovic @tmiller_unimelb @yoavgo
10
71
If you (like me) see a 600B model, and shriek, let me try to give you some consolation. Why should we care about ultra-large models?
1/n
1
15
74
Do you wonder if prompt-based finetuning & in-context learning can be extended to not only predict task labels, but to also generate explanations in plain English?
See our NAACL Findings paper (arxiv.org/abs/2111.08284) for answers! 👀
Code: github.com/allenai/feb
Details ⬇️
3
13
70
Rebranding is an effective way to overlook those who drove the original ideas into the mainstream.
Oana-Maria Camburu @oanacamb who popularized free-text explanations with e-SNLI should have her work cited in all your prompting papers that generate their reasoning + prediction.
2
9
69
20,164
To NLP/ML profs who still actively code for research:
1. What kind of coding tasks do you work on given the limited weekly time for this?
2. Do you code with your grad students? How does that affect their learning/stress?
3. Is a 20% industry role a best way to continue code?
4
5
67
17,514
My office right now 😍
ALT A photo of a view from a big office window. There are lots of colorful trees covered with snow.
1
64
If you don't know, now you know: transformers' attention matrix can be written as a sum of 2 parts. One is irrelevant for the output, the other is termed "effective attention" by Brunner et al
Are model interpretations from effective and original attention different? Answers👇🏻
Our paper “Effective Attention Sheds Light On Interpretability”(w/ @anmarasovic) was accepted into Findings of ACL2021 #ACL2021NLP #NLProc
Pre-print available at: arxiv.org/abs/2105.08855
Thread⬇️
1
4
61
Whoever is telling students to reach out to a prospective phd advisor and select one of their papers as a highlight of why they are interested to work with the advisor, is really not doing the service to the students
5
3
63
My keynote on contrastive explanations at BlackboxNLP will be streamed on Nov 11 (Thur) at 1:15 PM (Seattle time*). There is a QA session after, looking forward to it!
* Other times:
- Punta Cana: 5:15 PM
- London: 9:15PM
- Beijing: 5:15AM
1
13
62
Thrilled to see this work recognized at #EMNLP2025!
This framework and approach to measuring CoT faithfulness have been hugely influential for how I think about reasoning evaluation, and I'm so lucky to have worked with such brilliant collaborators. Huge credit to @mtutek
Very honored to be one out of seven outstanding papers at this years' EMNLP :)
Huge thanks to my amazing collaborators @fatemehc__ @anmarasovic @boknilev, this would not have been possible without them!
8
5
64
7,610
Y'all my dad got me a personalized jersey 🥹
ALT Croatian soccer jersey with Marasović on the back.
57
3,187
I'm recruiting students!
My interests include measuring usefulness of explanations for human-AI collaboration, addressing human factors that confound such measurements, & modeling interactive explainability (multimodality, few/zero-shot learning, dialogs, personalization, etc)
3
18
60
My reflections on approaches and ideas of @YejinChoinka, @dannydanr, and Percy Liang outlined at the @NAACLHLT workshop on generalization in #DeepLearning and #NLProc. Many thanks to @andrey_kurenkov, @tl0en, and @acganesh for taking the time and effort to improve this work.
Deep learning has made enormous strides in NLP, but state-of-the-art models are still spurious and brittle. @anmarasovic breaks down the problem and shares three ways we might solve it:
thegradient.pub/frontiers-of…
23
56
Utah is hiring tenure-track/tenured faculty and NLP is a priority area!
Please reach out over email if you have questions about the school and Salt Lake City, happy to share my experience so far.
utah.peopleadmin.com/posting…
14
54
9,928
Renders of the new building for @UtahSoC 😍 Seems like a piece of UW followed me [same architects].
Images provided by GSBS+LMN Architects.
ALT A render of the new building for the Kahlert School of Computing at the University of Utah, with a view of the valley. Images provided by GSBS+LMN Architects.
ALT A render of the new building for the Kahlert School of Computing at the University of Utah. Images provided by GSBS+LMN Architects.
