Very basic question

[Marketa PS]

Dear all, 

I am new to using Inception, and I don't know everything about its functionalities. I apologize for my very basic question. 

What we want to do in our project is quite simple: tag ritual actions in 2 layers: relationships and entities.

For example:

"You should consume bread". 

Relationship layer:

You-consume -> relationship called agency (actor does action)

consume-bread -> relationship called patiency (action is done to object)

Entity layer

you - > entity called practitioner

consume -> entity  called action (consume)

bread -> entity called food

My question is this: If I upload 100 texts and I want to see the relationship between bread-consume, how to I do this? Right now, I have created annotations, but I don't know how to make use of them across texts... Right now, my annotations are not connected to the Knowledge base layer, and I wonder if that is necessary. 

If you search in the demo for "CoMaF and Mageia", you will see two texts that I annotated. 

Many thanks!

Markéta

[Richard Eckart de Castilho]

Hi Markéta,

I've had a brief look at the "CoMaF and Mageia" on demo.

I have to be honest, I don't understand the annotations scheme quite well.
You have quite a few different features on both the your span and your relation
layers but most of the annotations only have a single feature actually set
and often that feature value is set to the annotated text. E.g. you annotate
the word "agarwood" with an "Entity" and set the feature "Incense" to "agarwood".
Have you considered annotating the word "agarwood" just with "Entity" and setting
e.g. a "label" feature of Entity to "Incense" instead?

Similar for relations. I would guess a Relation layer with a single feature "label"
that indicates the kind of relation between the two linked entities might suffice?

Regarding your question:

Do you imagine uploading 100 texts and then annotating those as well and finally
you would like to get an idea e.g. how often a certain type of relation exists
between certain entities? In that case, the "Explorer" view might help you, although
with the current annotation schema design, it would seem a bit tricky to extract
useful statistics.

Or are you asking because you would like INCEpTION to use the annotations you
already made to automatically generate similar annotation on the remaining 100
texts? INCEpTION has a few built-in recommenders for spans that you might use,
but two texts will not provide sufficient annotation examples for those and the
tasks may be a bit complex for them the built-in recommenders. Also, we only have
a very very basic string-matching relation recommender built-in and that's
experimental and not enabled on the demo machine. Also, it would only kick in
after you have already annotated the entities. So while you could use INCEpTION
to annotate your texts manually and possibly with some help from at least some
span recommenders, INCEpTION can't just auto-annotate your 100 texts based on
a few examples.

So for the moment, I would propose you try creating a project using the
"Basic annotation (span / relation) template and see if you can get by just
using the single simple "label" feature defined by those. That project also
already comes with a "Generic span recommender" enabled. You can add a second
one with layer "Span", feature "Label" and type "Multi-token sequence classifier".

I hope that helps. If not, maybe it helps getting to more questions :)

Cheers,

-- Richard

[Marketa PS]

Dear Richard, 

many thanks for your kind reply. I apologize again, I am completely new to this!

I have now created a new project (CoMaF and Mageia 2) as you recommended, using the Basic annotation (span / relation) template and I added the recommender as suggested. 

I see now that what you propose is much better than what I had. 

Regarding your questions: I did not expect Inception to do automatic tagging for us; I didn't even know much about the recommenders, but already in tagging these two texts, I see that they make the work much faster! We want to use it to tag potentially thousands of texts, so we can see cross-cultural similarities between rituals (for example, in which cultures do people bind amulets on their arms? When is wine drunk, and when is it poured on a surface?); a few experts would do the tagging manually. 

I have a few further questions: 

- For the case that the same entity appears multiple times in a text, I created a chain layer. But I am not really sure if that is useful for my purposes? For instance, I tagged "amulet" and "it" (which refers to it) as amulet, which I think suffices for my purposes. What would be the benefit of the chain layer? 