4
3
52
6,613
Slides available here: docs.google.com/presentation…
I'll talk about measuring faithfulness of verbalized reasoning today at Repl4NLP at *9:45* #NAACL2025
10
56
7,970
I'm on my way to Seattle ✈️ Reach out if you'd like to meet during NAACL – especially if you're interested in applying to the University of Utah and working with me!
1
52
If you truly believe both in the existential risk and that society would be much worse without AI technologies, why don’t just build specialized, narrow AI models targeted for specific applications? I’m confused.
6
8
51
8,883
Your paper is rejected? No worries, in 2 years you might win the best rejected paper award 🙃
Congrats to @fiandola and @judyfhoffman for winning the (very unofficial) ICLR 2017/18 Best Rejected Paper awards! 🥇🥇
Some interactive graphs of ICLR Submissions 3 years on, with submissions linked to citations via @SemanticScholar:
markneumann.xyz/visualessays…
11
50
The University of Utah Persian Student Association reminded us today on our privilege to tweet a single hashtag. Talk to your Iranian students. Listen to their stories. #MahsaAmini
ALT A table with a sign that say "LOOK AT THEIR FACES they have protested, fought, and DIED for the RIGHTS of women in IRAN". On the table are photographs of the victims, two candles, and a bouquet. Above the table are different posters with messages about Mahsa Amini or her face, the slogan "women, life, freedom", and wearing hijab. Long hair is dominant.
5
15
46
📢 New at Findings #EMNLP2020 📢
"Natural Language Rationales with Full-Stack Visual Reasoning: From Pixels to Semantic Frames to Commonsense Graphs"
w/ @_csBhagav @jae_sung_park96 @Ronan_LeBras @nlpnoah @YejinChoinka
📖 Paper: arxiv.org/abs/2010.07526
Thread 👇
1
13
50
I'll talk about measuring faithfulness of verbalized reasoning today at Repl4NLP at *9:45* #NAACL2025
1
6
53
11,125
Au revoir NAACL! It was great to meet old friends and make new ones :))
ALT A sailboat on a lake, with downtown Seattle in the background.
48
I'm finally at #NeurIPS2023 too 👋
If you are an NLPer & you heard you should be doing *application-grounded evaluation of explanations*, but you are kind of lost, we got you! Fateme will present what you should have in mind tomorrow at @XAI_in_Action
openreview.net/pdf?id=8BR8Ea…
Thrilled to present our work at #NeurIPS2023 Workshop #XAI in Action on Dec. 16th!🤓🥳
"On Evaluating Explanation Utility for Human-AI Decision-Making in NLP"
📘Paper details: openreview.net/forum?id=8BR8…
⏲️Oral presentation: 16:07-16:14 | Poster Session: 16:30-17:30
🗺️Room: 271-273
2
3
47
8,409
Weekend plans for the next 0.5 year
ALT A photo from a ski lift, with skis in front, and lots of trees with snow piled on them.
44
Looking for examples [for teaching] where converting text to a set of n-grams or tf-idf features is not worse than using embeddings, or it is the only thing you could do given the scale of corpora and compute you have 🙏
15
4
42
29,253
Dear NLP professors, from what I hear these days, current applicants for PhD programs in the US need you to debunk "you need to do a PhD in top-N [American] schools to get a tenure-track position [you'll be happy with]” 😬
(dunno what N is and which ranking)
8
3
43
I will be at #EMNLP2024! My student 𝙁𝙖𝙩𝙚𝙢𝙚 𝙃𝙖𝙨𝙝𝙚𝙢𝙞 𝘾𝙝𝙖𝙡𝙚𝙨𝙝𝙩𝙤𝙧𝙞 @fatemehc__ will present "On Evaluating Explanation Utility for Human-AI Decision Making in NLP" in the poster session on 𝗪𝗲𝗱𝗻𝗲𝘀𝗱𝗮𝘆, 𝟭𝟬:𝟯𝟬𝗮𝗺: arxiv.org/abs/2407.03545 1/
1
6
45
2,792
Happening now!
ALT Alexis Ross presenting a poster with the title Does Self-Rationalization Improve Robustness to Spurious Correlations?