- Could you please help me with the formulation of the search query in the search function in annotation? I would like to look for, for instance, "patiency" label within "Relation" layer with "amulet" as the target. In the screenshot, in the relation that I highlighted, the target is "it" and not "amulet". Would the search be able to identify the "it" as "amulet"? Or is this what the chain layer (which I have also created) is for?

I think these questions suffice for now. Many thanks again!

Markéta

--
You received this message because you are subscribed to the Google Groups "inception-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to inception-use...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/inception-users/D6271DF2-4789-4563-BF95-8F1EFCE003BE%40gmail.com.

[Richard Eckart de Castilho]

Hi,

> Regarding your questions: I did not expect Inception to do automatic tagging for us; I didn't even know much about the recommenders, but already in tagging these two texts, I see that they make the work much faster! We want to use it to tag potentially thousands of texts, so we can see cross-cultural similarities between rituals (for example, in which cultures do people bind amulets on their arms? When is wine drunk, and when is it poured on a surface?); a few experts would do the tagging manually.

Nice. Now with the scheme you now have, you can also build a nice overview of the relations using the "Explorer".

[Marketa PS]

Dear Richard, 

many thanks for this!

Thank you for the SemPred/SemArg suggestion, but I think I'll keep it as it is for now, due to the search and explorer functionality. 

I have a few other questions:

- how do I turn off a recommendation, if it is not correct? (see image) Or do I just leave it like that?

- concerning my question with the "chain". I see that I cannot create relations across layers. In that case, can I tag words in a knowledge base layer by default instead of a normal span layer? What I mean is that now, I tag in Span, in Relation, and then in KB layer so that things show up in the Knowledge Base. But I guess I could get away with just using KB layer instead of Span? See the attached image ("Bind it to yourself" - you see that I tagged amulet and beneficiary twice, one in the span and one in the KB layer). I guess what I would loose is the "Label" 

- because our texts are translations from Coptic, I was wondering how you deal - if at all - with original text and translations. I guess to have both visible, we would have to upload both the Coptic and the English in one txt file and organize it the way we want. I could then tag both the English and the Coptic (but I guess I would have to do that manually each time). Has anyone else worked with original and translation?

Many thanks! Have a good week, 

Markéta

--
You received this message because you are subscribed to the Google Groups "inception-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to inception-use...@googlegroups.com.

To view this discussion visit https://groups.google.com/d/msgid/inception-users/243977E6-F0AD-4B59-A529-18BCC277EF1E%40gmail.com.


Btw. I have tried to use the "SemPred" and "SemArg" layers in "InceptionT1329c.txt". I think would in principle be well suited for your kind of annotation task where you annotate activities and their participants, objects and cirumstances. However, I not that the "link features" that
SemArg uses are not yet supported in the explorer. So feel free to have a look at the sample annotation I made, although with the current version, keep in mind it does not work with the explorer.

If you want to try the SemPred yourself, here is an example:

- Switch to the SemPred layer
- Mark the trigger of the action/frame (usually a verb)
- Click on "add" in the right sidebar
- Mark an argument of the action (e.g. the beneficiary)
- Click on the added argument in the sight sidebar again so it gets orange
- Enter the label (e.g. beneficiary) into the field below and click "Set"

(I notice again this workflow needs to be improved... actually, it should be possible to enter the label before pressing "add"...).

Also, I see below you want to use search - that would also favor relations over SemPred/link features since the latter is supported in search at all.

For a theory behind this kind of annotation, please look at https://en.wikipedia.org/wiki/FrameNet and related resources.

Mind that theory and layers are independent. You could model this theory also using relations (as you do) instead of link features. But in contrast to your current annotation scheme, the theory recognizes that a frame is essentially an n-ary relation while in your scheme, you currently seem to be focussed on binary relations.

> I have a few further questions:
> - For the case that the same entity appears multiple times in a text, I created a chain layer. But I am not really sure if that is useful for my purposes? For instance, I tagged "amulet" and "it" (which refers to it) as amulet, which I think suffices for my purposes. What would be the benefit of the chain layer?