Does training models with free-text rationales facilitate learning *for the right reasons*? 🤔
We ask this question in our #EMNLP2022 paper, "Does Self-Rationalization Improve Robustness to Spurious Correlations?" arxiv.org/abs/2210.13575
W/ @anmarasovic @mattthemathman
🧵 1/n
1
44
The time for contrastive explanations of NLP models has come 🥳
In her ACL'21 Findings paper, @alexisjross proposes generating minimal edits of the input needed to explain "why [predicted label] instead of [another label]"
Her generators+code (khm, baseline) are available ⬇️
I'm happy to share that our paper "Explaining NLP Models via Minimal Contrastive Editing (MiCE)" was accepted into Findings of ACL 2021!
Updated paper: arxiv.org/abs/2012.13985
Code & models: github.com/allenai/mice
Work with @anmarasovic @mattthemathman
2
41
None of the students I've taken into my group had any of the requirements stated in the reddit post. I suppose many other advisors who are not at stanford, berkeley, cmu, uw, etc cannot recruit such students either.
So, if you are open to going to other schools, don't despair!
Folks here have no idea how competitive top PhD program admissions are these days, wow…
teddit.net/comments/1c2x5mx
37
17,453
This reminds it's time to share new datasets with explanation annotations that are added to exnlpdatasets.github.io/ 👇
For this week's @MilaNLProc reading group, @chiara_dibo presented "Teach Me to Explain: A Review of Datasets for Explainable Natural Language Processing" by @sarahwiegreffe & @anmarasovic
Paper: arxiv.org/abs/2102.12060
#NLProc #ReadingGroup #XAI
1
11
41
I want to have confidence of researchers who introduce a new dataset and don't report basic stats like the data/splits size
2
1
40
Someone is happy that spring is coming
ALT An orange cat lying on her back and sleeping on a sunny day.
38
1,795
Finally can enjoy afterwork hikes ☺️
ALT A short-horned lizard between dry leafs.
ALT Bring orange sunset from a dry, yellow hill. Mountains in rain and downtown Salt Lake City are in the background.
ALT Pink, orange clouds with an interesting "texture".
ALT More pink, orange clouds with an interesting "texture".
3
40
Universities be like "submit a diversity statement and when you do that, also specify your gender and you better be one of the two options we like". What a joke
1
1
37
Almost on my way to #emnlp2022! You can find me:
- Presenting CondaQA (arxiv.org/abs/2211.00295) on Sat 11AM
- Supporting @alexisjross who is presenting a poster on self-rationalization x robustness (arxiv.org/abs/2210.13575) on Sun 11AM
- Cheering for Croatia on Friday!!!⚽️🇭🇷
1
4
38
These streets, these mountains ❤️
ALT Omiš old town narrow street with stone buildings and a mountain at the end
ALT A photo of mountains surrounding the town of Omiš
3
37
Arriving to #ACL2025 #ACL2025NLP in a few hours!
See you at the welcome reception & catch me at the poster session on 𝐓𝐮𝐞𝐬𝐝𝐚𝐲 (𝐉𝐮𝐥𝐲 𝟐𝟗) 𝐚𝐭 𝟏𝟎:𝟑𝟎𝐚𝐦, where Jesse will present our work introducing new tasks for supporting legal brief writing
2
3
39
3,443
Utah/SLC area right now 😳
ALT Lots of very bright yellow colored trees. In the background is a big grey granite mountain with some red, orange, and green trees. Sky is blue with almost no clouds.
1
37
So large models have these nice properties, but they are accessible to the broader community only if they are compressed as well. Is now the time that every ultra-large model is released together with its compressed, matching version? 4/4 roberttlange.github.io/posts…
1
2
36
A job talk to the entire CS department: I can do this 👻
A lecture to a bunch of unknown students: 😳🥺🤯🥵😨😵💫😵🤕🤡
36
𝐻𝑜𝓌 𝑀𝓊𝒸𝒽 𝒞𝑜𝓃𝓈𝒾𝓈𝓉𝑒𝓃𝒸𝓎 𝐼𝓈 𝒴𝑜𝓊𝓇 𝒜𝒸𝒸𝓊𝓇𝒶𝒸𝓎 𝒲𝑜𝓇𝓉𝒽?
A new #blackboxNLP paper where we propose a supplemental measurement to contrast set consistency that enables discussion of whether a higher consistency was achievable with the same accuracy 🔎 1/
Some insightful recent works report model consistency across bundles of related instances. But since this naturally increases with accuracy, how should these consistency scores at different accuracies be compared?
Our paper for BlackboxNLP@EMNLP2023:
arxiv.org/abs/2310.13781
ALT Screenshot of the abstract for an academic paper "How Much Consistency Is Your Accuracy Worth?" by Jacob K Johnson and Ana Marasović "Contrast set consistency is a robustness measurement that evaluates the rate at which a model correctly responds to all instances in a bundle of minimally different examples relying on the same knowledge. To draw additional insights, we propose to complement consistency with relative consistency -- the probability that an equally accurate model would surpass the consistency of the proposed model, given a distribution over possible consistencies. Models with 100% relative consistency have reached a consistency peak for their accuracy. We reflect on prior work that reports consistency in contrast sets and observe that relative consistency can alter the assessment of a model's consistency compared to another. We anticipate that our proposed measurement and insights will influence future studies aiming to promote consistent behavior in models."
1
4
36
7,283
#nlphighlights 136: @alexisjross and I had a pleasure to talk with @kayo_yin and @malihealikhani about their position paper that calls for including signed languages in NLP. Thanks Kayo and Malihe for joining us!
soundcloud.com/nlp-highlight…
2
7
36
Academics, when we retire
ALT A booth at the SLC farmers market saying "OLD COOTS GIVING ADVICE It's probably bad advice, but it's free". There are four people standing in the line for it, and an elderly person sitting next to the booth's desk.
3
1
34
Preparing a lecture on detecting data artifacts with local explanations.
Is there a systematic study of whether current models (e.g. T5-11B, GPT-3, & SOTA multimodal models) use the same data shortcuts like models used in early 2018? Do they break with similar input changes? 1/
2
6
36
Hey #NLProc, if you have a good understanding of how BERT/Transformer works, but you keep forgetting the details, this could be useful: keen-goldberg-c5ab8b.netlify…
It is possible that I've misunderstood parts of this big architecture, please let me know if that is the case!
1
6
33
It makes me super happy to see contributions made to our repo with datasets for explainable NLP 🥰
exnlpdatasets.github.io/
Some new datasets for explainable NLP 👇
1
5
34
And actually live in Salt Lake City instead of visiting it all the time 🧗🏔️🚵⛷️ Where else can you snap all of this right next to the city? 😍
ALT A photo of the Antelope Island within the Great Salt Lake in Utah before sunset. Everything is colored in pink and light blue colors. There are two people standing in the distance.
ALT A photo from a hill next to Salt Lake City. There is a group of mountain bikers on the top of the hill, and the Great Salt Lake in the distance. It's obviously a sunny day with a clear blue sky.
ALT A photo of one of the canyons next to Salt Lake City, with the city, clouds, and other mountains in the background.
ALT A photo of Salt Lake City from a nearby hill.
2
33
I'm presenting our Findings paper at #blackboxNLP #emnlp2020 tomorrow (Friday) at:
(A) 10:30 AM - noon (Seattle time)
(B) 4 PM - 5:30 PM (Seattle time)
in GatherTown room K-N
Come talk to me about free-text rationales!
📢 New at Findings #EMNLP2020 📢
"Natural Language Rationales with Full-Stack Visual Reasoning: From Pixels to Semantic Frames to Commonsense Graphs"
w/ @_csBhagav @jae_sung_park96 @Ronan_LeBras @nlpnoah @YejinChoinka
📖 Paper: arxiv.org/abs/2010.07526
Thread 👇
3
10
33
Seems like it's time to re-share this excellent post:
theexclusive.org/2018/03/tou…
1
7
34
5,884
Our team opened applications for PYIs and interns! 🙌 Feel free to ask me any questions that you might have in DMs
4
9
33
#NLProc I'll lead a birds-of-a-feather meetup session on interpretability at #ACL2021NLP
What are the topics you'd like to see discussed?