I would recommend not using chain layers. In particular because these are (still) not supported during curation (relations and link-features are both supported). Also, chains cannot have custom properties, chains cannot create cross-document references, and are also not supported in search. So unfortunately, they are a bit limited.

If you want track entities appearing throughout your corpus, setting up a knowledge base of these entities and using a concept feature to link to them would seem more useful to me.
An alternative would be using a simple string feature with an open tagset (i.e. one which auto-adds new tags as annotators make them). However, here you'd need to be careful that you have good rule for tag naming to avoid the same tag accidentially being used for unrelated entities or duplicate entities being introduced. Also note that tags cannot be renamed. You can change the tag definition, but existing annotations will not be updated. In a KB, you can change a concept label and all the concepts will show with the new label afterwards. However note that in the actual exported data, not the label will be visible, but rather the concept ID from the knowledge base.

> - Could you please help me with the formulation of the search query in the search function in annotation? I would like to look for, for instance, "patiency" label within "Relation" layer with "amulet" as the target. In the screenshot, in the relation that I highlighted, the target is "it" and not "amulet". Would the search be able to identify the "it" as "amulet"? Or is this what the chain layer (which I have also created) is for?<Screenshot 2026-03-04 at 09.39.22.png>

So what you can currently do is e.g. this:

<Relation.Label="patiency"/> fullyalignedwith <Relation-target.Label="amulet"/>

That will find a relation with "Label" having value "patiency" that is at the same position as a relation that has a target which has the "Label" "amulet". Note that the two relations may not necessarily be the same if you have stacked relations at the same position.

Support for relations in the underlying corpus query tool we use is non-existent, so we had
to bolt this support on top. Certainly improvable. Or better, we would switch to another
search backend. I already have one in mind, but didn't find the time yet to make the switch.

Anyway, I hope this helps.

Cheers,

-- Richard





--
You received this message because you are subscribed to the Google Groups "inception-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to inception-use...@googlegroups.com.

To view this discussion visit https://groups.google.com/d/msgid/inception-users/243977E6-F0AD-4B59-A529-18BCC277EF1E%40gmail.com.

[Richard Eckart de Castilho]

Hi Markéta,

> On 9. Mar 2026, at 11:53, Marketa PS <preininger...@gmail.com> wrote:
>
> I have a few other questions:
>
> - how do I turn off a recommendation, if it is not correct? (see image) Or do I just leave it like that?

You can double-click on it to reject it. Or if you open the annotation overview sidebar, there are explicit accept and reject buttons.

> - concerning my question with the "chain". I see that I cannot create relations across layers. In that case, can I tag words in a knowledge base layer by default instead of a normal span layer? What I mean is that now, I tag in Span, in Relation, and then in KB layer so that things show up in the Knowledge Base. But I guess I could get away with just using KB layer instead of Span? See the attached image ("Bind it to yourself" - you see that I tagged amulet and beneficiary twice, one in the span and one in the KB layer). I guess what I would loose is the "Label"

You could maybe just add a concept feature to the "Span" layer?

In theory, there is an experimental mode for cross-layer relations, but it is not enabled on the demo instance.
There may also be unknown bugs when using this mode.

> - because our texts are translations from Coptic, I was wondering how you deal - if at all - with original text and translations. I guess to have both visible, we would have to upload both the Coptic and the English in one txt file and organize it the way we want. I could then tag both the English and the Coptic (but I guess I would have to do that manually each time). Has anyone else worked with original and translation?

INCEpTION doesn't have explicit support for parallel texts.

But you could either upload your data with a block of Coptic at the start and the translation below.
Or you could upload with interlaced sentences, e.g. one sentence in Coptic, then the translation, then an empty line.

In this case, I would strongly recommend to format your files such that there each line is a sentence (in either language)
and choose the format "Plain text (one sentence per line)" during upload. That ensures that each line is treated as a
sentence by INCEpTION - it will then not try to do sentence-detection.

Cheers,

-- Richard