3
2
30
fall 🤝 winter
ALT A photo of a backyard taken from a window. It's snowing but there are also colorful trees.
2
31
I'll co-organize SustaiNLP 2021 🌱 Consider voting for this workshop if you're interested in data/training/inference efficiency, justifying model complexity, model simplifications that give other benefits, and other related topics
Please vote for the workshop proposals for EACL/ACL-IJCNLP/EMNLP/ NAACL-HLT 2021 forms.gle/kkfsQZjjs2hFYiBfA
@ACL2020 @naacl @allenai_org @uwnlp @ACL_NLP
#ACL2020 #naacl #NLP #ACL_NLP
-- The EACL/ ACL-IJCNLP / EMNLP / NAACL-HLT 2021 Workshop chairs
8
30
#nlphighlights 127: Tosin Adewumi (@tosintwit) and Perez Ogayo (@a_ogayo) talked with Pradeep (@pdasigi) and me about "lowresourcedness" and how Masakhane (@MasakhaneNLP) is using participatory research to spur NLP in African languages. Thank you!!
soundcloud.com/nlp-highlight…
8
31
In between canyoneering, submitting a grant, preparing a tutorial, submitting a paper to ARR, and backpacking, I didn't highlight:
1. arxiv.org/abs/2311.09694 presented at NAACL
2. arxiv.org/abs/2402.14897 accepted to TMLR
Nate, an author of both works, is at NAACL!
More ⬇️
3
1
34
4,186
<rant> There are some excellent reviews from very junior researchers, and too bad there isn't a system where you can highlight them such that the future organizers know they can rely on them. </rant>
2
33
Just utah april after work things
ALT Skiers next to a giant pile of snow at the top of a ski lift.
ALT A sun setting behind snowy mountains.
ALT Tall trees covered with snow and mountains in distance.
ALT Photo of skies from a ski lift.
32
2,632
I'll moderate #emnlp2020 mentoring session on "Whether to do a PhD; and how to apply for and choose a PhD program"*
Leave your questions to the RocketChat channel #to-phd-or-not-to-phd
* Monday 10:00 AM (Seattle)
4
32
What kind of group meeting sessions besides ongoing project presentations, practice talks, paper discussions, low-key catchup, paper/abstract clinic, socials, do you have in your lab? Looking for ideas ☺️
7
30
Tomorrow @ #COLM2025:
1️⃣ Purbid's 𝐩𝐨𝐬𝐭𝐞𝐫 @ 𝐒𝐨𝐋𝐚𝐑 (𝟏𝟏:𝟏𝟓-𝟏:𝟎𝟎𝐩𝐦) on catching redundant preference pairs & how pruning them hurts accuracy
2️⃣ My 𝐭𝐚𝐥𝐤 @ 𝐗𝐋𝐋𝐌-𝐑𝐞𝐚𝐬𝐨𝐧-𝐏𝐥𝐚𝐧 (𝟏𝟐𝐩𝐦) on measuring CoT faithfulness by looking at internals
1/3
1
6
32
2,946
Looks like she is regretted her choice 😂
ALT More of Luca and Ana with the conference cat
4
1
31
1,508
It is striking to see the contrast between the coverage for static vs. interactive explanations in the taxonomy of AI explainability techniques proposed in Arya et al. (2019) arxiv.org/abs/1909.03012
1
6
29
Living in Seattle and visiting Salt Lake City regularly made the proximity to mountains a factor in academic job search for me, but simultaneously nothing seems to compare to WA and UT 😩
5
1
32
If you work on text generation from images+text, you (ironically) probably wonder what “general-purpose” multimodal model to use.
In our #EMNLP2022 Findings paper, we study this with self-rationalization and show that the answer is not simple. 1/
arxiv.org/abs/2205.11686
1
6
30